AI Center of Excellence: Your Intake Ranks Guesses, Not Hours

An AI center of excellence (AI CoE) is the central team that sets an organization’s AI standards, runs its governance, and decides which AI use cases get funded and in what order. That last job decides whether the rest works. The CoE’s intake, the ranked list of use cases it chooses to fund, is only as good as the data on how the work is done today.

In most organizations, that data is a guess. The value column on a typical intake form is filled in by the person asking for the money.

Take Priya, head of AI strategy at a regional property and casualty insurer. She is an illustrative composite, not a client. Her CoE’s quarterly intake drew 34 submissions, and the top-ranked one came from claims: an AI triage tool that sorts first notice of loss reports and routes each claim to the right adjuster, with “saves 30% of adjuster time” in the value field. The 30% was a show of hands in a team leads’ meeting.

The pilot ran for three months. When the results came in, sorting and routing new claims turned out to be about 9% of adjuster hours. The tool cut that roughly in half.

The tool worked as built and freed about 4.5% of adjuster time, against the 30% it was ranked on. The project ranked second never started.

This guide is for the CoE lead who owns that list and the operations leader who feeds it. It covers what the CoE does, the AI operating model choices in brief, why most intake ranks estimates, what a measured intake form asks for, and how to rank and re-rank the list. If you already have an intake queue and doubts about the numbers in it, book a 30-minute strategy call and bring it.

Quick Answer

An AI center of excellence succeeds or fails on its intake: the ranked list of use cases it funds. Most intake forms ask the submitter to estimate hours spent and hours saved, so the list ranks estimates. Require measured hours by activity, reconcile them to paid hours, and score only what is measured.

What an AI Center of Excellence Does

An AI center of excellence sets the AI strategy, standards and governance, builds skills, runs pilots, and maintains the prioritized backlog of AI initiatives. Microsoft’s Cloud Adoption Framework lists “create intake and prioritization workflows” as a core CoE responsibility and tells the team to “maintain a prioritized AI initiative backlog” (Microsoft Learn).

The usual list of responsibilities looks like this:

• Strategy: which business goals AI serves, and where to look for use cases
• Standards and governance: data quality, model life cycle, security, and responsible AI policies
• Skills: assessments and learning paths for the people who will use and build AI
• Pilots: proofs of concept that test an approach before it scales
• Intake and prioritization: collecting requests, scoring them, and keeping the ranked backlog
• Measurement: reporting adoption and business impact to leadership

Every item on that list serves the backlog. Governance decides what may be built, skills decide who can build it, and pilots test it. But the backlog decides what gets the money, and the CoE’s results are the results of whatever sits at the top of it.

The measurement half is where most CoEs are weakest. In IBM’s 2025 study of more than 600 chief AI officers, 68% said they start AI projects even when they can’t measure their effects, and 72% said their organization risks falling behind without AI impact measurement (IBM). A project you can’t measure afterwards was usually ranked on a number nobody measured before.

Choosing an AI Operating Model for the CoE

An AI operating model sets who owns AI decisions, where the experts sit, and how work moves from idea to production. The common choices are a centralized CoE, a hub and spoke model, and an embedded model where each function runs its own AI with thin central guardrails. IBM found that organizations whose chief AI officer runs a centralized or hub and spoke model can realize a 36% higher ROI on AI initiatives (IBM).

The operating model sets who owns the intake; the evidence rules for what goes in are a separate decision.

Microsoft’s guidance is to start centralized and move toward an advisory role as AI delivery matures. That matches the AI operations maturity model, where the last stage is a standing, data-driven operating model rather than a series of projects. Whichever model you pick, the evidence standard for the intake is the part you can’t delegate.

Why Most AI Center of Excellence Intake Ranks Estimates

Most AI use case intake forms ask the submitter for two numbers: how much time the work takes today and how much the AI would save. Both are estimates. A published intake template asks for “estimated time currently spent” and “estimated time saved” (Pertama Partners). KPMG’s CoE guide describes a framework that uses standardized mechanisms “to promote those with higher perceived value” (KPMG).

Perceived is the right word. The two-axis grid these guides recommend, value against feasibility or value against complexity, has a value axis filled in by the person who wants the project.

That produces three predictable failures.

The best-written submission wins

A department with a skilled writer and a confident number rises. A department with a messier problem and an honest “we don’t know” sinks. The same intake guide warns that without criteria, “the loudest voice or most senior requester wins” (Pertama Partners). Criteria help, but criteria applied to estimates still rank estimates.

Operations submits volumes, not effort

Operations submissions cover the most people. In a 400-person organization, 300 might work in operations, so that is where AI teams look first. The systems those teams work in count finished output, such as claims processed and tickets closed, and keep no record of the effort behind each one.

So when an operations manager fills in “hours today,” they start from a volume report and multiply by a handle time someone estimated years ago. Looking things up in a second system, rekeying data, and redoing returned work leave no trace in a volume report, so they never reach the estimate. Those are often the best AI targets, and they rarely appear on the intake at all.

The budget moved, and the competition changed

A few years ago, an operations improvement project was approved inside operations, out of the operations budget. Today it goes to the AI strategy team, which ranks it alongside every AI request in the company. Operations leaders tell us they now compete for the same budget against projects like a chemistry AI tool, not against other consulting bids.

That is a fair fight only if both sides bring the same kind of number. A specialist team with a vendor’s benchmark and an operations team with a workshop estimate are not on the same scale.

What a Measured Intake Form Asks For

A measured intake form asks for the hours an activity takes today, measured over a window long enough to include a month-end, for the people who actually do it. It also asks how much those hours differ from person to person, how much of them go to lookups across systems and returned work, and what must not break. Our guide to what a chief AI officer needs from operations explains each of those inputs. Here is how they become intake fields.

Three rules make the form hold up.

1. The hours must reconcile: Add up the hours claimed by every submission from one team. If three submissions each claim 30% to 40% of the same team’s week, something is wrong before anyone scores anything. Activity hours should reconcile to the hours you pay for, which is the check at the end of a workflow analysis.
2. The CoE computes the value, not the submitter: The submitter supplies measured hours and the problem. The CoE applies the same cost and automatable share to every submission, so value means the same thing across departments.
3. Unmeasured is not rejected: A submission without measured hours is logged and marked “needs a baseline.” It stays visible, and it is not scored on a guess.

Here is what changes when the form asks for measured hours. Marcus, VP of loan servicing at a credit union, is another illustrative composite. His team of 60 had an AI agent for payoff quote requests on the intake, valued at “40% of team time.”

Measured over 30 days, payoff quotes were 11% of the team’s hours. Status lookups across the core and the servicing system, plus fixing files sent back from underwriting, took 23%. Marcus withdrew one submission and filed two, both with hours attached. The second one, which nobody had named before, ranked higher.

If your CoE is redesigning its intake form, we can walk through it with you. Book a call with Wendy and bring the current form.

How the CoE Ranks and Re-Ranks the List

Once the hours are measured, AI use case prioritization is the easy part. Score every candidate on the same weighted criteria, rank by score, and let a short list of rules override the score where service or compliance is at stake. The automation prioritization framework gives six weighted criteria and a worksheet. Its heaviest criterion, addressable hours, is exactly the field a measured intake fills.

The five-step method in how to prioritize AI projects in operations covers sequencing the first pilots. For the wider question of which work is worth automating at all, start with how to know what to automate in operations.

What the CoE adds is the discipline around the list:

• One unit across departments: Hours and loaded cost, the same for every submission. That is how a claims project and a specialist project end up on the same scale.
• A fixed cadence: Re-rank every quarter, and again whenever a baseline is rerun. A list ranked once goes stale as the work changes.
• A reason for every position: Publish why each item sits where it does. When the reasons are visible, lobbying turns into better submissions.
• The after-read is funded with the project: Every funded item gets a read against the same baseline once it is live. Measuring AI ROI in operations shows the calculation.
• Parked items stay on the list: they are often where the next quarter’s best candidates come from.

This is how the CoE avoids the 68% problem. A project ranked on a measured baseline can be read against that baseline afterwards. A project ranked on an estimate can only be compared to another estimate.

What Operations Brings to the Intake

The operations leader’s best contribution to the intake is a baseline of the whole area, not a handful of submissions. One measured baseline produces many ranked candidates at once, including the ones nobody had a name for, and it gives the CoE the same measure to read every project against later.

That is the gap we built the Ground Truth AI² Platform to fill. A lightweight desktop client captures each click and the screen region around it, not the full screen, with no keystroke capture.

Patent Pending vision AI classifies each capture at the activity level, so you see “entering customer data into an account form,” not “in Salesforce.” Each activity is scored for automation potential. The result is ground truth workforce data: a measured baseline of the work, activity by activity.

Our Phase One engagement runs a fixed 90 days and covers at least 50 employees in the target operations area, with an initial findings summary in weeks three to four. The Ground Truth AI² Report brings the complete operational data set, an AI replacement assessment, a prioritized deployment roadmap, and a staffing model analysis. The final readout is with the operations leader, the AI team, and the C-suite together. See how the workforce intelligence platform works.

The specialized AI tools on your intake list still need a baseline to be measured against, which is why the baseline should come first.

Back to Priya. The next quarter, her CoE added one rule: no measured hours, no score. Claims operations ran a baseline across 120 people before resubmitting.

It produced seven candidates, and three of the new top five were work nobody had submitted before: rekeying between the claims and billing systems, status lookups, and reopened files. The triage tool stayed on the list, ranked at the size the hours supported.

Frequently Asked Questions

What does an AI center of excellence do?

An AI center of excellence sets an organization’s AI strategy, standards and governance, builds AI skills, runs pilots, and maintains the prioritized backlog of AI use cases. Its most consequential job is the intake: deciding which use cases get funded and in what order. The quality of that ranking depends on the data behind each submission.

Who should be on an AI center of excellence team?

An AI CoE team usually has an executive sponsor, a dedicated CoE lead, data scientists and machine learning engineers, governance and security specialists, and business leaders who bring use cases and know the work. For operations use cases, include an operations leader who can supply measured hours, not only estimates, for the work each use case would touch.

What is an AI operating model?

An AI operating model defines how an organization runs AI: who owns decisions, where AI experts sit, how use cases move from idea to production, and how results are governed and measured. Common models are a centralized center of excellence, hub and spoke, and embedded, where each function runs its own AI under shared guardrails.

Which AI operating model is best for an enterprise?

Most enterprises start with a centralized CoE to build standards and expertise, then move toward hub and spoke as more teams build AI. IBM’s 2025 study found that organizations whose chief AI officer runs a centralized or hub and spoke model can realize a 36% higher ROI on AI initiatives. Whatever the model, one evidence standard for the intake matters more than the org chart.

What information should an AI use case intake form capture?

An AI use case intake form should capture the problem, the specific activity the AI would touch, the hours that activity takes today (measured, not estimated), how much those hours differ from person to person, how much time goes to lookups across systems and rework, what must not break, and how the result will be read after go-live. The CoE, not the submitter, should compute the value.

How long does it take to build an AI center of excellence?

A typical first 90 days covers an executive sponsor, a CoE lead, a first set of standards, and two or three first use cases. The intake can run from the first month. A measured baseline for one operations area, the data the intake needs to rank fairly, takes 90 days with our Phase One, with first findings in weeks three to four.

Book a 30-Minute Strategy Call

If you lead an AI center of excellence with operations use cases in the queue, or you’re an operations leader about to submit one, the ranking goes better when both sides bring measured hours.

In 30 minutes we’ll go through your intake form, the use cases waiting for a decision, and the scope a measured baseline of that work would need. You’ll leave with notes the CoE and the operations teams that submit to it can both work from. Book a 30-minute strategy call.