Humans buffered the old system

Someone on your team is the layer.

You probably know who. The person others cc on everything not because policy requires it but because institutional memory requires it. The one whose vacation gets quietly canceled. Whose exit gets quietly feared. Not because they are irreplaceable in the heroic sense. Because the operation stops running on process the day they leave and starts running on whoever is willing to compensate by buffering the system manually.

That is not a talent problem. It is a structural one.

The Coasean inversion accelerated the runtime clock. AI systems execute at machine pace. Decisions propagate results many times over before the next review cycle. Drift compounds in the gap between what the system is doing and what the organization intended. The person absorbing that gap is not failing. They are doing the work the layer was supposed to do.

The layer was never built.

Cadence and Signal

This is not a new failure mode. It is an old one running at new speed.

The quarterly review was calibrated for a world where decisions propagated slowly enough that the next cycle could catch the drift. That world is gone.

AI systems do not wait for the calendar. They execute, adapt, and compound between check-ins. The slice width, the interval between your reviews, was set when the runtime clock moved at human pace.

The Nyquist principle in signal processing says you must sample at least twice as fast as the phenomenon you are trying to catch. Miss that threshold and the signal aliases. It looks clean. It is not. The drift is there. Your slice was too wide to see it.

People are now paired with AI primitives that carry the machine’s speed without requiring the human to match it. The machine runs at Nyquist-adequate frequency. The human does not have to. The alert arrives because the system already caught the frame. Deliberation starts from evidence, not from the anxiety of potentially having missed something.

That is the functional distinction between oversight done right and oversight as performance. The machine samples. The anomaly surfaces. The human deliberates on a prepared question rather than a raw feed.

The question is not whether drift occurs. Some variance is expected and allowable. The question is whether the seam is instrumented well enough to distinguish tolerable variance from compounding misalignment before the signal aliases into apparent stability.

That is a calibration question. Part science because the signal tells you when the slice is too wide. Part art because judgment decides what tolerable looks like. The calendar cannot answer it.

Same work. The clock changed. The slice has to follow.

Share

Cadence is not a meeting schedule. It is a design decision.

Multiple rhythms run inside the Hybrid Enterprise. Two types: machine and human.

Among humans there are three meaningful rhythms, each operating at a different altitude, each contributing to how the enterprise shows up externally in the market and holds together internally.

Build rhythm. Close to the machine, execution pace. F1 and F0.

Alignment rhythm. The cadence at which floors stay coherent with each other. F2 and F3.

Judgment rhythm. The slower, higher-stakes deliberation that sets direction and reads whether the portfolio is compounding or just accumulating. F3 and F4.

Each requires a different slice width. Each is vulnerable to a different failure mode when the cadence is wrong.

(We explored these rhythms in an earlier piece, “A Little Friction by Design.”)

The operating layer does not force these rhythms to match. It chooses when to couple them and when to deliberately uncouple. That choice is not administrative. It is architectural.

Coupled cadence is appropriate when the decision requires human judgment at machine speed. An agent flags an anomaly. The alert surfaces. The human acts. The loop closes.

Uncoupled cadence is appropriate when the human rhythm operates at a longer horizon than the machine. The VP does not review every transaction. She reviews the pattern the transactions produced.

Different slice width. Different question. Same operating layer underneath both.

Most organizations never make that choice. They simply inherit the rhythms and do not know or ask why. The calendar, the standing meeting, the quarterly business review. The machine compounds underneath it.

That default is not neutral. It is a structural decision made overtly, else by omission.

Bounded execution is the companion principle. An agent operating without a defined scope is not autonomous. It is unmonitored.

Scope defines the boundary within which the machine runs without interruption. The human reviews at the boundary, not inside every step.

The boundary is the design. The review is the checkpoint where judgment enters. With both, cadence is a mechanism. At the right intervals along the seams we capture the signals that matter.

Signal is what makes judgment actionable. The right cadence captures valuable signals.

A review cycle without a signal protocol is a calendar event. It produces reports. It does not catch drift.

Signal is the operating layer’s sensory system. It tells you what changed, what compounded, what crossed the threshold that matters. Not everything. The things that require judgment.

The foils are recognizable.

The strategy document reviewed on a schedule. Thorough, well-formatted, updated quarterly. It tells you what was decided. It does not tell you when the schedule is wrong or when the conditions that justified the decision have shifted. The artifact is not the valve. The calendar is not the signal.

The roadmap review that names the runtime problem and answers with a planning artifact. The right diagnosis, the wrong repair. Runtime requires a mechanism. A plan is a snapshot relevant under a few conditions. Good prep. Good wargaming, but snapshots do not catch drift.

And the most common one: the person holding continuity by hand. The one who knows what was decided, why, and what changed since. Not because the system carries it. Because they do.

That is compensation, a human buffering a fragmented system, not stewardship.

In the matrix management era humans buffered for an imperfect system. At human pace that held. At machine pace it does not. The cost is invisible until they leave.

Each pattern is a reasonable response to a real problem. None of them is a designed property. Without both Signal and Cadence working together, the review is a ritual and the drift is undetected.

Records carry forward

Cadence and Signal catch drift in the present. Records carry judgment forward in time.

A decision made well, documented poorly, is a decision made once. The next person inherits the outcome without the reasoning. The next cycle inherits the policy without the context that justified it. The organization re-litigates intent from scratch because the intent was never formally encoded. It lived in the room, in the relationship, in the person who has since moved on.

This is not a documentation problem. It is a continuity problem.

Records, as a primitive, are not artifacts. They are callable shared memory. A Record carries the decision, the basis, and the conditions that held at the point it was made. It is a snapshot of judgment at a moment in time, invokable by the next person, the next cycle, the next agent operating in the same domain.

The difference between a slide deck and a Record is not format. It is verbs. A slide deck describes. A Record enables. It is the difference between reading a map and having a guide who was there.

The memory continuity layer is what Records build over time. Not a database. Not an archive. A living substrate that carries institutional judgment across leadership transitions, reorgs, and the natural turnover that erodes tacit knowledge.

Without the memory continuity layer the organization is perpetually new.

Pilots run and are forgotten. Decisions repeat because the record of why they failed the first time does not exist. Leadership transitions reset the clock. The institutional memory that should compound instead evaporates with every exit.

The Agent Decision Record (ADR) is the runtime face of this primitive. When an agent acts, the ADR captures what it did, why, under what authority, and what the conditions were. Not for compliance. For continuity.

The human reviewing the next cycle does not reconstruct intent from outputs. They read the record of the judgment that produced them. That is non-reflexive deliberation with memory. That is the operating layer working.

Most accountability frameworks capture what an agent did after it acted. The ADR records the action and its conditions. That is already an improvement over nothing.

The Hybrid Enterprise goes one layer deeper. Before the agent receives its instruction, the human judgment that produced that instruction is recorded. Who made the call. What they knew at the time. What boundaries they set. What they were trying to achieve.

Human deliberates. Judgment is recorded. Instruction is issued. Agent acts. Not: agent acts, outcome is reviewed, accountability assigned.

A procurement agent is authorized to approve vendor contracts under a defined threshold. The ADR captures each approval. The Hybrid Enterprise also carries the record of the human judgment session where that threshold was set, why it was set there, what conditions held at the time, and who owned the decision. When the threshold produces a bad outcome six months later, the organization does not ask what the agent did. It asks what was the basis for the instruction the agent was given. That basis exists. It is traceable. It is immutable.

That sequence matters. Accountability that arrives after the fact is forensic. Accountability encoded before the fact is structural.

ROX, return on experiment, is the cadence argument from the firm’s seat. Records are what make ROX real. Without them, experiments are episodes. The learning evaporates. The portfolio accumulates activity, not judgment. With them, each experiment compounds into the next. The VP at F3 is not running a dozen pilots. She is building a judgment portfolio. The difference is Records.

The posture is yours

Wherever you already stand, the floor does not change. Our posture does.

That is the answer to the question the arc has been building toward. Not where should I be. Not what role should I hold. Where you are is where you start. The operating layer does not require a title change or a reorg. It requires a posture adopted at the altitude you already occupy. The titles by floor that follow are stand-ins for the horizon, not the point.

At F0 the posture is chassis. The infrastructure that builds, connects and governs every AI primitive, every agent, every access control, every data permission. Invisible when it works. That invisibility is the job.

At F1 the posture is authorship. The practitioner who flags the drift in the output before it compounds is not performing a compliance function. They are exercising judgment on the loop. That is stewardship at the execution altitude.

At F2 the posture is instrumentation. The manager who updates the Record after the decision shifts, who reads the signal before the incident, who asks whether the cadence is still matched to the rate of change underneath it. That is stewardship at the control altitude.

At F3 the posture is portfolio thinking. The VP who asks whether the experiment portfolio is generating learning that compounds or just generating activity. Who requires the ADR before the agent scales. Who holds the memory continuity layer as a management responsibility, not an IT one. That is stewardship at the management altitude.

At F4 the posture is fidelity. The C-suite that asks not just what the summary says but what it may have smoothed. That insists the Spine carries real signal, not curated signal. That owns the question, “Are we still right about what matters?” That is stewardship at the intent altitude.

None of these require a new function. None require a Chief AI Stewardship Officer. None require a governance committee standing up a module. They require a posture. Adopted. Practiced. Maintained without a terminal state.

AI stewardship is the visible signature of the operating layer. The layer itself is invisible when it works. No breach, no drift, no incident. Success looks like nothing happening. Stewardship is what makes the layer legible. It is the observable behavior of people exercising Fit at runtime. Each act is attributable. The pattern is the signature.

You do not declare the operating layer complete. You practice its stewardship. That practice is wearable at every floor, by anyone willing to adopt the posture. It does not require permission. It requires presence.

Share

Without end

The layer does not complete. It is the ongoing fit between what the enterprise is here to do and the form it currently runs.

Function does not change on its own schedule. Form does not hold without tending. The inversion did not create a problem to solve and move past. It created a condition to manage without end. That is not a burden. It is the nature of any system that operates in a moving world, and it is the opportunity we share now.

The six primitives are not a checklist. Intent and Policy set the direction and the boundaries. Roles and Records bind authority to judgment and carry that judgment forward in time. Cadence and Signal keep the layer alive between reviews, catching drift before it aliases into apparent stability. Together they are the substrate. The operating layer running.

The Stewardship Spine runs vertically through every floor. Intent flows down. Telemetry flows up. The stewardship signature runs as a continuous thread, not a phase, not a module, not a function standing up and standing down. It is the posture the enterprise maintains as long as it operates.

We began this arc with a feeling. Something had inverted. The outside world had changed faster than the inside could follow. The binding constraint had moved. Four posts later the diagnosis is complete and the repair is named.

The Hybrid Enterprise is not a destination. It is a practice. The floors are core functions. The six primitives express and compose the operating layer. The leadership posture of AI stewardship is adopted floor by floor, altitude by altitude, across functions by people who were already there.

First we must fall in love with the problem, not the solution. Subscribing to that discipline, Function informs Form. Combined, they inform Fit. We set the function and frame the form. From there the machine runs at scale and pace, inside the bounds we set. The human is the Fit.

Machine capacity pairs with humans to re-fit across each function, for each decision and each horizon. That residue is not a limitation. It is the point.

Leave a comment

Prologue

Why even name it a “hybrid” enterprise?

I landed on it as the future state but needed new words. I sensed it, then saw it. An archetype where people and agentic AI primitives work at scale and pace. Paired.

• Human: for authority, the seat, not the role. The human sings unbounded.

• AI: for capacity, to process at pace and scale under bounded conditions. The AI specs and repeats within the policy and the operating system (OS) that humans designed and framed.

Humans broadly for framing, coordinating, and overseeing. For what falls out of bounds, and the refit. The technologist's path endures: designing, building, tuning, integrating new machine capabilities that extend our own.

Humans everywhere for their judgment, especially the decision to deprecate.

So yes, you and I are mid-shift. The Hybrid Enterprise is not. It represents what we are shifting to. Once we shift, the pair is the norm. A dyad. On every floor.

We begin at “ground level,” our feet planted on the foundation (F0).

Forget what the floors usually mean. This building does not sort by rank. It sorts by horizon. Near to far. Small stakes to large.

An open house.

F0 through F4 floor by floor, we walk each of them.

How far out you have to see, and how long your calls take to land. The top floor holds the longest sightline. An enterprise architecture we can walk through, feel, and recognize by function, not job title.

F0. Foundation

Trust that it holds. The infrastructure runs, the AI primitives are wired in, and the people closest to each one carry the domain knowledge that makes each model’s output meaningful.

Down here, humans add value by not being needed in the live loop, only off it, building, tuning, and refitting the primitives the floors above run on.

Now the climb. We explore the view from each floor as we rise.

From here up, every floor pairs.

• A human holds the authority the floor requires.

• Capacity runs under it at machine pace and scale.

Capacity scales with the work. Authority stays seated; the seat is the authority, not the role.

Under every floor, bounded work runs at machine pace. The unbounded call rises to the human.

Every function is informed continuously. Intent is explicit. Durable memory captures each decision and informs the next. Not for gotchas. For refinement, point-in-time relevance, and easy referral. Supports compliance too.

The split (capacity to authority) holds all the way up. One floor at a time. The view widens as we climb.

F1. Execution

This may be the most familiar shape of work. You accept, reject, or escalate: the calls that reach you, shift after shift. Real-time focus. You orchestrate the primitives, catch what the model misses, own the call when it lands wrong.

For instance, prior authorization at a health system. The requests used to stack in a queue, worked one at a time by a nurse, each a small research project against the plan's rules. Now the primitive clears the in-bounds ones at machine pace. Approves what plainly qualifies, and brings her only the request it cannot resolve, with the rule it triggered, its recommendation, and the likely denial-and-appeal cost if she signs.

This is the floor of seconds, the nearest horizon. But the second stops being an urgent reflex. It becomes deliberation at human resolution, the estimate already on the table, known tradeoffs legible. She is making an informed call on the patient's care.

No job title here survives the decade; form and then fit flow downstream from function. The function survives. Someone meets reality the moment it arrives.

Conceptually like a pilot in their lane. Augmented, in command, one of many landing at once. Not Sully and the river. The systems fly the routine. You still own and select the approach when and where it matters.

F2. Control

You coordinate authority. You tune the thresholds, grant and revoke on the fly, call the intervention before drift becomes an incident.

Procurement across a manufacturer’s supply base. The buying agent works inside the rules it was given, reordering to meet pre-defined service levels. Then a supplier’s lead times stretch, and the agent, still within policy, starts ordering earlier and heavier. Nothing alarms. The dashboards stay green.

What moves under that in-policy decision is working capital. Cash tied up in inventory the business does not need yet. The control function is the human who reads that drift early, tightens the agent’s order authority, and intervenes before the quarter absorbs it. Grant and revoke, on the floor, in real time.

This is the floor of rhythm, your operating rhythm. Hit your stride, or things break. But it’s never just one rhythm.

The work moves at one pace. Oversight at another. Review and the board at others still, each at a different speed for good reason. Akhilesh Warik calls it a “cadence mismatch between runtime execution and organizational supervision.”

Control is managing the cadence of authority. You grant it and revoke it on the clock. Couple the clocks when work and oversight should move together. Decouple them when they should not. Let them fall out of time, and credibility is what goes first. When authority fragments under pressure, coordination fails before anyone is willing to say so.

Maryna Deundiak, PhD reads decision systems and leads decision-architecture sprints for executive cohorts. She names why so much of this surfaces now.

The air traffic controller does not fly the planes. She decides who lands, and when. Who, and when. One move.

On the next floor, more decision flows to oversee, coordinate, and control. But these now span multiple functions.

F3. Management

More complexity. More coordination. More rhythms now with lead and lag times to couple and decouple. Deliberation here too, but decisions with more constraints and some reversibility.

You manage by exception. You read nominal from off-nominal, steward the portfolio, decide what is worth a CEO’s attention. Otherwise, own it and run it.

Stewarding the portfolio includes pruning it. The management function reads when a capacity has been outrun: the market now offers a better one, or it duplicates a tower already running, and folds it down before it hardens into cost no one questions.

A services business running agents across pricing, renewals, and collections. Each performs on its own dashboard. The pricing agent holds win rates. The renewals agent holds retention. The collections agent cuts handle time.

Read alone, every tower is green. Read together, margin is thinning and days outstanding are creeping, because the discounts that won the deals are the same ones eroding the price, and no single tower shows it. Read only in the operating review, the cross-tower drift surfaces two cycles late. By then it is history, not a call anyone can still make.

This floor runs on telemetry. Signals flowing up, read together, across towers that used to never speak to each other. And not only the present.

The system harvests ambient signals and lagging indicators, and runs the near future in the background, hundreds of what-ifs at once. If the supplier slips, if demand moves above or below boundaries, even what-if we move capacity off this tower. Most never surface.

The work is learning and deciding which what-if is worth acting on, and which is worth a floor up. Many tradeoffs. Management, with machine capacity and telemetry in hand, sifts the signals that reveal new opportunities, reduce the risk profile, grow the top line, and protect the bottom line.

Mission Control watches every system at once. Not one cockpit. The whole board. Managing across functions and departments. Trust but verify.

F4. Intent

You choose the game. You set the bet. Fold, ante up, raise, or go all-in? You hold the longest horizon in the building.

You set the direction in the moments that shape the future, across horizons. You release the ideas that become the objectives. Those ideas become what everyone below is serving.

Not the next incident escalation or exception. Cross-tower management addresses and resolves those. You hold the ultimate authority and embody the decision system and leadership posture that cross-tower management subscribes to, operates, and enforces.

Is this a strategic bet, or an experiment? You decide. Whether to enter the market this year or cede it. Which platform to commit the company to, knowing the other gets stranded. You set what matters before anyone below knows they are already serving it.

The sharpest of these is the decision to end something. To deprecate a capability the floors below have built on, knowing what depends on it gets stranded. The augmentation shows you the better option has arrived. It cannot answer for pulling the legacy one out from under the people on it today. Few and final. Only the seat.

Up here the augmentation reads the world.

Continuous telemetry on the rivals, the anchor accounts, the suppliers you cannot afford to lose. It mines the other CEOs’ earnings calls for the signal before it is consensus, drafts the deck and the call script, and war-games the courses of action with a price on each. It hands you the board, fully read.

A floor down, the what-ifs are many and reversible. Here they are few and final.

What it cannot do is place the bet.

Everything it models, like the machine capacity on each floor and all the decision architecture that runs at machine pace and scale, is bounded. The commitment is not. And it cannot answer for the bet.

When the call lands wrong, someone has to face the board, the regulator, the counterparty across the table. No one accepts that answer from a model. Accountability does not transfer. So the unbounded call, the one only the seat can make, stays yours.

You own the bounded too. Responsible for everything you do, and everything you choose not to do. Both actions. The machine widens what you can see. It cannot stand behind what you decide.

And at the very top sits the thing that lets intent outlive its authors. Deepak Jha calls it the “memory continuity layer.” Judgment held across time, turnover, and changing conditions, so the decision made this quarter still reaches the execution floor two years on, long after the people who made it have moved on. Organizational memory was tribal. Now it's a live layer.

The poker floor. Kenny Rogers’ gambler, reading the table, deciding which hands to back and which to lay down.

Visible patterns, and a little math

Step back, and the four floors above the foundation (F0) are one move repeated. To see what remains, subtract what the machine now does, floor by floor:

• The output at execution.

• The routine at control.

• The aggregation at management.

• The reading of the world at intent.

What remains on each floor is not a smaller job. It is the judgment of when the work is right, plus the coordination the old bundle was quietly carrying.

New work appears too. Work that had no home before.

• Keeping intent legible.

• Catching drift early.

• Compounding judgment, not re-litigating it.

Subtract the doing on every floor, and what remains concentrates into judgment and coordination, then add the new work that grows on the structure and connects the floors.

Share

The spine

One structure runs the full height of the building, cross-cutting every floor. It is the formal encoding of what used to be absorbed by hand, made explicit at last and load-bearing.

Intent travels down, so the person meeting reality knows what the reality is for. Telemetry travels up, so control, management, and intent are set on what is actually happening, not only on what was true last quarter. AI stewardship runs alongside both, a dotted line, neither command nor report. The posture that shapes and sustains the operating layer.

Intent down. Telemetry up. Stewardship throughout.

The objection

A human on every floor, touching every call. Human in the loop (HITL) four times over, oversight stacked on oversight until it recurses. Nothing gained.

Look at the spine.

It carries intent and signal, not approvals. No floor signs off on the one below. The human is on the loop, not in it. Supervised autonomy, not stacked approvals.

The primitives clear the in-bounds work and surface only the exception, and the higher the floor, the fewer those are and the more each one costs.

And because each decision is captured and callable now, the work compounds instead of churning, and survives the turnover that used to erase it.

The recursion the skeptic fears is real in one case only: when every loop re-opens intent and decides it again. The spine is what prevents that. Intent is set up top and carried down, so the floor of seconds runs a decision instead of re-litigating it.

So four jobs at four horizons is a cascade, not a loop eating its tail. The top floor places the bet and owns it.

Nothing gained? Only if the human has stopped overriding, and fallen off the loop. That is not oversight; it is automation bias on a green dashboard. Watermelons are great on a picnic, not in your portfolio.

Carrying that intent across time, not only down the building, is its own work.

Maybe the building itself is the problem. Why staff every floor when one person can work them all?

Elena Verna names the role the “High-impact IC” and is living it: the ex-leader who runs a project end-to-end alone, now that the capacity scales to one. She has the doing-collapse exactly right.

“Cheaper to just try something than to debate it,” she writes, and the coordination layers thin. From inside one career, you cannot see the building. At scale, the horizons do not fold into one person. They distribute across the spine. The High-impact IC buys the capacity for every floor. She still cannot occupy them all at once.

That is the building. Humans on every floor, each working a different horizon, all of them on one spine. Each paired with capacity that scales with the work: built, hired, or both. You can build or buy the capacity. You cannot buy the seat.

From the engineer's chair

The engineering and coding bench felt the inversion first. Writing the code was the doing for entire professional lives, and today the cost of the doing collapsed.

A model produces good-enough code on demand now. The scarce act moved up. The work is no longer writing the function. It’s judging which output is good enough to keep. Intent and judgment, where hands on the keyboard used to be.

Then the part that surprises them. The reviews multiply, and most are not about code. They are about alignment.

Bringing a product manager into the reasoning. Bringing a stakeholder into why the system decides as it does. Or the reverse, pulling the engineer up to what the work is actually for.

The typing got cheap. The aligning did not.

Coordination is the cost that surfaces when production stops being the bottleneck. The CTO sees the same pattern across every bench at once.

Watch a real one

A four-person team sold three quarters of a million dollars of Bluetooth landline phones in six months. No electrical engineer. No compliance officer. No operations lead. Cat Goetze walked through how they did it for Shopify’s In Stock. Read it as a building, not a story.

Start on the execution floor. Josh, the CEO, diagnosed a firmware audio fault he had no training to diagnose, and carried the fix back to a manufacturer who had called it impossible. The function on this floor did not change. It reached higher. AI stood next to the practitioner as a domain coach. The practitioner stayed the one who asked the question and judged the answer.

Climb one floor. SOPs. Troubleshooting macros. A shipping QC checklist. Escalation thresholds. That is control-plane work, the process that lets the floor below run without a founder in the room. The control plane existed as output, produced on the execution floor by whoever hit the problem.

Floors are functions, not chairs. One person can stand on three of them in a morning. AI is what lets one person reach that far. It does not collapse the floors. It hides them.

Go to the top. The founder used AI to work through the hiring budget she had been avoiding, to surface revenue she had not considered, to shape a message to a supplier she could not afford to lose. Intent work. The kind that sets what everything below is for. Held in a couple of heads and their private chat histories.

Underneath all of it, the same ground. Shopify, Stripe, a few accounts. The foundation, rented by the month.

At four people the founder is the operating layer. Every floor runs through her attention and one shared thread. The building is real, compressed into a couple of heads.

That is not a flaw. It is the one regime where skipping the layer costs nothing.

It holds until it scales. Add people and the floors stop folding into one field of view. The shared thread stops being shared. The seats have to be filled, and the layer that was free now has to be built. What happens to the building when the person who held it in her head leaves the room is the next floor of this. We climb it next.

Leave a comment

Appendix: the audit stack

Decision trace. What happened, in what order. Immutable, machine-generated. Lives at F1.

Agent Decision Record (ADR). The why behind the call: inputs, alternatives, confidence, oversight. Julie Banfield’s term. Richer than a trace, and still not enough on its own; silent on tamper-evidence, retention, and automatic capture.

JCM. Judgment, aggregated. Deepak Jha’s Judgment Capital Management, reading the records as a portfolio at F3.

Stewardship. The leadership posture that surrounds the records with the system properties auditability actually requires, running the spine.

The forcing function. The mandate for auditable high-risk decisions has not gone away. The deadline has. On 7 May 2026, EU negotiators reached a provisional agreement to defer the Annex III high-risk obligations to 2 December 2027. (Council and Parliament provisional agreement, Digital Omnibus, pending formal adoption.) Auditability built as an operating-layer property during the runway, the trace and the ADR captured as the work happens, is cheap and compounding. Scrambled as a compliance project at the deadline, it is expensive and brittle.

The two loops. Human in the loop (HITL) fails twice. It recurses, oversight stacked on oversight with nowhere to climb. And the human goes numb with automation bias, unable to keep pace with the machine or match its scale. Human on the loop (HOTL) is the better fit. Human judgment paired with machine capacity, not stacked behind it. (Parasuraman & Manzey, "Complacency and Bias in Human Use of Automation: An Attentional Integration," 52 Human Factors 381 (2010).)

The gauge. Two numbers, read together, per floor. The governed ratio: how many agents run per human at this altitude. The override rate: whether the human judgment is still real.

Ratio rising, override rate healthy: the layer matured, the higher ratio is earned. Ratio rising, override rate collapsing toward zero: automation bias wearing a green dashboard.

A reading from the wild. In Project Glasswing, frontier detection found high- and critical-severity vulnerabilities faster than anyone could fix them. Finding stopped being the limit. Patching became it. Of an estimated 530 such bugs disclosed to open-source maintainers, 75 are patched. Detection scaled past what remediation, still a human job, could absorb. A green dashboard reads the find rate as capability.

The gauge reads the 455 still waiting on the human side. (Anthropic, “Project Glasswing: An initial update,” 22 May 2026. https://www.anthropic.com/research/glasswing-initial-update)

The safe ratio is relative, learned, and dynamic. Make it a target and it stops measuring anything.

A bonus dispatch, outside the Hybrid Enterprise series. The arc resumes with the floors. This one could not wait for them.

“Why don’t you just fix it?”

It’s the reasonable question. The flaws are found. The list exists. Patch the list.

Sit inside one of the orgs holding that list and the question falls apart. What’s left is overwhelm.

The queue scaled ten times over in a month. The cycle that clears it did not. Same intake. Same verification. Same sign-offs. Same people. All of it built for a trickle, now taking a flood.

Share

So the maintainers say the only thing left to say. “Slow down. We need to re-verify first.”

That is not resistance. That is a system reporting its own cadence. Time to resolution was already too slow at the old find rate. Multiply the find rate by ten and the cycle does not strain. It breaks.

The flood is not the problem. A cycle still turned by hand is.

That scene is not hypothetical. It is the open-source security world this month, after an AI system began finding critical flaws faster than the people who own them can verify and ship. But the shape is older than this month, and I have stood inside it.

I ran that work.

Years back I led enterprise security patching for a federal healthcare program. The work had a shape. Intent came down. Reduce the risk posture, these classes of vulnerability, this window. I did not set that. I received it. Then I triaged. Every critical and urgent patch available, sorted by what mattered against that intent. Then I sequenced implementation across every environment, inside a month.

That was never the whole job. There were always adjacent initiatives, do-not-disturb windows, people who could not be interrupted. So we ran a catch-all event to absorb what the standard cadence could not hold. And after, a report back into the firm. Progress, the improved posture, the exceptions, the new flaws that had surfaced mid-flight.

Then it started again. Progressive elaboration. The plan refined itself as the ground moved.

Three things ran underneath all of it. What gets fixed first. Who owns the call. What rhythm it ships against, and how the signal comes back. Intent. Roles. Cadence. Not three steps in a line. A loop that never closed, and was never meant to.

At a trickle, one capable operator can hold that loop together. I did. That is exactly why nobody saw it as a structure. It looked like competence. A person who was good at the job. Building the team, organizing the work, negotiating the windows, shipping on time, reporting up. Earning trust by showing up, again and again.

It was a person standing in for a system that did not exist.

Ten times the volume and the stand-in fails. Not because the operator got worse. Because the loop was never built to run without a hand on it. The hand was the architecture.

So the reflex is to reinforce the pieces. Better triage tooling. A cleaner ownership chart. Tighter patch SLAs. Each one is a real improvement. None of them is the fix.

The pieces were never the problem. Prioritization, ownership, cadence do not fail in isolation. They fail in relation. Triage that does not feed sequencing. Ownership that does not move at the speed of discovery. A report that lands a cycle too late to change anything. Each part can be excellent and the whole still fails.

What is missing is the thing that holds them in time with each other. The standing version of the loop I ran by hand. Always on, not stood up per incident. That is the operating layer. Not a tool. Not a team. The coordination itself, made durable.

None of this is mine to claim. The loop is the documented shape of vulnerability management. Identify, prioritize, remediate or mitigate, verify, report, and run it again, a closed cycle that adapts as new threats arrive. Treating it as anything less than continuous is the failure the field warns against.

So the federal program was not special. It was one instance of the standard cycle. And the standard cycle is now running out of room.

Two numbers tell it. The time from disclosure to exploitation has collapsed. Mandiant’s M-Trends 2026 puts mean time to exploit at roughly negative seven days. Exploited before the world is even told. Rapid7 counted exploited critical vulnerabilities more than doubling in a single year, with the median window from disclosure to known exploitation down to about five days and the old lead time gone.

Now the other number. The median time to fix a critical still runs into months, and most enterprises cannot clear more than fifteen to twenty percent of their queue in a month. The field already named the result. The Structural Backlog. A permanent state where disclosure outpaces the human capacity to remediate.

The human capacity to remediate. Not the tooling. The coordination.

The operators at fleet scale hit this wall first, and what they did about it is the tell. You cannot hand-patch millions of servers, so they don’t. They run it through automated pipelines, and at the limit they stop patching servers at all and rebuild them from clean images. Replace, not repair. They did not hire a faster version of me. They built the loop into the floor.

When volume forces the question, every serious operator lands on the same answer. The coordination has to become a system. The only choice left is whether you design it on purpose, or keep paying a person to be it.

And none of this stays inside security.

I posted a short version of this argument. The replies did something I did not expect. They finished the sentence from seats I had not mentioned.

Sharad Suri, who leads operations across multi-site healthcare, read it and said he sees the same pattern every time his world scales fast. His words, not mine. The trouble is rarely that the problems are new. It is that “the old operational cadence and decision pathways were built for lower signal volume and slower feedback loops.” Improve the visibility, and the exceptions stop appearing slowly. They become impossible to hide.

That is the security story with the security stripped out. Hospital operations. Same shape.

A founding CTO read the same post and moved the question. From “what is vulnerable” to “what actually matters and gets fixed first.” Different seat, same turn. The work is no longer finding. The work is deciding, owning, sequencing.

Neither was prompted. Neither used my words. An operations leader and a founding CTO, different altitudes, arrived at the same reframe.

That convergence is the evidence. When the shape shows up in vulnerability queues, in hospital operations, and in a founder’s architecture, it is not a security problem in different clothes. It is a coordination problem. And coordination does not care which industry it is starving.

So the repair travels too. Not a security playbook. Not a healthcare one. The same standing layer, wherever signal outpaces the pathways built for less.

So build the layer. Make the loop a system. Put the engine where the hand used to be.

Then comes the part the automation cannot do for you.

A system runs the loop. It does not decide what the loop is for. It will sequence whatever you point it at, faster than any of us, and keep pointing there long after the ground has moved. An engine has no opinion about what matters. It only turns.

So the hand comes off the crank, but it does not go in a pocket. It goes to the design. The layer is never finished. The domain keeps throwing new bounded cases, and every one is a judgment. Widen this loop. Retune that threshold. Stand up a new bounded loop for the case the first was never shaped to hold. And just as often, refuse to, because every new loop is another seam, and seams are the coordination cost the whole layer exists to contain.

That work has a name. Stewardship. It designs the loops, measures whether fixing fast is actually lowering risk or just clearing tickets, and refits them as the ground moves. It owns the exception the rules never foresaw, and catches the day the queue empties while the risk posture quietly slips. What it does not do is climb back in and grip the wheel.

That is the distinction, and it was never passivity. Enforcement polices the turn. Stewardship shapes the machine that turns. It informs the system and lets it run. It forces nothing and shapes everything.

The federal program worked because someone held the loop by hand. The next one works because the loop is built, and someone tends what the building cannot decide. The hand comes off the crank. The judgment stays, and goes to work.

Stewardship informs. It does not enforce.

Leave a comment

Sources

Mandiant M-Trends 2026: https://cloud.google.com/blog/topics/threat-intelligence/m-trends-2026

Verizon 2025 Data Breach Investigations Report: https://www.verizon.com/business/resources/reports/2025-dbir-data-breach-investigations-report.pdf

The Hacker News, which states the Mandiant and Verizon figures together: https://thehackernews.com/2026/05/most-remediation-programs-never-confirm.html

Rapid7 2026 Global Threat Landscape Report: https://www.rapid7.com/research/report/global-threat-landscape-report-2026/

The Structural Backlog framing and the 137-day-to-remediate against 5-day-to-exploit gap, Hexnode: https://www.hexnode.com/blogs/the-remediation-gap-why-patch-management-and-vulnerability-management-are-not-synonyms/

Monthly remediation rate of roughly one in six, NIST research via Appsecure: https://www.appsecure.security/blog/2025-vulnerability-exposure-severity-trends

Continuous-cycle mandate, CIS Critical Security Control 7: https://www.cisecurity.org/controls/continuous-vulnerability-management

Immutable infrastructure and the “replace, not repair” practice, golden images: https://www.wiz.io/academy/container-security/what-is-a-golden-image

Hybrid Enterprise arc. Part II of IV.

The captain at the console. Two seconds to approve.

The screen shows a recommendation. The system reasoned through a chain of upstream decisions the captain did not see. The captain signs. The work continues.

The signature took two seconds. The signature did not take any judgment.

Sol Rashidi senses this from a Board seat. She calls it automation bias with a human signature. The human is in the loop on paper. The human did not engage on substance. The signature satisfies the audit. The system continues to drift.

This pattern is older than AI. Managing by exception. Going numb to signals. Good enough this time, again. The new thing is not the pattern. The new thing is the pace at which the pattern compounds, and the scale at which the cost is paid.

The fix has never been a more vigilant captain. The fix is a system designed to hold the weight without the captain being heroic every two seconds. Not brittle. Not casual. Not legacy. Load-bearing for weather today, and many tomorrows. One deep on the org chart was never going to work. It works less now, faster.

Post 1 named the dissonance. The Coasean Clock named the sound. This piece names the absence under both.

Something that should be holding the weight of those decisions is not there. The signatures know it. The captains know it. Nobody has put a noun on it yet.

Many voices. Many signals. Multiplying.

Mireia Garcia Roca named the shape from one seat. AI Director at Hinojosa Packaging Group. Three years on a management committee. Fifty-plus deployments. AI Academy stood up. Recently completed an executive Master in AI. She has the standing.

She wrote about the AI Director as the least understood role. Three tensions made the post load-bearing. Mandate without authority. Startup pace inside a legacy organization. Adoption as emotional, not technical. Her keystone framing is that the role itself is the answer. CAIO, AI Director, Head of AI. Whatever the title, the role is what holds the work coherent.

She is not wrong. The view from her seat is sharp. She is also not alone.

Julie Banfield, from inside an architecture seat, is building Agent Decision Records (ADRs) as downstream-aware artifacts that do more than satisfy auditability.

Dr. Maryna Deundiak (PhD.), from a decision-architecture seat, names runtime authority and the ambiguity buffer humans used to be inside enterprises and no longer are.

Akhilesh Warik, from the execution boundary, pressure-tests admissibility, reversibility, and optionality as the runtime properties an operating layer must hold.

Chirantan Rajhans, from an engineering seat, prices drift as an agent management retainer at five hundred to two thousand a month, the agent that performed in Q1 drifting by Q3.

Five practitioners. Five vocabularies. One structural absence. All five reach for the same shape from a different seat.

The convergence is the signal. The naming is what comes next.

A shape in six pieces

The structural absence has a shape. Six pieces in three pairs. Garcia Roca's three tensions map across all three.

Intent and Policy. The codified what-we-are-here-to-do. The codified what-we-will-and-will-not-allow. Garcia Roca’s adoption-as-emotional tension lives here. Adoption stalls when no one has codified the operating intent for AI in the work. Every individual ends up improvising at the layer where the firm should be coherent. Sprawl and incoherence.

Roles and Records. Who has runtime authority. What artifacts capture the decisions made at runtime in a form the next person can act on. Garcia Roca’s mandate-without-authority tension lives here. So does Banfield’s ADR work. Records written for audit go to a drawer. Records written for downstream use stay alive. Audit and amnesia.

Cadence and Signal. The rhythm at which the system reads itself and corrects. The signals that surface drift before drift becomes incident. Rajhans’s Q1-to-Q3 drift lives here. So does Garcia Roca’s startup-pace-in-legacy-org tension. The cadence the work needs is not the cadence the legacy enterprise was built to run. Drift and incident.

Three pairs. Six primitives. None of them is software. All of them are operating discipline.

The captain at the console signed in two seconds because none of these primitives was load-bearing in their setting. The system was running. The operating layer was not.

Nine answers, one shape they share

The market waits for no one. Nine answers in the absence of a name have surfaced. Not one is wrong. Not one, or any combination, is enough.

Role-as-answer says the gap is filled by hiring a CAIO. Adoption-discipline says the gap is filled by Prosci-style change management. Maturity-ladder says the gap is closed by climbing a level. Use-case-portfolio says the gap dissolves with the right pilot mix. The full nine are catalogued as anti-patterns in a companion downloadable PDF below this article.

The shape they share is what matters here. Each is true work that real practitioners are doing well. Each is better-version work on the binding line the Coasean inversion already moved. Shahidi et al. (NBER, 2025) named this shift the Coasean Singularity. Howard Yu at IMD made it legible to practitioners.

The line between external transaction cost and internal coordination cost has moved. It will move again as trust in agentic solutions rises. Not one of the nine sits where the binding now is.

The operating layer is composable. It is built from the six primitives connecting across altitudes and to one another.

Records inform Cadence. Cadence surfaces Signal. Signal pressure-tests Policy. Policy clarifies Intent. Intent reshapes Roles.

None of the nine answers compose. Each is a freestanding partial.

Once the inversion has happened, the operating layer becomes load-bearing whether the firm has built it or not. The nine answers are individually partial. The layer is not.

The architecture has a name

The Hybrid Enterprise is the form the firm takes when classical organizational geometry holds the strategy floor and AI primitives support the work below. The planes-and-floors vocabulary owes its seed to Sanjeevan Srikrishnan, whose framing of networking as an AI primitives array at Equinix helped me describe what I sensed. The floors are connected. People continue to add value at every plane.

Five floors carry the work. F1 Execution… F2 Control… F3 Management… F4 Intent… Below F1, the Foundation carries data and infrastructure.

F1 Execution, where humans and AI primitives do the work together. F2 Control, where managers steward the local conditions. F3 Management, where skip-level leaders read across surfaces. F4 Intent, where the C-suite codifies what the firm is here to do.

Below F1, the Foundation carries data and infrastructure. Above F4, the Board Pavilion floats. Deepak Jha’s Judgment Capital Management work lives at this beat, capturing decision value at the moment of intent so the firm does not pay it again to excavate it later. He calls the result a memory continuity layer.

Humans and AI primitives are present together, serving the functional intent of each plane. Different time horizons. Different relative quantities by design. The Hybrid Enterprise multiplies and augments contributions. It does not replace them. A companion PDF below this article holds the time-horizon detail and caliber of work at each plane.

Through every plane of the Hybrid Enterprise runs the Stewardship Spine. The Spine is present. The Spine is not yet animated. That is where Post 3 picks up. We will turn on the building. Spine animated. Telemetry visible. Screens running.

Stewardship: the discipline that keeps the layer load-bearing through compounding weather, and alongside the regulatory scaffolding arriving in 2026.

The cutaway shows the building. The screens are dark. The telemetry is held. You are looking at the Hybrid Enterprise: people plus deterministic and generative systems.

The J-curve. When do we harvest?

The framework is testable.

The Brynjolfsson productivity J-curve has begun its harvest phase. Firms that shape the operating-layer primitives at design and sustain them through use will reach harvest earlier and compound farther than firms working only on the nine answers.

Two empirical anchors will read the result.

The Stanford HAI production-gap: 89% of enterprise AI agent deployments do not reach production.

The Davenport and Bean production-scale rate: 39% of Fortune 1000 firms have AI at production scale, up from under 5% two years ago.

Both will be measured again in 2028.

The framework either survives the next two years of empirical observation or it does not. The precise prediction lives in the research brief PDF below.

The Steward cohort (S) shaped the six operating-layer primitives at design and sustained them through use. The complementary intangibles the J-curve requires for harvest were already being built while AI was being deployed. S1 sits above baseline by 2026. S2 reaches harvest by 2028.

The Non-steward cohort (N) worked only on the nine answers. Tools, roles, pilots, governance modules. Each is better-version work on the binding line the Coasean inversion already moved. The complementary intangibles that compound never get built. N1 sits at the trough in 2026. N2 is still below baseline in 2028.

The 2028 re-measurements will read the cohorts back onto the curve. The framework either survives or it does not.

Sources: Brynjolfsson, Rock, and Syverson, “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies” (American Economic Journal: Macroeconomics, 2021). Stanford HAI 2026 AI Index. Davenport and Bean, MIT Sloan Management Review (2026).

An aside on intangibles (for the CTO, and skeptics)

For the technologist reader who has weathered enough waves to be skeptical of the intangible kind.

Deming estimated 94% of variation in performance comes from the system, not the worker. Out of the Crisis (1986), page 315. Forty years on the shelf.

PwC’s 2026 AI Performance Study surveyed 1,217 executives across 25 sectors. 74% of AI’s economic value flows to 20% of organizations. The discriminator named in the same study: workflow redesign over tool layering. Technology delivers 20% of an initiative’s value. Work redesign delivers the other 80%.

Forty years apart. Same finding. Different vocabulary. Deming called it management’s responsibility. PwC calls it work redesign. McElheran and Brynjolfsson call it structured production-management practices. The operating layer is what they have all been naming.

The operating layer is the 20% of effort that compounds 80% of the value AI offers. It is not a product or artifact to purchase. That is not a reason to discount it. That is the reason it is load-bearing. It is alive. It is continuously active. It is lived.

The framework was synthesized from lived practitioner experience first. Pressure-testing against the academic literature surfaced nine independent research pieces that lend cross-cutting support.

Monday morning

This piece arrived at the cutaway treatment because many voices have been pointing at the shape, including some not yet named.

Sanjeevan Srikrishnan named networking as an AI primitives array at Equinix.

Jacob Lauritzen of Legora named and illustrated the new shape of work within firms at AI Engineer Europe: more planning and reviewing, less doing.

Each pointed me toward the same North Star.

The Coasean Clock ticks.

The cutaway connects what was familiar to what is different. For me, and I hope for others. So we can chart our next course better informed, and contribute new value to our collective professional experience, and our personal lives.

The self-diagnostic is one question. Which primitive pair is the binding gap in your setting right now?

If meetings keep ending without anyone able to say who decided, look at Roles and Records. If the system found a problem in March that nobody acted on until June, look at Cadence and Signal. If your AI initiative has executive air cover and frontline confusion, look at Intent and Policy.

Pick the pair. Name the gap. Write it down where someone else can read it.

That is what makes the building load-bearing.

Shape at design is the half this piece named. Sustain through use is the half Post 3 picks up. The Spine animates, the telemetry rises, stewardship becomes the discipline visible at every floor.

Composable, yes. Optional, not anymore.

Leave a comment

PDF and PNG files that travel with this piece

Downloadable artifacts sit beneath this piece. None is required reading. Each one is for a different kind of reader.

Research Brief Many Signals Converging V1 3

541KB ∙ PDF file

Download

Download

• Title: Research Brief, Many Signals Converging

• Description: The independent academic, government, commercial, and practitioner sources behind this post, organized by plane. Several pages of methodology and synthesis, with the interferometric diagram embedded.

Nine Anti Patterns V1 2

131KB ∙ PDF file

Download

Download

• Title: Nine Anti-Patterns

• Description: The common enterprise AI responses, diagnosed. Each one is better-version work on a binding line the Coasean inversion already moved.

Humans On Every Floor V1

118KB ∙ PDF file

Download

Download

• Title: Humans On Every Floor

• Description: What humans do at each altitude in the Hybrid Enterprise. Practitioner through Board Pavilion. With characters most readers know.

Next (below) are two versions of the Interferometric Diagram

v1.7a (bare, cleaner visual), and

• v1.7b (much busier, including plane specific references horizontally, along with cross-cutting independent references vertically)

Fall in love with the problem, not the solution. Uri Levine, Waze founder, put it plainly.

It’s timeless. Design Thinking innovator Carlos Solana, who champions technology transfer and teaches in Barcelona, holds it as a discipline.

Examine the problem, resist the confirmation bias that might urge us to accept an ill-fitting or partial solution.

W. Edwards Deming framed the management version of it in Out of the Crisis (MIT Press, 1986). It is still aging well.

Management’s job is not to push people through a process. Management’s job is to design the system inside which people work.

The environment. Each firm is its own system-of-systems, the inside complement to product-market fit. Each enterprise operates one of two ways: coherently with fewer seams and fewer human APIs, or heroically with more effort and cognitive translation; humans buffering what the system-of-systems left incomplete.

How do we know? Microsoft Excel and Google Sheets. For three decades, the requisite skill for ICs and platform owners across the enterprise. The translation pad.

Leaders design for the market externally and for the firm internally. People do what the system rewards, whether the reward signal makes sense on its surface or not.

A dual stance: outside-in and inside-out. Both must reject premature solutions, even well established ones. Forty years apart. Still visible. A North Star.

Share

Executives place bets. Operators deliver transformations. Both love the solution and the story after. The solution is what they got hired for, promoted for, paid for. A transformation. The level-up story.

The heroics were real and the credit for artful execution is well earned. The reward signal, real. Loyalty and pride, both earned. Credibility, lifted for both parties.

Neither loyalty nor pride in a solution is bad. But another artifact, another deliverable, a new org chart taxonomy is not what’s required now.

New capabilities have already arrived. What was scarce is becoming widely available at acceptable quality.

The exchanges, across the market and within the firm, shift and re-value. So do the promises that hold them together.

Firm to customer. Firm to employee. Employee to customer. What we bring to market changes as much as what we bring to employers.

Development capability is no longer in short supply. New specialized skills land every week. Some multiply what we can do.

That’s how the Coasean Singularity shows up. Professor Howard Yu of IMD is making the term legible. The Coasean Clock is what ticks inside it. That intangible angst people continue to talk about after work, at the ballgames, each morning in this season.

It's audible outside the firm. The deep vibrations come through the walls. The weather has shifted. We can ignore it, “This too shall pass.” Or, we can face it and go deeper.

Ostrich isn’t a yoga pose, and it’s not a leadership posture either.

Some operators and firms feel this as a threat to relative competitiveness. Brake or accelerate into the turn?

Deming’s moment all over again, with a twist.

The 1986 alternative was higher-quality cars. The 2026 alternative is good-enough AI capability arriving from the market two ways at once.

Inorganically, through custom skills, Gems, agents, third-party tools. Organically, from the native LLM providers themselves (Anthropic, OpenAI, Google, hundreds more at the next tier).

External alternatives and internal capabilities don’t line up cleanly. The misalignment is the new bottleneck. Even when we can’t see the systemic seams at every level from our seat, we already see the shape of work has changed.

A fair skeptic will hold that AI lowers production cost, not transaction cost. Drafting, coding, summarizing, iterating. She is right about the distinction. Search, negotiation, monitoring, enforcement. Those are what move the firm boundary, and those depend on agents that act, not assistants that suggest. When trusted, coordination costs collapse.

The Clock is ticking, not striking. None of which changes what's already in production. The misalignment between deterministic and non-deterministic systems is observable today, regardless of when the Singularity arrives or whether it arrives at all.

Look closer. There's a second thing. A crisis of operational coherence. The systems that delivered on the firm's promises, to customers, to employees, between them, can't carry the next round.

Attention! Levine's line gives you the disposition. Deming's mandate gives you the work. Neither tells us what to design as we meet, use, and sustain emerging generative AI and agentic capabilities.

The Operating Layer is what you design and steward.

Not a deliverable. Not a maturity ladder. Not a module bolted onto an existing PMO. It's where Levine's disposition and Deming's mandate become something you can run on Monday morning.

If you have loved a hero, or a solution that served you well, that’s fine. If you have loved a way of working that the market rewarded for a decade or three, that’s fine too.

None of our fondness gets discounted. The point is not to shame a seat, a solution, or a hero.

Management answers for what it designs and for what it ignores. Reorgs might be helpful, but alone they are superficial, like repainting the lobby. Nothing new here. Carry on. Go back to work.

The next round won't be all-deterministic or all-generative. It will be hybrid, and it will need an operating layer that holds both. A few will lean into the challenge (the opportunity) before certainty arrives. Data accumulates. New insights and labels arrive. The rhythm normalizes.

The work becomes routine. Tick-tock.

Here’s a hint from a mentor of my own: “Steady pace, wins the race.” - Neal Huysman

Levine gives you the disposition. Deming gives you the mandate. The floor is yours to build, with whoever’s doing the work alongside. The Coasean Clock keeps ticking. Your work picks up.

Subscribe now

Hybrid Enterprise arc. Part I of IV.

Something feels off inside the enterprise right now. Not broken, exactly. Just misaligned. The org chart has changed a few times. So have the tools. The tools are better than they’ve ever been. But getting to done feels harder, not easier. Constant heroics. So much in flux every month, every week.

You’re not imagining it.

What you’re feeling is an information architecture problem at the enterprise level. Among incumbents and start-ups for certain. Unicorns too, I expect, although I don’t have direct experience there. Yet.

The nav system of the enterprise, the structures, signals and ways of working that help associates find things, make decisions, and move from “I need this...” to “It’s done.” That enterprise system was designed decades ago in an environment that no longer exists.

Writer’s note: My favorite rollercoaster of all time was The Beast at Kings Island (now part of the Six Flags family). You can Google it; a good one! I once saved up for a summer pass and spent an entire day riding only The Beast over and over from the moment the park opened to close. It began with that clack-clack-clack climb up the first, biggest, steepest hill, then over the top into the big drop, then racing through the rest of the course. That’s what I want this series to feel like. This post is the slow climb. Bear with me on the concepts and signals. The payoff (the thesis, the synthesis, the hypothesis) is coming. And I’d love your help pushing on these ideas along the way.

Let me borrow three concepts from a discipline most people associate with websites, mobile apps, public libraries, not corporations. They’re more useful than they look.

1. Taxonomy is a classification system built on hierarchy. It answers one question: where does this belong? Parent, child, sibling. Every item sorted into a branch of a tree. Mutually exclusive at each level.

The firm’s org chart is its taxonomy. The folder structure in SharePoint, the chart of accounts, the JIRA project hierarchy are all taxonomies. They are clean. They are also, by definition, incomplete. A taxonomy tells you which drawer to open: top, middle or bottom. It does not define the things inside the drawer, nor does it convey the relationship to the next thing in the next drawer over.

2. Ontology defines what things are, what properties they have, and how they relate to each other. Not just vertically, but across every boundary the taxonomy draws. An ontology is a web, not a tree. It captures meaning, not just position.

In an enterprise, the ontology is the 𝘰𝘱𝘦𝘳𝘢𝘵𝘪𝘯𝘨 𝘮𝘰𝘥𝘦𝘭 in practice. What the firm’s terms mean, how its entities relate, and the processes that run against that map. Not the org chart, but the actual network of relationships, dependencies, rules, and influence patterns that determine how work gets done. Who has authority over what. What counts as an approval. When an exception becomes a precedent. How a “customer” in one system maps (or doesn’t) to a “customer” in another. Most of this lives as tribal knowledge. It was never written down because it didn’t need to be. People carried it. People negotiated it in conversation. People were the buffer. We were the human APIs. We navigated the unwritten expectations, protocols and rules, and absorbed them as the way things work.

3. Information architecture (IA) is the design discipline that encompasses both taxonomy and ontology, plus navigation, labeling, search, and the overall structure that helps people find and understand things. It is the practice of making a complex body of information usable.

Day to day within the enterprise, IA is the felt structure of the employee’s day. The onboarding flow that teaches you who to talk to. The Slack channel naming convention that tells you where the real conversation lives. The unwritten rule about which Confluence space, SharePoint folder or Teams Channel is actually current and still in use. It’s usually terrible, because nobody designed it. It simply accreted.

These three layers have coexisted in tension for decades. And for decades, it worked well enough. The taxonomy was maintained. The ontology was tribal. The IA was accidental. And humans bridged the gaps.

Then the cost and value structure shifted across markets, not only one or two large firms’ customer bases. What IMD Professor Howard Yu made legible, “the Coasean Singularity,” is the inversion of an old equation. Most of the twentieth century and this early twenty-first, the cost of coordinating work inside a firm was lower than the cost of coordinating it outside. Cheaper to build inside than to buy outside the firm. That justified the hierarchy. That justified the headcount. That justified the org chart and operating model. Once established, little changed, however imperfect, however tribal. And customers returned, were loyal, so long as “good enough” was not widely available elsewhere.

AI is collapsing external coordination costs below internal coordination costs, not only development costs. And when that happens, the enterprise’s information architecture, the nav system designed for the old stable environment, stops making sense. You are walking hallways designed for a building whose floor plan just changed underneath you. Inception, but at work. What was down is now up.

This is the disorientation feeling. Not a skills gap. Not a technology gap. Technology adoption and enablement are not enough, not even with OCM to promote the change continuously with leader-led behavior. The gap is structural. The firm is still organized for how work used to flow. The work has changed shape and moved on.

It gets worse when you add the trust dimension.

Deterministic systems hold trust passively. You deploy an ERP, CRM or ITSM solution (SaaS or on-prem). It behaves the same way every time. Or it did, until about 2022. Confidence persisted because the output was predictable. The operating model was designed for that world. Design it, implement it, revisit it during a reorg. Trust the blueprint before, during and after.

Non-deterministic systems behave differently. They consume trust actively, and over time. Every interaction is a small withdrawal or deposit. The model drifts. Context shifts. User behavior evolves. A response that felt appropriate six months ago becomes clearly wrong, or worse, tone-deaf. And this is happening on both sides of the house simultaneously. Customers experiencing AI-informed products and services externally. Associates relying on AI-assisted workflows internally. Trust degrading on two fronts if nobody is watching. We can expect this from non-deterministic systems. Let’s own it and organize for it proactively.

The operating model cannot catch this. It is a snapshot. Another artifact. Static like the org chart. It describes the arrangement at design time. What we’re describing is a 𝘳𝘶𝘯𝘵𝘪𝘮𝘦 𝘱𝘳𝘰𝘣𝘭𝘦𝘮. Brand confidence, associate trust, customer experience quality. These are continuous signals. They require continuous stewardship. Not a quarterly review. Not an annual audit. A running discipline that detects drift before it becomes damage.

This is why the enterprise needs something it doesn’t have yet.

Not a better org chart. That’s taxonomy. Necessary, but it was never sufficient.

Not a better 𝘰𝘱𝘦𝘳𝘢𝘵𝘪𝘯𝘨 𝘮𝘰𝘥𝘦𝘭 That’s ontology committed to a slide deck. Useful at design time. Silent at runtime.

Not a better tool. Tools operate inside the existing architecture. If the architecture is misaligned, better tools just accelerate in the wrong direction.

What the enterprise needs is an 𝘰𝘱𝘦𝘳𝘢𝘵𝘪𝘯𝘨 𝘭𝘢𝘺𝘦𝘳. A running system, not a blueprint. A living discipline that governs how intent becomes action and how the organization knows whether it’s working. Something that keeps the business model’s promise to customers and the operating model’s promise to associates honest over time, in a world where the systems underneath them no longer behave the same way twice.

The operating layer is what the operating model was supposed to become and never did. Because until now, the human buffer absorbed the gap. Humans navigated ambiguity in conversation. Carried the tribal knowledge. Translated between systems that defined “customer” differently.

AI removed the buffer. The gap is visible now.

AI surfaced our ontological technical debt.

The question is not whether your enterprise has this debt. It does. Every enterprise does. The question is whether you’re going to address it with another blueprint, or build something that actually runs.

That’s what this series is about. Not the technology. The layer underneath it. The one that was always missing.

Next: what the operating layer is made of and where AI primitives and humans each fit. Not a replacement story. A hybrid one. Where humans continue to show up, add value, and lead at every altitude of the enterprise. Where AI primitives handle what they’re built for. And how the two pair, working together, inside something I’m calling the 𝘏𝘺𝘣𝘳𝘪𝘥 𝘌𝘯𝘵𝘦𝘳𝘱𝘳𝘪𝘴𝘦 .

Post 2 (next) names the shape. Post 3 climbs its floors, and Post 4 lays out the mechanism that keeps it running. Now we ride over the top into the big drop.

Subscribe now

The Best Solution Architect

The best Solution Architect I ever worked with conceived and shipped a working prototype in an afternoon.

It took another three weeks for the client and team to agree on what they had built, and elicit more of what would make it stick and still play well with others: adjacent groups, practices, and third-party solutions.

Now countless devs, engineers, and vibe coders can also deliver prototypes in an afternoon. But do they stick the landing?

I recently commented under a post by Lenny Rachitsky that the bottleneck is not just getting six people to agree. It is keeping them aligned after the decision ships.

AI accelerates many tasks, but it cannot read our minds when someone asks "Does this still hold? Is it still relevant? Can we do better? Should we still pay a premium for this from vendor Z if we can now do that for ourselves?"

That is not a kickoff problem. That is not another brief or artifact. It is an operating rhythm challenge. Most organizations have not found their rhythm yet.

I also published six questions about the AI components in their portfolio that most enterprise operating models cannot answer yet.

Questions like:

Who owns whether an AI solution is still delivering value six months after launch? Who looks for the signals when two AI solutions should consolidate into one? And when user behavior shifts or the market does again, who's keeping tabs on that?

The response confirmed what I suspected.

The gap is not knowledge. It is that no one owns the recurring practice of asking and bringing a consistent frame.

The hybrid enterprise

Has a manual transmission with three gears

Today's enterprise is a hybrid: humans and bots (or agentic solutions, if you will) working in concert, sometimes seamlessly, sometimes less so.

There are three gears turning inside every organization adopting AI, and they are turning simultaneously, but always at different speeds over the same ground.

Share

Gear #1:

Build velocity has gone from days to hours. Tools like Claude Code and agentic development environments let a developer produce working software in hours that would have taken a cross-functional team one or several iterations.

Executives notice this. They get excited. Understandably.

Gear #2:

Alignment speed is harder to pin down. Sometimes a team clicks with a single conversation. Other times it takes weeks of demos, feedback, revisions, and do-overs before there is consensus on what they are shipping, and when it is actually ready to ship.

Executives are still human.

Because development is now so much faster, we want it by Friday! Frequently we don’t get that. We still have to confront our own ability (or inability) to describe what we want clearly enough for anyone, human or bot, to deliver it on the first pass.

Progressive elaboration describes how we actually work: describe a thing, develop a thing, learn more about a thing, then share emerging insights to better define that-same-thing, and go again.

From beginning to end our understanding and ability to express what we actually require grows over time. That’s a function of us - human communication and learning.

Expectations also rise steadily over time. Say it, see it, say it better, have an ah-ha moment, then land on what is truly impactful. New value.

AI and agentic solutions have not collapsed that. The constraint on speed to outcome is not development. It is us.

Gear #3:

Judgment is best exercised when we have sufficient context. Not certainty but enough. Understanding whether a deployed AI capability is actually delivering value, whether the ground around it has shifted, whether the reasoning behind the original investment still holds.

One month of data is a blip.

We require strong and weak signals to connect. Over a quarter patterns become evident. There is a story now.

With a manual transmission in an auto we use the stick shift to engage one gear at a time. It requires skill. One that we can develop with practice and wield expertly.

Across the enterprise, all three gears (Build, Alignment, Judgment) are turning at once, but at different speeds for good reason.

Many organizations are indexing their expectations and commitments casually to build speed. That is a leap too far.

Build speed is good for rapid development, but iterations are still valuable and have their place in the Alignment gear.

AI stewardship is the clutch: the deliberate act, and acquired skill, that intentionally and selectively brings Build, Alignment, and ultimately Judgment to bear on the enterprise's power train.

There is an order of operations to AI stewardship in this season of change. Those that find their rhythm will go farther and fare better.

A simple diagnostic

A recent conversation catalyzed by IMD Professor Howard Yu made the Coasean Singularity legible in this AI market moment. Executive practitioners and Business thought leaders including Professor Stefan Michel (IMD Dean of Faculty & Research), Deepak Jha (Founder & CEO, Quantum Mosaic), Boy Kester (Founder, DecisionDNA and AI-scaling Advisor), and Juan Jauregui Becker, PhD. (Expert in Design Engineering, Robotics, Automation and Transformation Leadership) helped me see a pattern and frame a response. Becker also extended the control-theory training maps directly to the KPIs-as-signals, managers-as-controllers, processes-as-plant framing.

(Edit: paragraph and hyperlinks updated May 15, 2026 for readability and credit.)

Helped me see a pattern and frame a response.

One Inch Ahead

Coase vs. Claude and The Future of the Firm

In 1931, a twenty-year-old commerce student at the London School of Economics received a modest traveling scholarship and boarded a ship to the United States. His name was Ronald Coase. He had worn leg braces as a child. Economics had found him almost by accident, in a seminar where a professor unpacked Adam Smith’s theory of the invisible hand…

Read more

4 months ago · 210 likes · 92 comments · Howard Yu

Perhaps the sharpest two-step diagnostic:

Before you can design the operating rhythm, you need to know where you stand. Two questions frame it.

First: how much decision-making have you delegated to AI?

On one end, deterministic systems that execute rules humans defined. Inputs lead to uniform, repeatable outputs.

On the other, probabilistic, agentic solutions interpret context to respond meaningfully but may not arrive at the same result twice.

Agentic solutions are fundamentally different from almost everything enterprise IT has supported and governed before.

Given:

Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks. […]

When more complexity is warranted, workflows offer predictability and consistency for well-defined tasks, whereas agents are the better option when flexibility and model-driven decision-making are needed at scale. - Anthropic.

Engineering-wise:

Treating the Large Language Model (LLM) not as a chatbot, but as a nondeterministic kernel process wrapped in a deterministic runtime environment. - [Praetorian].

(Edit: Author’s additional reference added May 15, 2026 to be precise about the nature of the problem, when a system of systems - a firm - contains both, we must design and sustain them over time)

Operator’s Thesis:

The operating layer must continually address both challenges, at inception (design) and over time (sustain).

Enterprise software until recently was broadly deterministic. Testing strategies could be linear or simple math: addition and subtraction. Same input, same output. Consistency.

LLMs are probabilistic. Generative. Non-deterministic.

To incorporate flexible decision making (not only if-then, and/or/not), we constrain where and how the non-deterministic capability gets to decide. Safely. Confidently.

Sprawling, unbounded agentic solutions increase exposure. They risk tone deafness as much as non-compliance. Unacceptable. At scale, prohibitively costly.

By Design: We box the agentic capability inside a deterministic envelope so the decisions land under acceptable conditions.

This is the Coasean inversion at the engineering layer.

That grants us confidence to operate with the benefit of automated tasks and workflows leveraging AI decision making at pace and scale within bounds we define. A tolerance. And a risk profile we can accept and defend, personally and professionally.

Without a recurring practice to track and monitor how agentic solutions behave in production over time, trust erodes quietly at first, then loudly. Long before the next annual governing audit or exercise.

Therefore: We take a new leadership posture. We commit to stewardship.

Continuously looking outside the firm and in, balancing Product Market Fit (PMF) with operational coherence.

Stewardship preserves trust with customers and associates. It shows as market coherence outside the firm and operational coherence inside.

Second: does someone in your organization have the authority and skill to examine AI activity across the enterprise, both official and shadow IT? And do the work of making sense of how it holds together?

To earn a "yes" you need both coherence and visibility.

To earn a "no," either gap is sufficient.

Most firms today are operating in the Exposed quadrant (bottom-right).

Legend:

Stable: Traditional governance fits. No stewardship gap yet, because all enterprise software is deterministic, no AI. Basically circa 2022 and prior. (top-left quadrant)

Stewardship: Agents have significant delegated authority within a designed operating rhythm. Value compounds from regular inspection, informed decision-making, and growing coherence. (top-right quadrant)

Brittle: Functioning today. One disruption away from breaking. Nobody is assembling the full picture or bringing it to leadership consistently. (bottom-left quadrant)

Exposed: Agentic solutions proliferating officially and unofficially. No one is empowered to consistently ask and synthesize a response: do the portfolio components where AI is in play make sense together? Is the strategy the agents are executing still relevant to us?

I call the Exposed quadrant (bottom-right) the race to irrelevance. Within the firm it feels disconnected, like pedaling a bike after the chain falls off.

The effort is genuine.

The execution can be flawless.

But the connection to forward motion has been severed. The pedaling looks productive.

IMD’s Professor Yu highlights the urgency. Without traction the enterprise is adrift, as the market reshapes.

The test for where your organization sits can be settled in a single meeting: Can someone satisfactorily characterize the portfolio components where AI is in play and offer a coherent narrative about how it fits together?

If yes, top row.

If not, bottom row.

And if the answer is ‘I don’t know who would do that,’ read on.

The operating rhythm

AI stewardship is not a framework on a shelf. It is the act of having fundamental conversations. The monthly and quarterly cadences do not produce governance as their primary output. They produce shared understanding.

The recommendations, the scorecards, the briefs: those are just artifacts.

The real value is that people sat in a room, looked at the same evidence, gained insight and talked about what it means in this moment for the next, and the target outcomes on each subsequent horizon committed to.

The marketplace is confronting the enterprise with change faster than it can be interpreted and absorbed comfortably. Faster, in many cases, than the enterprise's own SaaS providers can keep up.

Much of the AI capability bolted onto SaaS products in 2025 is already being superseded by what LLM providers shipped natively in Q1 of 2026.

Ouch.

Not having those conversations regularly?

The signals pile up unprocessed and leadership acts on potentially expired assumptions, so more-of-the-same momentum becomes expensive and throwaway work.

Monthly, at the operational layer. Technical Product Managers list the capabilities each AI solution unlocks, and communicate relative performance along a baseline frame.

Not individual agent decisions, but the patterns: is this capability still aligned with intent? Has context shifted? What is approaching deployment that may need stewardship attention, support, or guidance before it ships, not after?

Not all agent outputs deserve the same scrutiny. Calibrate the lens to the stakes.

Applying board-level rigor to routine meeting summaries is governance theater. Applying casual standards to an Earnings Call script or board financials is negligence.

Upstream assumptions, decisions, tradeoffs, and alternatives get surfaced and preserved as a reference point. Not for future blame. For current basis in light of new information, market movement, and sharpening intent. So the organization can easily and effectively revisit next quarter and stay the course, tune, or re-assess and respond to the market.

Monthly enables and respects the investment in learning, communication, and growing alignment.

Quarterly, at the portfolio layer. This does not need to be a new meeting. It can layer onto an existing strategic portfolio rhythm as a standing segment. It’s familiar.

Someone with portfolio-level visibility assembles findings and presents to the CEO, CTO, and CFO. Not a committee. A person with the mandate to see across all AI activity and the authority to access, inspect, and baseline it.

They synthesize and offer relatable recommendations directly to the leaders who set strategy, enable it technically, and fund it. They may enlist SMEs to clarify specific dimensions of particular use cases and emerging (or aging) solutions.

The standing open questions are strategic: What do we fund, keep, uplift, combine, deprecate, or subscribe to? Has the make-or-buy math shifted? Are we paying SaaS premiums for capabilities the market now presents to us natively? Are we honoring our core strategy, or diluting it?

The quarterly AI Stewardship review serves as a dual forcing function.

The learning frame asks "What do we know now that we did not know before? What does that mean to us? Have we unlocked a competitive capability or achieved more leverage?" Disarming for honest operators.

The evidence frame is where a balanced scorecard makes capability performance visible across consistent dimensions. Constraining for political operators who would rather control the narrative than present the data.

The learning frame is generative. The evidence frame raises the cost of politics. Both operate in the same room.

Deepak Jha's work on Judgment Capital Management sharpens this further.

His Board Test offers standing questions of every consequential decision: “What was decided? Why? What alternatives were rejected?”

The gap between what organizations record and the reasoning they actually retain is what Jha calls Institutional Amnesia. It will cost eight premium executive hours of C-suite decision reconstruction next quarter, or it could be avoided in ten minutes today.

Quarterly promotes disciplined transparency and accountability.

The window is wide enough to distinguish signal from noise, revisit the basis of standing bets, track progress against multiple benefit horizons, and guard against drift before trust erodes and opportunities expire.

Subscribe now

Not every enterprise needs this

The need for formal AI stewardship follows a barbell distribution.

At the hyper-niche end, two-person AI-native startups, the founder is the governance. Stewardship cadences would be overhead.

At the hyper-scale end, large enterprises, regulated industries, portfolio companies, stewardship is essential.

The market's middle is collapsing. That’s where the importance and urgency are most acute.

These organizations are being pushed toward commodity pricing by the rapid emergence of good-enough alternatives monthly, and would prefer to identify and earn a premium where the market values and rewards coherence.

AI stewardship enables CxO leadership to not only bet on a strategy that maximizes their ability to protect value and generate new, but to chart a course and tack to it as the market shifts.

Most everything is clearer in hindsight.

The internet collapsed distribution.

Mobile collapsed access.

When the consumer-grade interface arrives for AI, the scope of change will be genuinely broader. It will collapse the middle of how work itself gets organized and delivered, faster and wider than anything most of us have seen in our working lives.

Stewardship is the clutch

The instinct in most organizations is to remove friction everywhere possible. AI is good at that. Remarkably good.

But the enterprise still has a manual transmission and runs best when we operate its stick shift skillfully. AI stewardship is our clutch.

It does not slow us down. It preserves momentum achieved in one gear to compound in the next gear. It holds space for smooth transitions so timing and rhythm meet and serve us better where we are.

Without it, we get stuck in one gear or grind a couple gears and get frustrated. The memories of grinding gears as a student-driver still make me cringe. Not a good experience for a firm either.

A well-designed operating rhythm lets the enterprise shift between Build, Alignment, and Judgment without grinding. It connects the engine's power to the wheels' traction through a deliberate human act.

AI stewardship creates the moments where that designed friction brings visibility and recommendations to leadership so they can act accordingly: tune strategy for market relevance and action across the enterprise in a coherent manner instead of incoherent.

The organizations that compound value with AI do not need to be the fastest builders. They will be the ones that recognize a little friction by design promotes their ability to find and validate external market relevance. And creates regular opportunities to level-up internal coherence.

The constraint is not development. It is us.

Leave a comment

If your organization is addressing this, I would value your comment or a conversation. The discipline gets sharper with more of us in the conversation.

I am Allan “Al” Dean, a Staff Technical Program Manager with 20+ years across enterprise strategy, operations, and program delivery. I write about AI stewardship, operating models, and leadership. I try not to learn everything the hard way, and I enjoy helping others do the same.

I'm Allan Dean. I lead across the executive table.
I've spent twenty-plus years in enterprise strategy, operations, and program delivery: building PMOs from scratch, recovering programs others wrote off. Before that I led a tank platoon and coordinated forward-deployed task force operations as a U.S. Army captain.
The through line is the same question: what holds under pressure.

Why this, why now

The enterprise AI conversation has moved past what to build. The harder question is what happens after it ships? Or after many things have shipped already?

Who owns whether they still deliver value? Who assembles a full snapshot across the enterprise? Who cascades that to leadership with enough signal, consistency, and a relatable narrative?

I’ve been calling that discipline AI stewardship. Not governance as a framework on a shelf. A recurring operating rhythm that enables leadership to keep enterprises coherent internally at speed.

What to expect here

How organizations stay coherent while moving fast.

How designed friction beats designed enforcement.

How the constraint on opportunity and growth is no longer development. It’s us.

I publish when something is ready, not on a schedule. Quality over cadence.

My first essay drops next. It’s called “A Little Friction by Design.” Now in this era of AI and hybrid enterprises composed of people and bots, if you’ve witnessed a good team build much, much faster yet struggle to stick the landing, this next one’s for you.