M&T Bank's Copilot population has been published three different ways in twelve months, and none of the three is the number that decides whether a per-seat licence is earning its price. Sixteen thousand used it in September 2025. Seventeen thousand had it rolled out to them by December. More than fifteen thousand have it deployed as of this month. Those are three measurements of three different populations, reported by three different outlets — and across the most recent twelve months M&T has reported, its own quarterly filings show the headcount underneath them fell by 928 people.
M&T is the reference case everyone reaches for when they need proof that a regional bank can do enterprise AI properly. It probably is one. But if you are the CIO renewing a per-entitled-seat agreement this quarter, the M&T story as published gives you nothing you can put in a business case, and the reason is worth more than the anecdote.
The Same Deployment, Three Different Populations
The three figures are not a correction sequence. They are three different questions, each answered once and never asked again.
American Banker reported in September 2025 that 16,000 of M&T's 22,000 employees use Microsoft Copilot, following a pilot that started with 800 people. In December 2025, chief data officer Andrew Foster told CDO Magazine, "We have rolled out Microsoft Copilot Chat and Microsoft Copilot 365 to 17,000 employees." On 4 September 2026, AI News reported that the bank "has deployed AI copilots to more than 15,000 employees."
Use, rolled out to, and deployed to are three different things. Rolled-out and deployed are entitlement counts — a licence exists against a user object in a directory. Use implies activity but carries no time window, so it could mean once, ever. And "more than 15,000" is a floor, not a measurement: 17,000 also satisfies it. A 2026-09-06 review of the coverage read the sequence as a decline in penetration from 77% to 68%. That reading assumes a fixed denominator, and the filings say otherwise.
None of this is M&T behaving badly. Nobody in that sequence was asked for a comparable number, so nobody produced one. Which is exactly the failure mode you inherit when you cite the case in your own board deck.
The Denominator Moved Too, By 928 People
The "22,000 employees" that every version of this story divides by is not a constant, and M&T publishes the real series itself.
Its third-quarter 2025 results put full-time equivalent employees at 22,383 at 30 September 2025 — the closest reported count to American Banker's "22,000." Its fourth-quarter and full-year 2025 results, released 16 January 2026, put FTEs at 22,080 at 31 December 2025. Its second-quarter 2026 results, released 15 July 2026, put FTEs at 21,662 at 30 June 2026, against 22,590 a year earlier — a decline of 928, or 4.1%, with the release attributing part of the drop in salaries expense to "lower average staffing levels in the recent quarter."
Run the three published figures against the nearest reported FTE count and the penetration line reads 71.5%, then 77.0%, then 69.2%. That is an eight-point swing in nine months produced by numerators that measure different things and a denominator that shrank underneath them. It is not a trend. It is noise with a decimal point.
Here is the part that should change how you read every seat-count announcement: M&T's FTE count fell by 418 between the December figure and the June filing. Applying the December entitlement rate of 77% to that decline — arithmetic, not a disclosed figure — puts roughly 320 Copilot entitlements walking out of the count on their own. An entitlement count falls by 320 and the story writes itself as flagging adoption — when what actually happened was attrition in a bank that got smaller. Headcount reduction and abandonment look identical in a licence count. They look nothing alike in an activity count.
Microsoft Already Ships the Number You Want
You do not need M&T to publish weekly actives. You need to publish your own, and Microsoft built the report.
The Microsoft Copilot usage report in the Microsoft 365 admin center defines exactly three metrics that matter here. Enabled Users is "the total number of unique users in your organization with Microsoft Copilot licenses over the selected timeframe" — your entitlements. Active Users is "the total number of enabled users in your organization who tried a user-initiated Microsoft Copilot feature, in one or more apps in Microsoft 365 over the selected timeframe." Active users rate is one divided by the other. The report supports the last 7, 28, 90 or 180 days. Weekly active users is a dropdown.
The bar for "active" is real but low. Microsoft's own FAQ: "if a user selects the Copilot icon in the Word ribbon to open the Copilot chat pane, this action doesn't count towards active usage. However, if the user interacts with the chat pane by submitting a prompt, this action counts."
The dashboard your executives actually see is looser than the report. The Microsoft Copilot Dashboard in Viva Insights defines "Active Microsoft Copilot users" as "the number of Copilot licensed users who used Copilot at least once in any of the following Microsoft 365 apps during the last 28 days," and it is fixed at 28 days with "a delay of up to six days" and a further seven-day lag before newly licensed employees appear at all. If you have only ever seen the Viva tile, you have only ever seen monthly actives, lagged.
Two more traps in the same documentation, both of which move your ratio without anything changing in the business. The admin-centre user table "shows all users who were licensed for Microsoft Copilot at any point over the past 180 days, even if the user later removed the license or never had any Copilot active usage." And in Viva, "if users lose their Copilot licenses, they aren't included in reporting" — except for employees who also hold a Viva Insights licence, for whom the measured period "includes all employees who were enabled for Copilot at any point during that time period," while anyone holding "only a Copilot license" is removed from every analysis once that licence goes. Whether a leaver stays in your denominator turns on which second licence they happened to have. Two Microsoft tools, three populations, same tenant.
And you cannot roll your own. Microsoft states plainly that audit log data "is not, however, intended to be used as the basis for Copilot usage reporting," and that metrics built on it "might not be consistent" with the official reports.
What "Active" Buys You at $360 a Year
At list price, one prompt every four weeks costs about $28 a prompt and scores as a fully active seat.
Microsoft prices Microsoft 365 Copilot at "$30.00 user/month, paid yearly" on an annual commitment, as an add-on assigned per user on top of a qualifying base plan. That is $360 per assigned user per year. A user who submits exactly one prompt inside each rolling 28-day window is counted 100% active on the dashboard for all thirteen windows in a year — thirteen prompts for $360.
Do the arithmetic on the M&T numbers and the stakes get concrete. Fifteen thousand seats is $5.4M a year at list; 17,000 is $6.12M. The 2,000-seat gap between the two published figures is $720,000 a year — a rounding error against technology spending that exceeded $1.2 billion in 2025, and a number no CFO should have to guess at. The problem is not that $5.4M is large. It is that nobody outside the tenant can tell you what it bought, and if your only KPI is entitlements, nobody inside it can either.
This is the same defect the industry keeps rediscovering under different names: a containment rate that isn't a resolution, a 30% time saving that turns out to be a target rather than a measured result, a 743,000-seat rollout reported as an outcome. Deployment is an input. It has been reported as a result for three years.
The Only M&T Number With a Real Denominator
Six minutes per call is the one figure in this entire story you can multiply by something you already know.
Foster said generative AI summarisation of call-centre conversations saves about six minutes per call. That number has a unit attached to it — a call — and you know your own call volume and your own loaded cost per agent-minute. It converts into a business case in one line of arithmetic. A seat count does not convert into anything, because a seat is a subscription, not a transaction.
Note also what Foster says the tool actually delivers: "Gen AI gets you 60% of the way, then a human reviews it and takes it the other 40%." That 40% is labour, it is recurring, and almost nobody prices it into the savings — the same review-cost gap that showed up when NHS clinicians stopped catching what the scribe dropped. Six minutes saved minus the minutes spent checking is the honest figure, and it is still probably positive. Publish that one.
What Made the Rollout Land Predates the Model
The transferable part of the M&T story is a seven-year platform programme that had nothing to do with generative AI when it started.
Between 2018 and 2025, annual technology releases went from about 15,000 to 65,000, with outages down more than 80%. On the data side, M&T runs Solidatus and Monte Carlo for lineage and data quality, grounds retrieval on an internal repository of authoritative documents and bank policies called Edison, and has put roughly 2,000 employees through a data academy. Monte Carlo was not bought to make Copilot work. It was bought because a regulated bank needs to know where its numbers come from.
That ordering is the actual lesson, and it is the same one behind Santander's five-figure-headcount rollout: the release cadence and the lineage came first, and the assistant landed on top of them. If your change-failure rate is high and your lineage is a spreadsheet, buying 15,000 seats does not get you M&T's outcome. It gets you M&T's invoice.
What to Do Before You Sign
This Week: Open the Microsoft 365 admin centre, go to Reports → Usage → Microsoft Copilot → Copilot, and set the window to 7 days. Write down Enabled Users, Active Users and the active users rate. That single screenshot is the number your renewal turns on, and most organisations have never taken it. Take the 28-day and 90-day views in the same session so you can see how much of your "adoption" is one prompt a month.
This Month: Pin the denominator in writing. Agree with HR and your identity team on one entitlement source and one reconciliation date each month — the 180-day user table and the Viva measured group are different populations, so pick one and stop comparing across them. Then subtract leavers explicitly: report entitlement changes as net of headcount movement, or your next attrition wave will show up in a board pack as an AI adoption problem.
Before Renewal: Put three things in the agreement. A named activity threshold that defines a productive seat for your organisation — weekly active, not monthly, and more than one prompt. A quarterly true-down right on seats that fall below it, because an annual commitment with no true-down means you fund the gap for twelve months. And at least one per-transaction metric with a unit, on the model of six minutes per call, that you commit to publishing internally whether or not it flatters the programme. If your vendor's team resists the weekly window specifically, you have learned something about what the 28-day number is hiding.
Before Renewal, part two: Price the review labour. Take Foster's 60/40 split as a planning assumption until you have your own, apply it to every workflow you claimed savings on, and see which business cases survive. Some will not, and finding that out before you sign is cheaper than finding it out in the run cost after you have already skipped the build-versus-buy analysis.
The Bottom Line
Per-seat AI pricing survives because the buyer and the seller have agreed to measure the same thing: seats. Entitlements are easy to count, they only move when procurement moves them, and they make a rollout look like a result. The industry has spent three years reporting deployment as an outcome, and the pricing models are already moving on — toward consumption and outcomes — while the reporting has not. Whether you are buying Copilot, Glean or something else on a per-seat model, the discipline is identical and it is not technical.
M&T did the hard part. It rebuilt release engineering, it bought lineage tooling before it needed it, and it produced one honest per-transaction number. What it has not published — what almost nobody publishes — is the ratio that says whether the licence is working.
Count the people who used it this week. Divide by the people you paid for. Everything else is a press release.
Continue Reading
- DBS Put 1,500 Bankers on Agents. The 30% Is a Goal.
- 743,000 Employees Later: What Accenture's Copilot Rollout Reveals
- Agentic AI Pricing: Don't Buy Consumption Without a Cap
- Glean vs Copilot vs Dust: Buy the Permission Model
- Best AI Contact Center Platforms: Containment Isn't Resolution
- A Third Skipped a SaaS Buy. Now Price the Run Cost.
- 40% of Bank Code Written by AI: Santander's €1B Bet