A Central AI Team Buys Demos; Embedded Capability Buys Work

Which two jobs belong in the center?

Tooling and the standard. Tooling means contracts, security review, data handling terms, access management and a single written statement of what may go into which system. The standard means one definition of what counts as a checked deliverable, applied identically across functions, so that a marketing brief and a claims decision are held to the same evidence rule even though the content shares nothing.

Both are genuinely central work because both are worse when duplicated. Five functions negotiating five contracts get five sets of terms and no bargaining position. Five functions writing five verification standards get five, which is the same as none, because nobody can then say across the company what a checked piece of work means.

Keep the group small enough that it cannot absorb delivery. Three to six people is enough at mid size: someone who owns vendor and security, someone who owns the standard and the training design, and one or two who can sit with a function for a fortnight and leave.

What does not belong in the center is the thing every consultancy proposal puts there, which is a portfolio of use cases the central group builds on behalf of the business. That arrangement makes the center the customer of its own work, and the measure of success becomes the portfolio.

Why does delivery belong to the function?

Because the expensive judgment is occupational and does not transfer. Knowing that a comparable-sales figure is the number a valuation turns on, or that a coding modifier decides whether a claim is paid, is domain knowledge held by the people doing the job. A central specialist can build the workflow and cannot tell you which line in it will be costly if the model gets it wrong.

The Boston Consulting Group field experiment is the cleanest illustration. On 18 realistic consulting tasks inside the model's capability, consultants using GPT-4 completed 12.2% more tasks, finished them 25.1% more quickly and produced work graders rated more than 40% higher ¹. On one task deliberately placed outside that capability, the same tool made people 19 percentage points less likely to be right ². One firm, one sitting, a 2023 model. The transferable part is that the boundary was invisible to capable people, and the only defence against an invisible boundary is someone who knows the domain well enough to be suspicious in the right places.

The same reasoning explains why the work reshapes rather than disappears. The Burning Glass Institute found skills exposed to automation were 16% more likely than baseline skills to see demand decline in postings, while skills exposed to augmentation were 7% more likely to see demand rise, with the most automated occupations also the most augmented ³. Those are relative likelihoods against a baseline rather than the size of any change, and posting text is employer language. The shape still matters for org design: capability that reshapes a job has to live in the job.

Work out where that applies before staffing anything, because it is rarely everywhere at once. Which roles actually need AI skills right now is the question that sizes this decision.

Write the sunset condition before the first hire

State in the founding document what has to be true for the central group to shrink, and put a date on the review. Something checkable: tool contracts signed and renewing on schedule, the verification standard published and in use by every function, and two consecutive quarters where operating teams ran their own reviews without central help. Then review it on the date whether or not anyone raises it.

Without that clause the group does what every permanent function does, which is to find more work. The failure is not bad faith. A central team measured on activity will produce activity, and the activity that is easiest to produce is another pilot in another function, which also happens to justify another hire.

Three clauses worth having in the founding document.

• A named end state. What the organization looks like when the center is three people instead of nine, described concretely enough to argue about.
• A review date, not a review trigger. Triggers are never pulled by the people they affect.
• A rule about delivery headcount. If the central group is building deliverables for a function after the first two quarters, either the function is not staffed for it or the center has absorbed the work.

If a mandate arrived from above rather than from a function's need, the sequencing is different and tighter, and what has to be built in the first ninety days is a shorter list than the announcement implied.

What should the exec team ask for at the quarterly review?

One artifact per operating function, held next to what that function produced a year earlier, with the change named by whoever is accountable for it. Not slides about pilots. A quarterly report, a claims decision file, a campaign brief, a forecast, and a specific difference in each: a brief written before generation, a claim checked against a source, a draft rejected for a stated reason.

Be careful with the adoption numbers that arrive alongside the request. Published US adoption rates disagree by construction rather than by error: about 18% of firms had adopted AI at the end of 2025 on the Census measure, about 41% of the workforce reported using it at work in the same period, and an employment-weighted estimate put 78% of the labor force at firms that had adopted something ⁴. Those are three different units, and the 78% describes employers rather than users. Any of them can be quoted to make a program look ahead or behind.

The firm-level picture is also strongly size-dependent. Census put AI use among US businesses at 19.8% as of May 2026, against 37% among firms with at least 250 employees ⁵. Useful for sizing expectations, useless as a target, and the survey asks about use in the past two weeks rather than about depth.

Then ask two questions that no dashboard answers. Whether the verification standard is being applied by people who do not report to the central group, and whether the functions could keep going if the group were halved next quarter. If the answer to the second is no after two years, the capability was never embedded, and the headcount plan is resting on an assumption nobody has tested. The program that survives this review is the one where the upskilling ran inside the functions rather than beside them.

See the benchmarks >