More prompts, more data stored, more tokens and credits used. The bill keeps going up. As companies scale, LLM usage scales with them. Even after negotiating unit costs down, you're looking at 40–50% margin at best. So how do you get back to 80%+ margins?

A few options: increase prices, increase revenue to cover the cost of AI, or decrease costs. Most teams reach for the first lever. Almost none reach for the right one.

The reality is 5–10% of users are actually driving 90% of costs. A price increase for 100% of your base when 90% aren't the problem is a tough sell. A company's cost increase is a company problem — not a customer one.

So design a lever that only the 10% will feel.

The math

You have 1,000 customers on a $50/month plan. Your AI feature costs you roughly $2 per customer per month in compute at median usage. That's manageable — $2K in variable cost against $50K in revenue.

But the P95 customer isn't using $2 worth of compute. They're using $25. And the P99 customer? $80. That top 1% — just 10 accounts — might cost you more in compute than the bottom 500 combined.

You have two options. Raise prices across all 1,000 customers to cover the tail. Or set a fair use threshold at something generous — say the 95th percentile — and only trigger a conversation, overage, or throttle for the 50 customers above it.

Option one risks churn from 950 customers who were never a problem. Option two protects your margin without touching the experience for anyone using the product normally.

One large customer can move product COGS by multiple percentage points without anyone noticing. I've used fair use to cut product COGS by 33% in six months.

4 real examples

Grafana Labs
Grafana Cloud Logs uses a 100x fair use query policy. Query up to 100x your ingested log volume at no extra cost. Go beyond that, and billing kicks in. Transparent, generous, and most customers never get close.
OpenAI
ChatGPT uses rolling message caps by model. Hit your limit on GPT-5 and the system drops you to a lighter model until your window resets. The experience degrades instead of stopping — a meaningful difference from a hard cutoff.
Atlassian Rovo
Bundles Rovo AI into paid subscriptions with a credit-based quota. Not enforcing limits yet — but they've published quotas, committed to 90 days' notice before enforcement, and will include dashboards for usage monitoring.
ClickUp
AI Super Credits with "Super Fair Billing." They pass on savings when AI model costs drop, subsidize spikes when costs rise, and provide grace periods if you burn through credits faster than expected. Fair use plus an explicit economic compact.

Four different approaches to the same problem. Here's how to figure out which one fits you.

How to build your own

1

Inspect

Start by pulling up usage by user. If you have 1,000 teams with an average of 10 users per product, look at the distribution. If you see the 90/10 pattern — 10% of users driving 90% of costs — you have a strong candidate for fair use.

2

Analyze

Take the P95 of your usage. There's a percentile function in Excel and Sheets — use it. The 95th percentile becomes your fair use threshold. You can position this to customers as: "95% of customers will be unaffected by this policy."

P95 makes sense for dev tools with a long tail of usage. For infra or spiky usage, P90 may make more sense. For productivity tools, you might go as high as P99.

3

Apply

Enforce fair use in every contract and apply it to your entire customer base — even if it only affects 5%. Then decide: hard limit (rate limit, stop usage, or automatically charge overages) or soft limit (notice to customer but allow continued scaling).

The benefit of soft limits: you have the right to renegotiate pricing without disrupting customer scaling. Win-win.

Rollout by deal type

Renewals

Roll out the policy for net new customers first, then existing at their next renewal. All current contracts should be grandfathered in. In renewal conversations, make sure customers have a clear way to track how close they are to the threshold. Proactive alerting is a bonus.

Expansion

A customer exceeding fair use is a great signal. Either they're rapidly growing and it's time for a renegotiation, or they're misusing the product and need support. Either way, the conversation is worth having.

Enterprise contracts

Proactively flag fair use as part of the sales conversation. I've used it as a competitive advantage — other companies may charge for dimensions like queries as an alternative. Your fair use policy can be the more transparent, more generous option in the room.

Self-serve

For credit-card customers, having fair use clearly called out on the pricing page plus examples in the documentation is enough. Help them understand what it is, how to track it, and how to optimize so they stay under.

The irony

Fair use is actually better for the customer it's enforced on. That customer running 150x their expected usage? They're either getting massive value and should be on a bigger plan, or they're doing something wrong and need help. Either way, the conversation is worth having.