38 standalone questions on the machinery that turns pilots into production — the commissary, the three failure modes, 10-20-70, and pilot to platform to portfolio.
Audience: senior business leaders. No technical background assumed.
Source: the CoE and scaling handout — the adoption-versus-impact gap, the commissary kitchen, the three failure modes, 10-20-70, and the pilot → platform → portfolio progression.
How to use this: every question stands alone. Pick an option, then read the answer. The ten sections run in the order the argument builds: the problem, what a CoE is, how it fails, the effort split, how scaling actually works, then ownership, staffing, measurement, funding and timeline.
A · Step 1 — The gap that a CoE exists to close
Q1Research across 300+ enterprise GenAI deployments and $30–40 billion of investment found about 95% produced no measurable P&L return. What did the report itself blame?
Answer: (2)The models were not the problem. That single finding is why the answer is organisational machinery rather than better technology, and it is worth saying immediately so the room does not conclude the technology is a fraud.
Q2A global survey found 88% of organisations using AI regularly in at least one function, but only 39% reporting enterprise-level EBIT impact and around 6% reporting more than 5% EBIT impact. What is the shape of that problem?
Answer: (2)Near-universal adoption, rare impact. Note the same pattern in agents specifically: 62% experimenting, 23% scaling. Everyone is doing something; very few are doing it at a size that shows up in the accounts.
Q3What is the diagnostic question that reveals an organisation's real AI maturity?
Answer: (2)The ratio between those two numbers is your real maturity. Most rooms go quiet at this point, and the silence is the most useful thing that happens in the first ten minutes.
B · Step 2 — What a Centre of Excellence actually is
Q4A restaurant group with forty outlets cooks everything in one central kitchen and ships it out. What goes wrong?
Answer: (2)Consistent, controlled, safe — and a bottleneck. This is the fully centralised CoE, and its failure is not incompetence; it is structural.
Q5The same group instead lets all forty outlets cook independently. What goes wrong?
Answer: (2)Fast, local, responsive — and chaos with a compliance incident attached. This is the fully federated model.
Q6What does the third option — the commissary — actually make?
Answer: (2)Each outlet then finishes the dish locally, adds its regional variation, and plates it for its own customers. The CoE is the commissary, not the kitchen. It makes the components everybody needs; it does not cook every dish. This is the hub-and-spoke model, and it is the consensus default — not because it is elegant, but because the two alternatives fail in documented ways.
Q7A business unit wants to build something with GenAI. What single question tells you which model you actually have?
Answer: (2)Permission → you have a committee. Parts → you have a Centre of Excellence. It is a one-sentence test that cannot be answered with a deck.
C · Step 3 — The three ways a CoE fails
Q8A CoE has 22 staff and a nine-week intake queue. Three business units have quietly bought their own tools on departmental cards. Which failure mode is this?
Answer: (3)Every idea must pass through people who are already fully booked, so business units either wait or — far more likely — go around you and build shadow AI on a corporate credit card, which is the outcome you were created to prevent. Symptom: a backlog measured in weeks and a growing list of tools nobody approved.
Q9The CoE lead in that situation asks for more headcount to clear the backlog. Why is that the wrong fix?
Answer: (2)The correct move is to federate: publish an approved tool list and a self-service path so low-risk work never enters the queue, and reserve the central team for platform components and genuinely high-risk builds.
Q10A team ships a technically excellent claims-triage model. Six months later, 14% of claims handlers use it; those who do not say it "doesn't fit how we actually work." Which failure mode?
Answer: (2)Symptom to look for: high build quality, low usage. Anything under about 20% adoption ninety days after launch is in this category — and accuracy was never the problem.
Q11A bank's CoE has published a governance framework, a model risk taxonomy, an ethics charter and a quarterly assurance report. In eighteen months nothing has gone into production. Which failure mode?
Answer: (3)The sharpest line in the literature: a CoE built as an IT governance function with vague authority and no accountability for business outcomes will produce governance artifacts, not business results. Symptom: the deliverables list contains no user-facing systems.
Q12All three failure modes share one root cause. What is it?
Answer: (2)If nobody on the CoE has a number in the company's plan that they are personally on the hook for, you will get one of these three. The fix for the governance-theatre case is exactly this: change the mandate from "assure AI" to "deliver £X of value by Q4, safely."
D · Step 4 — Where the effort actually goes
Q13The 10-20-70 rule splits AI deployment effort. What does the 70% cover?
Answer: (3)10% algorithms and models, 20% technology and data, 70% people and process. And the observation that makes it land: most AI programmes invert this, spending 70% of their attention on the model and the platform and treating people and process as a communications afterthought.
Q14A hospital buys an MRI scanner. What determines whether it improves patient outcomes?
Answer: (2)The scanner is the easy part — you write a cheque and it arrives. Buy the scanner and skip the rest and you have an expensive room. The model is the scanner, and every organisation can buy the same scanner. That is precisely why the scanner is not where advantage comes from.
Q15How do you apply 10-20-70 as a diagnostic to your own organisation?
Answer: (2)If the answer is nowhere near 70% on the second, you have just found your most likely failure mode — and you have found it before spending the money rather than after.
E · Step 5 — Scaling is three different activities
Q16What is the output of the pilot stage?
Answer: (2)One use case, one team, a hand-built solution, generous tolerance for mess — and the output is a decision, not a system. Most pilots should end in "no", and a CoE where every pilot succeeds is not being ambitious enough.
Q17What changes at the platform stage?
Answer: (2)It is the shift almost nobody plans for, because pilots are exciting and platforms are not. Nobody's promotion case says "I built the shared evaluation harness."
Q18What is the single best test of whether you have a platform?
Answer: (2)If yes, you have a platform. If no, you do not — you have a collection of projects that happen to use the same technology. It is devastatingly effective to ask a CoE lead, because it cannot be answered with a deck.
Q19At the portfolio stage, where does the constraint move?
Answer: (3)And a new competence becomes essential: killing things. The owner also changes at each stage — an enthusiast owns the pilot, the CoE owns the platform, the executive team owns the portfolio.
Q20Fewer than 30% of GenAI pilots ever reach production, and only about 25% of organisations have moved 40% or more of their pilots across. Where does the 95% failure rate actually live?
Answer: (2)Organisations stall between stages 1 and 2 because the work that closes the gap is unglamorous and nobody's career depends on it — and then they try to do stage 3 with stage 1 machinery.
F · Step 6 — Ownership
Q21Who chooses which use cases get built?
Answer: (2)Read the business-unit column out loud in the ownership table: every item that determines whether anything actually changes sits with the business, not the CoE. That is deliberate, and it is the whole design.
Q22Which decision does the CoE genuinely own outright?
Answer: (2)Along with data access and security patterns. It advises on workflow redesign but the business owns it, because it is their process; it recommends killing a project but the business decides.
Q23Why must a CoE refuse accountability for adoption?
Answer: (2)The CoE is accountable for making it possible. The business is accountable for making it happen. Blurring that line feels collaborative and reliably produces the Ivory Tower.
Q24In the ownership split for measuring benefit, who owns what?
Answer: (2)The CoE defines what "cycle time" means so it is comparable across the organisation. The business is on the hook for whether the number actually moves — which is the only way the number ever moves.
G · Step 7 — Staffing
Q25What is the most common staffing error in building a CoE?
Answer: (2)A useful CoE is small and cross-functional. The lead must be senior enough to kill things; the product lead translates business problems into buildable scope; the engineers build shared components rather than every solution.
Q26Which role is described as not optional and usually cut first?
Answer: (3)They are the 70% of the 10-20-70. A CoE without one is a CoE that has decided to fail in a specific, predictable way — and it will call the result "resistance to change."
Q27When should risk, legal and compliance be involved?
Answer: (2)Compliance arriving at the end is how projects die three weeks before launch — and how the CoE acquires a reputation for saying no, which then makes people route around it.
Q28What does it mean if CoE headcount is growing faster than production deployments?
Answer: (2)A twelve-person central team supporting a 20,000-person organisation is normal, and the spokes matter more than the hub. A healthy CoE's share of the work goes down over time; if the centre is still building everything in year two, the model never federated.
H · Step 8 — Measure impact, not activity
Q29"We ran 23 pilots this year." What does that measure, and what should replace it?
Answer: (2)Activity is not impact, and the distinction is where credibility is won or lost with a board. Similarly: licences bought → weekly active users as a percentage of licences; employees "trained" → percentage using it in their work 60 days later.
Q30Which single metric best predicts whether a CoE is working?
Answer: (3)If your tenth deployment takes as long as your first, the CoE has produced nothing durable. If it takes a third of the time, the platform is real. It is the second-use-case test expressed as a trend line.
Q31"AI-enabled processes: 30." What should be measured instead?
Answer: (2)"AI-enabled" tells you nothing at all — it describes a technology present in a workflow, not a change in what the workflow produces or costs.
I · Step 9 — Funding decides the outcome
Q32Each use case is funded separately, on its own business case. What does that reliably produce?
Answer: (2)Nobody's individual business case includes a line for components someone else will use next year, so those components never get built.
Q33The CoE is funded centrally and its services are free at the point of use. What happens?
Answer: (1)Free things attract unlimited demand and no scrutiny. But pure chargeback fails the other way — it produces real prioritisation while starving the platform work nobody wants to pay for.
Q34What is the recommended funding split, and what is the logic?
Answer: (3)Nobody will ever voluntarily pay for the shared evaluation harness, so it must be funded centrally or it will not exist. But if the centre also pays for use cases, the business has no skin in the game — and a business unit that has not paid for something will not do the hard work of changing how its people operate. In one line: fund the road centrally; make each department buy its own vehicles.
J · Step 10 — Timeline and health
Q35How long does a fully mature CoE take — governance in place, shared infrastructure live, federated model working, metrics reported?
Answer: (3)Anyone promising it in a quarter is selling you a slide. Days 0–90 should produce a written charter with a named executive owner and a budget line; months 3–6 something real in daily use outside the CoE; months 6–12 the second-use-case test passing; months 12–18 finance agreeing the numbers.
Q36What are the two counter-intuitive markers of a healthy CoE?
Answer: (2)A CoE that has never cancelled a project is not prioritising. And if the centre is still building everything in year two, federation never happened — the hub-and-spoke model became a hub with spokes drawn on.
Q37A retailer has eleven live deployments across six business units, each built by a different integrator with its own pipeline, model contract and monitoring. The eleventh took longer to build than the third. What is the diagnosis and the fix?
Answer: (2)Eleven projects, not a programme, and the second-use-case test fails badly. The fix is painful, unglamorous and the only thing that works. Note that this case and the nine-week-queue case are opposites — one is too much centre, the other too little — and almost every real CoE sits somewhere on that line.
Q38Three questions to take back. Which is not one of them?
Answer: (4)Model choice is the 10%. The other three interrogate the 20% and the 70% — and each can be asked on Monday, by a non-technical leader, without any preparation.