Clearing the Gate Is Not Permission: Charge Back One Copilot Line, Show Back the Rest

The model does not get fixed. It gets quietly abandoned, and the whole metered line goes back to sitting in an IT cost pool where it grows unwatched.

The failure is almost never arithmetic. It is scope. Teams treat the materiality gate as a green light for full chargeback, when all it licenses is building a model at all. Those are different decisions, and running them together is what puts you in a meeting you cannot win.

Here is the split that holds up.

The metered layer is additive. Cowork jobs, Work IQ grounding for custom agents and Copilot Studio agent runs bill in Copilot Credits, on top of the flat per-seat licence. The per-seat price did not move.

Microsoft's own billing documentation is unusually clear on the shape of it: in some scenarios, licences act as an entry point enabling access to AI services billed on a pay-as-you-go basis. Read that twice if anyone in your organisation has been saying Copilot went metered. The seat is still a seat. What has been bolted on is a consumption meter next to it.

That single change breaks per-seat allocation, because one invoice now carries two lines that split by two different logics.

Headcount logic still splits the seats perfectly. It cannot split a credit pool, because credits are not consumed in proportion to who holds a licence. That is the entire problem. Everything below is about the second row.

Before you design anything, run this:

metered_line / invoice_total

Under 5%, stop. Do not build chargeback machinery for a rounding error. Run showback for two quarters instead: publish each unit's number, charge nobody, and re-run the division at every monthly close.

Now the part that gets skipped. Clearing the gate tells you the metered line is big enough to be worth allocating. It does not tell you that every dollar in it can be allocated defensibly. Those are separate questions, and the second one is answered line by line, not invoice by invoice.

The numbers below come from an invented organisation. They are illustrative, sized to match what a mid-market estate actually looks like.

Brightwell Insurance, roughly 500 Copilot seats. August invoice:

$3,800 over $15,800 is 24%. Comfortably past a 5% gate. Under the reading most teams apply, Brightwell now builds a full consumption-share model across every unit and charges the lot.

What Brightwell actually does is charge back one line.

$2,400 of $3,800 gets charged. That is 63% of the metered line moved onto the budget that caused it, on the strength of an attribution nobody can argue with: one agent, one workflow, one department. The remaining $1,400 shows up on statements, named and visible, charged to nobody.

Arithmetic where attribution is clean. Patience where it is not.

Why the partial model is the stronger position

Two reasons, both practical.

Full chargeback on day one buys you a number. Partial chargeback buys you a number plus a mechanism.

Some agents serve everyone. The IT helpdesk bot. The HR policy bot. The all-hands meeting summariser. Their credits are real and billed, and they are unownable by any single cost center.

Pool them and reallocate by seat share, not consumption share:

unit_shared_cost = shared_pool_total * (unit_seats / total_seats)

The instinct is to split shared cost the same way you split everything else in the metered line, by credits consumed. Resist it. A shared agent's cost scales with who is entitled to call it, not with who runs Cowork hardest.

This rule holds whichever allocation method you land on. Direct attribution, consumption share, seat weighting, fixed percentage with a true-up: all four need the shared pool carved out first and reallocated by seats.

Brightwell's credit line was $600 a month one quarter earlier. It is $3,800 now. A 6.3x rise in three months.

Nothing about the estate changed structurally. Cowork went generally available worldwide in mid-June 2026 and the tenant enabled it, one department built one useful agent, and the meter did what meters do.

So the 24% is not the interesting figure. The trajectory is. A gate you fail at 3% today can pass by autumn, and the only way to know is a dated series. Which produces two unglamorous habits worth more than any model:

Pull your last three invoices and run the division on each. If the ratio is climbing, you have your answer about whether this is worth twenty minutes.

Four months from now nobody will remember why the split works the way it does. The person who decided it may be on another team. Put it on one page:

COPILOT COST ALLOCATION DECISION

Gate result:      metered / total = ____%      as of ____________
Verdict:          [ ] showback only   [ ] chargeback

Method per line type:
  Per-seat layer:       ______________________________
  Attributed credits:   ______________________________
  Shared-agent pool:    seat-share reallocation

Reopen this decision when:
  1. The metered line crosses the gate threshold in either direction
  2. Attribution coverage crosses 90% in either direction
  3. ______________________________

Decided by: ______________    Date: __________    Next review: __________

Those two reopen triggers are the ones that actually fire. The third line is yours, and it is usually a reorg or a contract renewal.

A signed page like this is the difference between a chargeback model and an opinion. It is also the document that ends the second-month conversation before it starts, because the answer to why does my unit carry this is written down, dated, and signed by whoever owns the budget.

That is a defensible allocation for the Copilot metered layer, and none of it requires a tool.

If you would rather not do the arithmetic by hand, the companion script for this post runs steps 1 through 4 off your own invoice and usage exports and prints a verdict per agent line: Copilot Credit Chargeback Gate Check. Standard-library Python, no credentials, no API calls, and it writes nothing except the output files you ask for. It measures and triages. Step 5, the signature, stays yours.

Related reading

The seat-level question that comes before this one, whether the licence pays for itself at all, is worked through in The Real Cost of Copilot. The renewal-math version, where the metered layer meets a bundle negotiation, is in the E7 renewal post.

And if what worries you is the meter itself rather than who pays for it, the ceiling logic for token-denominated spend is in The Agent Tax.

If the answer came back build it, the build is four things we could not fit in a blog post: a calculator whose totals tie to the invoice to the cent, a policy page finance will sign, a repeatable monthly close, and the reconciliation discipline that survives a disputed statement. That is The Copilot Chargeback Playbook, $49, with the formulas live in the workbook cells so you can check the math against your own export.

No sponsors, no vendor partnerships. The Playbook is my own product, so treat the recommendation accordingly. Prices cited are Microsoft list as of June 2026, and if you are on an enterprise agreement your negotiated rates differ. This is cost-allocation guidance, not tax, accounting or legal advice.