Every AI answer costs money to produce, and every AI answer can be wrong. How a company charges tells you who it expects to absorb both.
ChatGPT charges for acesss
Plus costs $20 a month whether you send one message or a thousand. Usage caps protect OpenAI from runaway compute bills, but a confident wrong answer still counts as service delivered. OpenAI carries the cost risk; you carry the accuracy risk.
GitHub Copilot charges per person
Copilot Business is $19 per developer per month. Since June 2026, usage beyond each seat’s AI credits costs extra. Compute risk is now shared, but buggy or insecure code is still the developer’s problem. Seats also tie revenue to headcount, an odd fit for a tool sold partly on doing more with fewer people.
Intercom Fin charges for an outcome
Fin costs $0.99 per resolved conversation. Paying only when it works sounds fair, but Intercom decides what “resolved” means, and a customer who simply gives up can count. That rewards conversations that end, not problems that get solved. The people hurt most, frustrated customers, never see the bill.
What we would choose
Our team is building a Slack bot that keeps PMs updated on project progress. Outcome pricing would push us to send more updates, not better ones. We would charge per PM seat with a fair-use cap on model costs, and let users flag any wrong status so we fix it at no charge. Price the access, own the mistakes.
