Forecast based prepay pools expired 10 to 30 percent unused, against a discount capped at 20
Microsoft sells Copilot Credits three ways, but the real decision is binary: stay flexible on pay as you go, or commit up front for a discount that tops out at 20 percent and expires. The arithmetic is less favorable than the headline.
Prepared by Redress Compliance · August 15, 2026 · Microsoft advisory. Based on Microsoft EA and Azure commitment renewals, 2024 to 2025.
Executive summary
Prepay pools sized on a vendor forecast were the single most common source of waste, expiring with 10 to 30 percent unused.
The discount ceiling is the whole problem. The Pre-Purchase Plan runs 5 to 20 percent, so a pool finishing a fifth unconsumed has given back more than it ever saved.
Pay as you go is one cent per credit, billed in arrears, with no up front purchase and no expiry, which is why it is the buyer side default.
Undersizing is the safer error, because overage simply falls back to pay as you go at the standard rate.
Run pay as you go first. Buyers who did so for a quarter or two sized their eventual commitment far more accurately.
Three models, one real decision
| Model | Commitment | Discount | Expiry risk |
|---|---|---|---|
| Pay as you go | None, billed in arrears | None, one cent per credit | None |
| Capacity Packs | Monthly, per tenant | Fixed price for 25,000 credits | Monthly, credits reset |
| Pre-Purchase Plan | An annual pool, bought up front | 5 to 20 percent, tiered | Annual, unused credits expire |
The three differ on commitment and flexibility, not on what a credit does. Every credit is the same unit of metered agentic work, pooled at the tenant, whichever way you bought it. So the choice is purely a choice about who carries the risk of being wrong about volume. Pay as you go leaves it with Microsoft, who are only paid for real usage. Prepay moves it to you, in exchange for a discount. Capacity Packs sit between, converting a variable line into a fixed monthly one that resets rather than accumulates.
When each model is the right answer
- Prepay pays on high, proven, steady volume, where the base is large enough that even a single digit discount is material and the pool will genuinely be burned.
- Pay as you go wins on uncertainty, covering pilots, rollouts, seasonal demand and modest volume, because it bills only for what ran.
- Capacity Packs fit steady monthly consumption that finance would rather see as a fixed line, accepting that unused credits reset each month.
- Optimization argues for flexibility, since shifting heavy workloads to direct models lowers credit volume and strands a pool you already paid for.
- Size to the floor of the proven range, never the middle and never the vendor forecast, per the true cost analysis.
- Treat overage as the cheap failure, because exceeding the pool simply bills at the standard rate rather than penalizing you.
What Microsoft Copilot Cowork really costs
What a task actually costs in dollars, the prepay floor, and the same work priced on a direct model.
Read it free →A discount you do not consume is a deposit you forfeit
In the commitment renewals we worked, prepay pools sized on the vendor forecast were the single most common source of waste, and the discount rarely justified the lockup. Pools expired with 10 to 30 percent unused. Set that against a discount that runs 5 to 20 percent and only reaches the top at very large commitments, and the arithmetic answers itself: a pool that finishes a fifth unconsumed has handed back more than the discount was ever worth.
What makes this trap effective is that the discount is real and the loss is invisible. The percentage appears on the quote, gets reported internally as a saving, and is banked in the business case on the day of signature. The expiry happens twelve months later in a different reporting period, shows up as nothing at all rather than as a charge, and is attributed to demand coming in below plan. Nobody experiences the forfeited credits as the reversal of the discount they announced, which is precisely why the pattern repeats.
The asymmetry between the two failure modes should settle the decision. Undersize and you fall back to pay as you go at the standard rate, so the penalty for being wrong is that you pay the normal price for the extra usage. Oversize and the surplus is simply gone at term end. One error costs nothing beyond the discount you did not earn on the incremental volume; the other costs the full value of the unused pool. When the downside is that lopsided, the correct target is the floor of your proven range rather than the middle of your forecast.
Which points at the sequence rather than the model. A forecast is not proof, and the account team is paid to book committed revenue early, which a prepay does perfectly. Your interest is the opposite: keep the dollars flexible until consumption is demonstrated. Run pay as you go for a quarter or two, watch where the volume actually settles, and only then commit, at the floor, on a base large enough for the discount to matter. Buyers who did that sized far more accurately than those who committed on a projection. The agentic model sits in the Cowork brief, the seat economics in the Copilot licensing brief, and the full cost model in the true cost analysis.
Watch the briefing · 4:51Negotiating Microsoft E5, E7, and Copilot Cowork: The Two-Layer BillThe governance floor and the execution meter are two separate bills, and agent execution runs through Copilot Credits with no rollover.
- Copilot seats matched to actual usage, not the vendor adoption pitch
- Credit pools sized from your real consumption curve rather than a forecast
- Shelfware surfaced and priced at your contract terms
The sequence that sizes a commit
Run it variable
Pay as you go bills only for what ran and never expires, which turns the first quarter into measurement rather than a bet.
Find the floor
Establish the level consumption does not drop below, separating steady agentic work from pilot and seasonal volume.
Commit at the floor
Only if the base is large enough for a single digit discount to matter, and only to the proven floor rather than the forecast.
What the commitment file shows
Across the Microsoft EA and Azure commitment renewals worked in 2024 and 2025, two patterns stood out:
On pools sized from a vendor forecast rather than from measured consumption, wiping out the headline discount.
What separated buyers who sized their eventual commitment accurately from those who committed on a projection.
The patterns: the discount banked at signature, the expiry absorbed as a demand miss, and the pool sized to the middle of a forecast rather than the floor of a measurement.
The buyer side move is to keep the dollars flexible until consumption is proven. The wider library sits in the Microsoft practice.
Your first five moves
- Default to pay as you go until you have measured consumption rather than forecast it.
- Separate steady agentic work from pilot and seasonal volume, because only the first belongs in any commitment.
- Size any commit to the floor of the proven range, never the middle and never the vendor worksheet.
- Test whether the discount is material on your base, since 5 percent on a small pool does not offset expiry risk.
- Accept undersizing as the safer error, because overage bills at the standard rate. The Microsoft practice sizes the pool with you.
Frequently asked questions
What are the three ways to buy Copilot Credits?
Pay as you go at one cent per credit billed in arrears, Capacity Packs at 25,000 credits for 200 dollars per tenant per month that reset monthly, and the annual Pre-Purchase Plan bought up front at a tiered discount. They differ on commitment and flexibility, not on what a credit does.
How large is the prepay discount?
It runs 5 to 20 percent and only reaches the top of that range at very large commitments. That ceiling matters, because unused credits expire at term end, so a pool that finishes with more than a fifth unconsumed has given back more than the discount was ever worth.
How often do prepay pools go unused?
Forecast based pools expired with 10 to 30 percent unused in the commitment renewals we worked. That was the single most common source of waste we saw, and the discount rarely justified the lockup that produced it.
When does prepay actually pay?
In a narrow band: high, proven, steady volume on a base large enough that even a single digit discount is material, and where you are confident of burning the pool. Notice that a vendor forecast is not on that list, because a forecast is not proof of consumption.
When does pay as you go win?
Most of the time. It wins whenever volume is uncertain, seasonal, early stage or modest, because it bills only for real usage and never expires. It is also the only model that lets you optimize freely, since you are never holding a pool you have to drain.
What are Capacity Packs for?
Steady, predictable monthly volume that you would rather see as a fixed line than a variable one. They sell 25,000 credits for 200 dollars per tenant per month and reset each month rather than rolling over, so unused capacity is lost monthly rather than annually.
Is undersizing a commit risky?
It is the safer error. Exceeding the pool simply falls back to pay as you go at the standard rate, so the penalty for undersizing is that you pay the normal price. The penalty for oversizing is forfeited credits, which is why the floor of your proven range is the right target.
How should a commit be sized?
Run pay as you go for a quarter or two first, then size to the floor of the proven range rather than the middle or the forecast. Buyers who did this sized their eventual commitment far more accurately than those who committed on a projection.
Cowork, Agents and Copilot Credits
Session 7 of the Microsoft EA Renewal 2027 Series. The metered layer arriving beside the seat: Cowork billed per task, Agent 365 governing the agents, and Copilot Credits underneath both. How to forecast a consumption tail nobody can predict, and what to negotiate before you sign it.
