In our experience this is rarely one big mistake. It’s four small defaults that nobody chose on purpose, stacked on top of each other. The tools you built with ship with them switched on, and nobody gave you a reference point for what normal looks like.

Here are four checks. You can run all of them in an afternoon, and the first two usually explain most of the gap.

Check 1: What happens when the AI call fails?

Look at: the retry setting on every step that talks to an AI provider. In visual automation tools it’s often a toggle called “retry on fail” or “error handling”.

A bad answer: “It retries up to five times, I think.” Each retry sends the full request again, including the whole document or conversation, and you pay for every attempt. The nasty part is timing. Failures cluster when the provider is slow or overloaded, so your spend spikes at the exact moment your product is struggling. A step that fails one time in five at normal load can fail most of the time on a bad afternoon.

Do this: cap retries at two, wait longer between attempts, and make sure a failure shows you an alert instead of looping silently. A loop that nobody sees is the most expensive kind.

Check 2: Which model is doing each job?

Look at: the model name on every AI step. Write them in a list.

A bad answer: the same biggest, newest model everywhere. That’s what most tutorials and templates default to, because it works best when you’re proving the idea. But the largest models often cost ten times or more per word than the small ones in the same family. Sorting an email into “billing / support / sales” does not need the model you’d use to draft a legal summary.

Do this: keep the big model for the one or two steps where customers would notice a drop in quality. Move classification, extraction, reformatting and “is this message spam” to a smaller one. Test on twenty real examples before you switch. If you can’t tell the outputs apart, you’ve found free money.

Check 3: Are you paying to answer the same question twice?

Look at: your last 100 requests. How many were identical or near-identical?

A bad answer: more than a handful. The usual version is a long fixed block of instructions, or the same product documentation, sent along with every single request. You pay to re-read it every time. Another common version is regenerating something that hasn’t changed, like a summary of a document nobody edited since yesterday.

Do this: save results you’ve already paid for and reuse them when the input is the same. Ask whoever built it about “caching”, which just means remembering the answer. Most providers also discount repeated instruction blocks heavily if you structure the request correctly. This one is a real build task, not a toggle, but it often cuts the bill by a third or more on document-heavy products.

Check 4: What stops it at €5,000?

Look at: the billing page of your AI provider. Is there a monthly spending limit, and is there an alert at 50% and 80% of it?

A bad answer: no limit, or a limit so high it’s decorative. Without a ceiling, one customer pasting in a 400-page file, one bug in a loop, or one leaked key is limited only by your card.

Do this: set a hard monthly limit at about twice your expected bill, and an email alert at half. Then add a per-customer limit inside your product, so one account can’t eat everyone’s budget. It takes an hour and it is the cheapest insurance you will ever buy.

What normal looks like

Founders overpay partly because they have nothing to compare against. These are rough monthly AI-usage ranges for a typical text feature (summaries, drafting, classification) used a few times a day per customer. Yours will differ, but if you’re far above the range, something above is leaking.

What moves you up the range: long documents instead of short messages, images or audio, and features that run automatically instead of when a customer clicks. What moves you down: smaller models, remembered answers, and trimming what you send. Cleaning up the first three checks typically puts a bill back inside its row. Check 4 keeps it there.

The position

Cost is a design decision, not an invoice you discover. If your product only works economically at the biggest model, no limit and unlimited retries, you don’t have a margin. You have a subsidy, and you’re the one paying it.

Product Blueprint — €2,500, 1 week. architecture, data model and a deployed skeleton, plus a fixed quote for the full build — credited in full against a Sprint booked within 60 days.

It starts with a free product teardown: two hours on what you are building, what already exists, and what is actually blocking launch. You leave with a written plan and a realistic number, whether or not you work with us. Book a teardown · What it costs