Is an AI centre of excellence worth it, or does it slow you down?

Usually not yet. An AI centre of excellence earns its overhead only once you cross two of three thresholds: AI running in production in 3 or more functions, roughly 1,000 or more staff, or a named regulatory exposure. Below two of three, it mostly adds a review queue to work that was already slow.

• What it is: a funded team that sets AI standards, reviews deployments and holds the model and vendor inventory.
• The test: two of three — 3+ functions in production, ~1,000+ staff, a named instrument. One of three, don’t build it.
• The cost nobody budgets: the queue tax — extra days per deployment × deployments a year.
• Below the line: one accountable owner plus a written AI register — practice 1 of Australia’s national AI guidance.
• Bought systems: an outsourced, outcome-priced deployment sits outside the build queue and inside the register.

What an AI centre of excellence actually is — and what it is not

An AI centre of excellence (AI center of excellence in the US; CoE for short) is a standing, funded team that owns AI standards, reviews proposed deployments and maintains the inventory of models and vendors. The defining feature is not expertise. It is standing budget and a veto.

Three things it is not — conflating them is where the argument goes wrong.

• Not a governance framework. A framework is a document; a CoE is a cost centre with headcount. You can hold the first without funding the second, and most organisations under 1,000 staff should.
• Not the team that runs a live AI system day to day. Whoever answers the pager at 7pm sits in the function that owns the process, not in a central lab.
• Not one shape. Three live patterns: centralised CoE (builds and approves), hub-and-spoke (central standards, practitioners embedded per function), federated guild (no budget, shared standard, recurring forum). Only the first two carry real overhead.

A centre of excellence is a veto with a payroll attached — so the question is never “is it good practice?” but “does our volume of AI decisions justify making them centrally?”

Is an AI centre of excellence worth it? The two-of-three test

Our decision rule, deliberately blunt. Count how many are true today:

1. Breadth. AI in production in 3 or more distinct functions. Pilots do not count — a pilot is something you can switch off on a Friday without telling anyone.
2. Scale. Roughly 1,000 or more employees, or 200 or more people affected each month by an AI-influenced decision — the same line the table below uses. Below that you cannot staff a CoE without stripping the function that was delivering.
3. Exposure. A named regulatory instrument, not a general sense of risk. “We handle personal information” is not exposure; “we make credit decisions” or “we deploy an Annex III high-risk use case into the EU market” is.

One of three: do not build it — you need an accountable owner, not an institution. Two of three: probably, in hub-and-spoke form. Three of three: yes, and you are already late, because the standards you should have set centrally have been set locally, three different ways.

The thresholds where the answer flips

Each row changes the answer on its own merits. These numbers are our decision rule, not a published standard — the point is that the answer flips on measurable conditions, not on ambition. The one externally fixed date sits in the last row.

The sign-off row is the fastest diagnostic in the set: if you cannot name the person who approves an AI deployment inside sixty seconds, your gap is accountability — and a CoE is only one of several ways to close it.

When an AI centre of excellence slows you down: the queue tax

Every CoE inserts a review step. Price it as the queue tax:

Queue tax = (median days from request to approved deployment with the CoE − median days without it) × AI deployments per year.

Measure both halves before you charter the team; afterwards there is no counterfactual left. Then apply the rule: if the CoE’s median review adds more days than the deployment takes to build, it is not a standard-setter, it is a bottleneck. A ten-working-day review on a six-month platform build is cheap insurance; on a three-day workflow change it turns three days into thirteen — roughly a 4× slowdown, and functions stop asking — which is how you get unregistered AI tools bought on a corporate card.

Does the centre of excellence pay for itself? The arithmetic

Work it with your own numbers. Ours are illustrative:

• AI initiatives started per year: 10
• Baseline failure rate: 8 in 10 — an assumption, and a pessimistic one. RAND’s 2024 study of AI project failure opens by noting that “by some estimates, more than 80 percent of AI projects fail — twice the rate of failure for information technology projects that do not involve AI”. RAND is repeating an outside estimate there, not measuring one — its own finding, from 65 practitioner interviews, is five root causes. Substitute your own rate.
• Internal cost of a failed initiative: 3 people × 4 months = 12 person-months
• CoE standing cost: 4 FTE = 48 person-months per year

Each failure prevented returns 12 person-months, so covering 48 person-months of salaries means preventing four failures a year — taking the failure rate from 8-in-10 to 4-in-10. A four-person AI centre of excellence running ten initiatives a year has to halve the organisation’s AI failure rate just to break even on its own payroll, before a single day of queue tax is counted.

Two caveats. This counts avoided waste only, not the upside of choosing better projects. And it cuts the other way at volume: across 30 initiatives the same team need prevent only four of roughly 24 expected failures. A CoE is a fixed cost defending a variable benefit, so it gets cheaper per unit exactly as your AI volume grows.

What to put in place below the threshold

Score one of three and the work still exists — it just does not need a department. Australia’s National AI Centre published its Guidance for AI Adoption on 21 October 2025, evolving the Voluntary AI Safety Standard’s ten guardrails into six essential practices, and the first is “decide who is accountable”: assign a senior leader as the overall AI governance owner, with enough authority and understanding of AI capabilities and risks to oversee all AI use, and clarify who is responsible for each part of the supply chain where vendors and integrators are involved. It does not say “stand up a centre of excellence.”

The minimum viable version takes a day to build and about an hour a month to maintain:

• One named accountable executive — a person, not a committee, able to stop a deployment.
• An AI register: one row per system — owner, vendor, data touched, decision influenced, date last reviewed. Five columns.
• A pre-deployment checklist for where you are actually exposed — consent basis, disclosure, opt-out handling and record-keeping are the recurring four in customer contact. Our AI outbound compliance checklist for enterprise buyers sets those out for Australia and the US.
• A quarterly register review, with a written decision per row: keep, fix, or retire.

A five-column AI register with one accountable name against every row does more compliance work than a centre of excellence that has not yet agreed its terms of reference.

Where an outsourced AI deployment sits relative to a CoE

This gets missed because CoE literature assumes you are building. A bought outcome sits outside the centre of excellence’s build queue and inside its register. The CoE does not design, staff or maintain it, but still owns three things: who signs off, what gets monitored, and how you exit.

That changes the economics in a way finance notices before IT does. A built system converts into headcount and a multi-year platform commitment; an outcome-priced one converts into a variable line — on our own revenue-share model that line is a performance fee of 5–20% of the sales generated, and on the pay-per-appointment alternative it is charged per booked qualified appointment rather than per seat or per month. Different approver, different budget, different review path — and whichever way it is bought, the register entry and the named sign-off still apply.

The trade is honest both ways. Outsourcing removes build risk, hiring and model maintenance; it adds vendor assurance work — and at five or more vendors that work is itself a condition pushing you toward a CoE. The assurance question worth asking any vendor is for the denominator behind their headline number. Our methodology page defines our 7× average sales lift as trailing three-month closed-deal revenue at month six against the three months before launch, averaged across clients who supplied both — and discloses that the median is closer to 4×. Demand that level of definition from everyone in your register, including us.

Customer contact and pipeline — qualification, follow-up, AI appointment setting — is the narrow lane where an outsourced deployment is easiest to govern, because the outcome is countable. For where procurement and security review sit, see enterprise lead generation services.

Frequently asked questions

What is an AI center of excellence?

An AI center of excellence is a standing, funded team that sets AI standards across an organisation, reviews and approves proposed deployments, and maintains the central inventory of AI models, vendors and use cases. It is distinguished from a working group by two things: permanent budget and the authority to stop a deployment. Spelling varies — “center” in the US, “centre” in Australia and the UK — but the function is identical.

Do we need an AI centre of excellence to comply with AI regulation?

No regulator currently requires one. Australia’s Guidance for AI Adoption, from the National AI Centre, sets out six essential practices, the first being to decide who is accountable — voluntary guidance specifying roles, not org charts. In the EU, the AI Act entered into force on 1 August 2024 and became applicable on 2 August 2026, with AI literacy obligations applying from 2 February 2025 and Annex III high-risk obligations extended to 2 December 2027. Those rules attach to systems and to roles such as provider and deployer, not to whether you run a CoE. General information, not legal advice — check your obligations at source.

How many people should an AI centre of excellence have?

Size it from the work, not the ambition. By the arithmetic above, a 4-FTE team across 10 initiatives a year must halve the failure rate just to break even on salaries; the same team across 30 initiatives need prevent only 4 of roughly 24 expected failures. Below 6 initiatives a year, a fractional owner plus an external reviewer on call is the honest structure.

Is a centre of excellence the same as an AI governance framework?

No. A framework is the written set of rules and controls you apply to AI systems; a centre of excellence is one of several structures that can apply them. The others are a named owner with a register, hub-and-spoke, and a federated guild with no budget. Every organisation using AI needs the framework. Only some need the CoE.

We already have a data and analytics CoE — can it just take AI?

Often yes, and it is usually the cheapest correct answer. The overlap is real in data lineage, model monitoring and vendor review. The gap is three specific things: generative model behaviour testing, disclosure and consent in customer-facing systems, and agent permissions — what a system may do without a human in the loop. Add those three to the existing charter before chartering a second team.

How do I know if our AI centre of excellence has become a bottleneck?

Track two numbers monthly. The queue tax: median days from deployment request to approval, times requests per year. And the shadow rate: AI tools found in use that are not in the register. A rising shadow rate is the leading indicator — people route around review before they complain about it, and by the time the complaint reaches you the unregistered tools are in production.