Personally Faster, Collectively Unchanged

The Holiday Test

Anna is on holiday for two weeks. What happens to the work she does with AI?

If the answer is it waits, you have personal productivity.

If the answer is someone else runs her skill, you have shared personal productivity.

If the answer is nothing happens — it just runs, and she looks at the exceptions when she gets back, you have a process.

Most organisations are firmly in the middle answer and believe they are in the third.

The data says the same thing at scale. McKinsey's 2026 State of AI survey, fielded in May and June among 1,719 leaders, found that 80 percent of respondents say AI has improved their individual productivity. Half say it helps them make better decisions. And 37 percent report any EBIT contribution at all — essentially unchanged from the year before. The share of AI high performers sits flat at about 6 percent. ^[1]

Over the same period, the share of large organisations scaling AI agents rose from 27 to 40 percent. More agents, more scaling, more spend — and the same 6 percent. ^[1]

Individual productivity is not adding up. That is not a technology problem, and it is not an enthusiasm problem. It is structural, and it sits at a single, identifiable step.

Four States, One Task

Simon Green makes the operational version of this argument precisely, and his ladder is the most useful diagnostic I have seen. ^[2] Rather than define it abstractly, here it is running through one ordinary task.

Every enterprise answers security questionnaires. Customer due diligence, ISO 27001 annexes, TISAX, supplier audits. Two hundred questions arrive with a two-week deadline, and three or four people in the company actually know the answers.

State one — manual. Anna writes the answers. She is one of the few who can.

State two — assisted. Anna pastes the questions into a chat window and edits the drafts. Considerably faster. Entirely private, entirely hers.

State three — structured personal leverage. Anna builds a skill: the house phrasing, the approved wording for the awkward questions, pointers to the current certificates and evidence. She shares it. Six colleagues now use it, and the answers are noticeably more consistent.

This is a real gain, and it is where most organisations stop and declare victory.

It is still state three. Nothing starts until someone remembers to start it. It runs under a personal account, with that person's access to the evidence repository. And the quality control is that the person reading the draft happens to know what a good answer looks like.

State four — an operating capability. A questionnaire landing in the mailbox starts the process. It authenticates as itself. It drafts what it can source from approved evidence and routes everything else to an exception queue. A named person owns it. Anna reviews the exceptions.

State five — a managed one. Performance is measured, and the process is improved on purpose rather than by accident.

The jump that matters is three to four. From the inside, states three and four both feel like using AI properly. Only one of them survives Anna's holiday.

Why Three Doesn't Become Four — And Why It Has To

Here I want to be uncomfortable about a view I hold myself.

I am convinced the right way to start is bottom-up. Give people the tools. Let them build skills. Let them share them. Find the high-impact patterns in what they actually do, then operationalise those. Do not build a data platform for eighteen months before anyone is allowed to touch a model. That sequencing is a category error — it treats agentic AI as a warehouse project.

But a claim hides inside that conviction which the evidence does not support: that the rest follows naturally.

It doesn't happen on its own. If bottom-up enablement converted into operational capability by itself, the high-performer share would not be flat at 6 percent while individual productivity sits at 80. That gap is the natural outcome. It is a plateau, not a ramp.

The same finding comes out of software engineering. DORA — Google's long-running research programme into what makes development organisations fast and stable — observes that individual speed gains are frequently lost further downstream: what finishes faster at the desk then queues up in review, testing, security and deployment. Where AI adoption grows only from below and nobody picks it up, it stalls at the team boundary. AI amplifies what is already there; it does not supply what is missing. ^[3]

Sharing helps less than it looks. It genuinely removes the dependency on one particular person — if Anna is away, a colleague can run her skill. What it does not remove is the dependency on a person. Nothing starts until someone decides to start it, and quality is still whatever the reader happens to know. Six people sharing a skill have improved consistency and discovery, which is worth having. They have not created a process.

It isn't safe to depend on. When personal leverage scales, agents end up acting under human credentials. That may be perfectly reasonable for one person's productivity. It is a poor foundation for a business process.

The numbers are now hard to ignore. Non-human identities — service accounts, API keys, OAuth tokens and the credentials AI agents wield — outnumber human users by roughly 45 to 1 in the average enterprise, and by as much as 144 to 1 in cloud-native environments, up from 92 to 1 in the first half of 2024. ^[4] Meanwhile IBM's 2026 breach research found that the share of security incidents involving shadow AI more than doubled year over year, to 43 percent, and that more than two-thirds of breached organisations had no governance process in place to limit it. ^[5][6]

The question is not whether one person can do ten times more. It is whether the organisation should depend on them doing it.

And it can't be measured. Matt Wood argues that a usable ROI calculation needs a unit both sides can be expressed in: value per accepted result, and full cost per accepted result. ^[7] State three has no such unit. The output is a draft that a person absorbs, and accepted is never defined, so there is nothing to count. You cannot compute cost per accepted result for Anna's skill, because acceptance exists only in her head.

That is the reason the 6 percent stays flat while the 80 percent keeps rising. The individual gains are real. They are simply not in a form the organisation can see.

The honest sequence is: start naturally, then promote deliberately. The promotion is a decision someone has to make and own.

State 4b: Governed Non-Determinism

Green's destination is deterministic automation — the same input under the same conditions producing the same outcome, or at least following the same defined decision path. For a large class of work that is exactly right, and workflow engines are not made obsolete by agentic tools.

An objection from practice. Determinism means two different things: a fixed result, and a fixed path to it. Hard-wire the path and you hard-wire every assumption about schemas and endpoints along with it — and those are exactly what change. An agent that works out the integration for itself is often built in hours rather than weeks, and it survives the next interface change. That shifts the threshold: processes running thirty times a month never justified an integration, and they are now within reach.

The price is in the table below. A broken workflow announces itself; an agent finds a way around — sometimes brilliantly, sometimes quietly and wrongly. It is the integration layer, not the decision layer. Which is precisely why checking the result is not optional here.

The most valuable agentic cases are in any case the ones where determinism isn't available at all. Judgement-heavy, context-dependent, high-variance work is where the value is, and it will not compile down to a flowchart.

The answer is not to keep a human watching every execution. That is state three wearing a governance badge. The answer is a rung the ladder does not draw — call it 4b: governed non-determinism.

What 4b Consists Of

• Service identity instead of a personal badge. The process authenticates as itself, with scoped permissions and a lifecycle independent of any employee.
• Evals instead of unit tests. You cannot assert equality on an output that varies. You can assert quality against a curated case set, on every prompt change, model update and configuration change. This is real engineering work rather than a checkbox: the case set has to be built, curated and kept current, and using a model as the judge adds meaningfully to what each run costs.
• Bounded variance instead of guaranteed reproducibility. Variability is acceptable where it is scoped, logged, monitored and explainable. Failure is designed, not discovered.
• Machine-readable boundaries with escalation. The agent acts on its own inside defined limits and stops the moment it reaches them. The Cloud Security Alliance describes six autonomy levels (0 to 5) and recommends level three — independent action within clear boundaries, anything beyond them handed back to a human — as what is defensible in enterprise use today. Full autonomy, in their view, is not a practical option. ^[8]
• A designated principal. Thoughtworks makes this a hard requirement: every deployed agent has a named human accountable for its outcomes, with the mandate written by a human and the escalation thresholds enforced by the platform. ^[9]

That last one is the cheapest and the most neglected. It is a single line of policy, and it converts an orphaned workflow into an owned one.

The evaluator is what makes the rest of it work. It checks the result and routes failed, unusual or high-stakes cases down a different path. ^[7] Which reframes what an eval harness is for: not a governance nicety, not a compliance artefact, but the component that defines what accepted means — and therefore the component that makes the whole thing countable. This is the answer to the measurement problem above. Until something other than a person decides what good looks like, there is no unit, and without a unit there is no ROI case, only a feeling.

What The Promotion Actually Costs

Back to Anna's questionnaires, because the interesting part is how small the step is.

The signal to promote. Not enthusiasm — recurrence. A second team asks for the skill. Volume crosses a threshold. The same work happens on a rhythm rather than by coincidence. That is the moment worth acting on, and someone has to be watching for it.

The work. A trigger on the mailbox. A service identity with scoped read access to the evidence repository. Two hundred previously approved answers become the evaluation set — that archive already exists, which is the point. Explicit boundaries: questions answerable from approved evidence get drafted automatically; anything with no source, anything touching contractual or legal language, anything genuinely new routes to a queue. The security officer, or whoever actually owns the answer, is named as principal.

The measurement. Human involvement becomes a number. Before: every answer touched by a person. After: perhaps one in five. That ratio is the metric, and it should be visible.

This is weeks, not quarters. The boundaries are a conversation, not a project.

And what stays human on purpose. The judgement calls. Anything that shades into negotiation. Anything that will end up in a contract. The goal was never zero humans; it was humans on the twenty percent that needs them, instead of on all of it.

One more thing the example is meant to show: not every skill should be promoted. If the value of a skill was one person's judgement applied freshly each time, industrialising it destroys exactly what made it good. And if the team already handles the work quickly and cheaply, automating it returns less than the finished result suggests — the baseline matters as much as the outcome. The test is whether the work is genuinely repeatable, not merely whether it is frequent.

You Don't Have To Boil The Ocean — But Defer The Right Things

The most common paralysis I encounter is the belief that the data platform, the AI platform and the governance framework must all be finished before anyone can start. That belief costs more than any tooling decision.

So: defer the warehouse, not the rails.

The enterprise data platform can be built lazily, against real use cases that have proven their value. Building it speculatively is how organisations spend two years producing infrastructure nobody asked for.

Identity, logging and a paved path are different. They are cheap, they are early, and they are load-bearing. Every number in the section above is what deferring them looks like at scale. And DORA's finding is that a high-quality internal platform is what stops individual gains being eaten downstream — which makes it a precondition for compounding, not a later phase. ^[3]

Pure bottom-up discovery has one more gap, and it is the expensive one. Individuals find friction inside their own role. They do not find cross-functional process redesign — and that is where the EBIT sits. Amazon's internal 4.5x median gains came from teams restructuring workflows, not from better prompts. ^[10]

So the journey runs on two tracks at once:

1. Bottom-up discovery. Everyone gets access, permission to experiment, and the ability to build and share skills. Treat every useful skill as a candidate for promotion, not as the finished article.
2. Top-down selection. Leadership picks three or four end-to-end processes to redesign properly. BCG's guidance is the same: align on a small number of central priorities rather than spreading across hundreds of use cases. ^[11]

Bottom-up finds the candidates. Top-down decides which ones become processes.

On the virtual expert team. A cross-functional group supporting the whole organisation through this is the right structure — with one condition. Give it a decision right, not just a mandate to help. If it only enables and evangelises, it is a community of practice, and everything stays at state three. If it owns the promotion decision — which skills become processes, who the designated principal is, what the paved path looks like — it is the organ the organisation is currently missing. High performers are twice as likely as others to have defined processes for measuring the impact of AI initiatives. ^[1] That is the difference between a centre of excellence that owns something and one that runs lunch sessions.

The Counterargument: Premature Industrialisation

The obvious failure mode of everything above is over-correction: pushing everything to state four immediately, and killing the experimentation that discovers what is worth building.

That risk is real, and it has a name. MIT Sloan frames it as minimum viable governance, against organisations that overcorrect in both directions — locking generative AI down with heavy controls, or letting experimentation run entirely unsupervised. ^[12] Governance designed for auditability rather than velocity produces process theatre: forms, reviews and approvals whose measured risk prevention is worth a fraction of the productivity they consume.

Discovery has to stay cheap and largely ungoverned. That is not a concession, it is the mechanism. You cannot select three high-value processes to industrialise if nobody was allowed to find them.

Funding it doesn't require inventing a business case. Wood proposes three questions: what useful result will this produce, what evidence would show the result is good enough, and what does the whole path to it cost? A proposal that cannot yet answer them still gets money — as an experiment, where finding those three answers is the deliverable. ^[7] Jeff Bezos put the same idea more bluntly: if you already know the outcome, it isn't an experiment. That is a sentence a CFO can approve.

And one caution in the other direction. In a controlled trial, experienced developers using AI tools were 19 percent slower than without them — and believed afterwards they had been 20 percent faster. METR published new data on late-2025 tools in February 2026, and the specific number is contested. ^[13] What matters is not the number but the gap between feeling and measurement. State four is what closes it: what runs as a process leaves logs, and logs can be argued with. Impressions cannot.

This Has Happened Before

It is worth knowing how long this particular mistake can last.

In 1900, factories ran on a single steam engine, a steel shaft along the ceiling, and a leather belt down to every machine. When electric motors became available, most owners bought one large motor, bolted it where the steam engine had been, and kept the shaft and the belts. It worked. It also changed nothing about the layout, and the productivity gains took roughly two decades to appear. They arrived only when a later generation scrapped the shaft, gave each machine its own motor, and rearranged the floor around the flow of work.

They had the best motor of their time and the old floor plan. What made them wait twenty years was not the technology. It was that the belts were still doing the routing.

Today the belts are people.

A Note On Timing

For European companies there is a scheduling argument on top. The EU moved the AI Act's deadlines in summer 2026, through a bundled amendment covering several digital laws at once — hence its name, the Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026). High-risk obligations now bite from December 2027 rather than August 2026, and from August 2028 for systems built into products. The transparency duties arrived on the original schedule. ^[14] Nothing was lightened; only the runway was extended.

That runway is worth something only if it is used. Risk management, logging, human oversight, documentation, post-market monitoring — the evidence trail the Act will ask for is very close to the machinery that makes an agentic process operable in the first place. Built as compliance work it is expensive and joyless. Built as the operating layer for governed autonomy, it produces the compliance evidence as a by-product.

The Honest Test

For any task where AI is now involved, six questions:

1. Does it run on a trigger, or does someone have to remember to start it?
2. Does it execute under a named user's credentials, or a governed service identity?
3. Is there validation and an audit trail, or does someone read the output and trust it?
4. Who owns it when the person who invented it changes role?
5. What happens when the volume doubles?
6. How much human involvement does one execution actually require — every time, or only on exceptions?

Human-in-the-loop should be a measured ratio, not a reassurance. A department where a human must inspect every output has not automated anything. It has an extremely good word processor.

One caution on that ratio: it is gameable. A process rewarded only for handling more without escalation can learn to escalate less rather than to be right more. Wood makes the same point about support systems rewarded for avoiding repeat contact — they end conversations quickly and leave people confused. ^[7] The definition of an accepted result has to include the thing you were actually protecting, not just the proxy standing in for it.

Where To Start

Three moves. None of them requires a platform, a budget cycle, or a reorganisation.

Run the holiday test on five tasks this month. Pick the five where AI has visibly changed how your team works, and answer the question honestly for each. Most will come back as state three. That is the diagnosis, and it takes an afternoon.

Name who decides what gets promoted. One person or one small group, this quarter, with an actual decision right — not a mandate to encourage. Their job is to watch for recurrence and to say which skills become processes. Without this, everything else here stays an interesting read.

Promote exactly one task, end to end. Choose something recurring, high-volume and reversible, where an archive of previously approved outputs already exists. Give it a trigger, a service identity, an eval set from that archive, an exception queue and a named principal. Measure the human-involvement ratio before and after. One completed promotion teaches an organisation more than any framework, and it produces the thing the second move needs: evidence that the step is weeks rather than quarters.

The most expensive transformation is the one decreed from above. The cheapest is the one enabled from below. Both remain true. What the last year of data adds is a third thing: the enabled-from-below transformation stops at the person unless someone is accountable for taking the next step.

The companies pulling ahead do not have the best AI, and they do not have the most enthusiastic users either. They have decided who owns the moment when a good habit becomes a process.

AGENTIC AI OPERATING MODEL GOVERNANCE PRODUCTIVITY