A failed job should never cost the customer

When I defined pricing and packaging at ReelUse (three paid tiers, team plans, an upgrade path and free-plan limits), the line I spent the most time on wasn't a price. It was a policy: a failed job never counts against a customer's allowance.

It sounds small. I think it's the most important line on the page.

AI fails in ways customers can't see coming

Customers send us long recordings and get short clips back. Most of the time that works. But AI processing can fail for reasons the customer can't predict and couldn't have prevented: a file that won't decode, a job that times out, a model that returns nothing usable.

From the customer's side, all of those look the same. They waited, and they got nothing. If that also uses up part of what they paid for, we've turned our technical problem into their financial one.

Allowance is a promise, not a meter

An allowance tells a customer what they can count on getting this month. If failures eat into it, the number stops meaning anything, and people start rationing. They upload less, try less and hedge more, and that's exactly the behaviour a growing product doesn't want.

Protecting the allowance protects the habit. It tells the customer to go ahead and try, because if it doesn't work, that's on us.

Making it real, not just written down

A policy like this only matters if the product enforces it without anyone having to ask. That means the definition of "failed" has to be agreed before launch, the billing logic has to follow it automatically, and the help centre and support answers have to say the same thing the system does.

That last part ties into how we release everything: pricing rules ship with their explanation in the same change. A rule a customer can't find is a rule they won't trust.

What I'd tell another PM

If you price an AI product, write your failure policy before you write your tiers. Customers forgive a failed job. They don't forgive paying for one.