The AI SaaS repricing playbook: fix margins, keep customers

AI Pricing Lab · Updated 2026-07-17

If your margins say you have to reprice (and if you have not checked, start with margin per customer), the risk that keeps you from acting is churn. The evidence from the 2025 to 2026 repricing wave is reassuring in one way and brutal in another: customers absorb higher or restructured prices, and they do not absorb surprise. Cursor's June 2025 change was economically defensible and still ended in a public CEO apology and refunds, because users found out from their invoices. The playbook below is built to make sure nobody finds out from an invoice.

Step 1: Simulate before you announce anything

Take your last 6 to 12 months of per-customer usage and replay it under the candidate pricing. For each customer: what would they have paid, what does their margin become, who crosses a pain threshold. This tells you three things no gut feeling can: total revenue impact, the exact list of accounts whose bills jump more than about 20 percent (your personal-outreach list), and whether the new model actually fixes the margin leak or just moves it. This simulation is the core of our Pricing Reset, but a spreadsheet version beats not doing it at all.

Step 2: Shadow-bill for a month

Run the new meter in production without charging it. Generate the invoice both ways and compare. Shadow billing catches metering bugs, edge cases (the customer with the batch job at 2 a.m.), and the gap between simulated and live behavior, while every mistake is still free. If your billing system cannot do this, even a nightly script computing what each account would have been charged is enough.

Step 3: New customers first

Launch the new pricing for new signups two to four weeks before touching the installed base. New customers have no anchor, so this is a clean read on whether the structure sells, and it stages your support load. It also means that by the time existing customers hear about the change, the pricing page already says it, which reads as consistency rather than a raid.

Step 4: Grandfather with a date, not forever

The two failure modes are symmetric. Zero grandfathering reads as a bait and switch and drives your angriest churn. Permanent grandfathering freezes your unprofitable accounts in amber, and in an AI product those are precisely the accounts the repricing exists to fix. The pattern that works: existing customers keep current terms for a defined window, 3 to 12 months, or until their annual renewal. That is how GitHub handled the Copilot migration to usage-based billing in June 2026: monthly plans moved at the next cycle, annual subscribers held their terms until renewal. For your handful of extreme whales, skip the form email and have the conversation directly; a custom ramp for two accounts is cheaper than the noise they can make.

Step 5: Announce with 60 to 90 days and their own numbers

For a structural change (flat to usage, seats to credits), give 60 to 90 days notice. For a simple rate increase, 30 days is workable. The announcement must contain each customer's own projected bill under the new model, computed from their real usage. "Most customers will pay less" is vendor arithmetic; "your last three months would have been $41, $38, and $44 against your current $49" is a fact the customer can check. The AI-cost framing is also unusually sympathetic, so use the honest version of it: your prices are changing because heavy AI usage has real cost, and the new model makes heavy users pay for their own weather.

The announcement email, copy-paste

What not to do: the 2026 casualty list

• Do not let the invoice announce it. Cursor's users discovered new charges on their bills. Refunds and an apology followed, and the story is now the canonical cautionary tale in every pricing thread.
• Do not price after the fact. Replit's effort-based checkpoints show the price only after the work is debited. Whatever its economics, it generates a steady stream of bill-shock posts. If your unit price varies, show an estimate before, not a receipt after.
• Do not dress a price increase as a generosity. Copilot's messaging drew the response "you will get less, but pay the same price" precisely because the framing oversold. State the trade plainly; skeptical users respect arithmetic and punish spin.

Signs your pricing is broken enough to justify all this

• Gross margin under 50 percent while peers run 60 plus (see the 2026 benchmarks)
• Any customer whose attributed AI cost exceeds their subscription
• Your P90 customer costs more than 5x your P50 under a flat price
• Revenue grows and gross profit does not (the growth paradox: more customers, less money)
• You hesitate to promote your own product to heavy users because they lose you money

Two or more of these and the repricing conversation is overdue. The structures to move to are covered in the pricing models guide and the seats versus usage versus credits comparison.