Do AI agencies guarantee results? What a fair guarantee looks like

Most writing on the subject treats an outcome guarantee as a warning sign. SEO.com's "7 AI SEO Agency Red Flags", updated 18 September 2026, puts guarantees of visibility first, because AI search results shift from one searcher to the next. The Black Sheep calls guaranteed rankings or returns "sales tactics, not performance commitments". Wellows, in an article updated 26 October 2026, says "AI is probabilistic, not deterministic" and lists platform changes, model drift and the client's data as reasons no agency can promise a fixed result.

Three things make outcome guarantees on an AI build hard to honour.

The model and its price can change after launch

AI providers change models on their own timetable. Google's Gemini API pricing page says that from 1 December 2026 every request to gemini-2.5-pro will route automatically to gemini-3.8-flash. On the same date Gemini 4 Argon's input price doubles from $2.00 to $4.00 per million tokens. An agency that guaranteed an outcome on the old model would be held to it on a different model with a higher bill.

Results depend on the client's data

An agent that answers from the client's documents is only as good as those documents. SpotDev, which sells a £5,000 data readiness assessment, says that if data is "scattered across spreadsheets, duplicated or full of gaps, that groundwork has to happen first". Wellows makes the same point. The agency rarely controls the data.

Hours saved depend on staff using it

A support agent saves hours only when the team sends enquiries through it. Adoption sits with the client, so a guarantee of hours saved asks the agency to promise something it can't enforce.

A fair guarantee names something the agency controls, measures it in writing and says what happens if it is missed.

1. It covers a controllable promise: a delivery date, agreed acceptance tests, a fixed price or a response time.
2. The success criteria are written before work starts and tested on your real inputs.
3. The remedy is specific, such as a refund, a waived payment or a fix at no charge, and any cap is stated.
4. It says what it excludes. Augustova's line on hours saved, revenue and conversions is an example.
5. You can claim it without a long minimum term.
6. You own the code, the data and the accounts, so you can leave if it fails.

UK law sets a floor under all of this. Section 13 of the Supply of Goods and Services Act 1982 implies a term that a supplier acting in the course of a business "will carry out the service with reasonable care and skill".

Yarigai's guarantee is one sentence: "If the build isn't profitable for you, we won't build it."

Before the client pays anything, the one-week blueprint puts the maths in writing. It states what the system replaces, what that costs a year and what the build costs. If the build doesn't come out ahead within the first year, Yarigai doesn't take the job.

A worked example uses the figures on Yarigai's homepage. A $500 a month tool is $6,000 a year. Ten hours a week of admin at $20 an hour is $10,400 a year. A build that replaces both has $16,400 of first-year cost to beat ($6,000 + $10,400), plus its own running costs.

A refund guarantee pays out after the work fails. Yarigai's rule applies before the work starts, so it doesn't depend on how a model behaves months later. The client owns the code, the data and the accounts from day one, and Yarigai quotes a fixed price after the blueprint. The AI agency cost survey and the guide on how to choose an AI agency cover the other terms worth comparing.