Fixed price earns its place when you can freeze the specification and write the acceptance test before anyone starts work. In AI and data engineering that describes a minority of engagements, perhaps one in five, which is why so much of what gets signed as fixed price is really time and materials with a contingency buffer and a slower argument attached.
The choice is not about trust. It is about who prices the unknown, and what that insurance costs you.
What each contract transfers
A fixed-price contract moves estimation risk to the supplier. The supplier prices that risk, and since a supplier who guesses low twice goes out of business, the loading is not small: 20% to 40% on top of the same work sold on time and materials is a normal range for anything with real discovery in it. You pay that premium whether or not the risk materialises.
Time and materials leaves estimation risk with you. You get the unloaded rate, you see the burn weekly, and you can stop. What you give up is the comfort of a number you can put in a board paper in January and never revisit.
The same work, priced two ways
Take a 60-day data platform migration for a team of three.
| Line | Fixed price | Time and materials |
|---|---|---|
| Contingency inside the price | 20% to 40% | none |
| Cost of a scope change | change order, 1 to 3 weeks of negotiation | next sprint |
| Who sees the burn rate | supplier only | you, weekly |
| Supplier’s incentive | finish early | keep the account |
| Your exit | full amount, or a penalty | notice, usually 2 to 4 weeks |
The last row moves more money than the first. On a fixed-price engagement that is going badly you are committed to the whole figure, because the one direction a fixed price never travels is downwards. On time and materials with two weeks’ notice, a bad first month costs you a month.
Work that does price well fixed
Four shapes hold up:
- A lift-and-shift migration where source and target schemas are documented and acceptance is row counts and reconciliation.
- A defined integration: two named systems, a known interface contract, an agreed error-handling rule.
- A time-boxed assessment or audit whose deliverable is a document.
- A repeat of something the supplier has built before at comparable size.
Notice what they share. The acceptance criterion is mechanical, so somebody can run it and get a yes or a no.
Why AI work usually fails that test
Retrieval quality, extraction accuracy and agent reliability are not deliverables you can specify honestly in advance. You do not know what proportion of your questions your document set can answer until it has been chunked, indexed and run through an eval set. You do not know whether 82% extraction accuracy needs a different prompt, a different parser or cleaner source data until you have read the failures. A supplier asked to commit to “90% accuracy” for a fixed fee either prices the worst case or writes the acceptance test loosely enough that it means nothing. Both outcomes cost you.
Sizing this kind of work is a question of ranges, not a single figure. Our notes on the components that drive the size and cost of a retrieval system set out the variables that actually move the number.
Capped time and materials
Most sensible contracts land here. You buy time and materials at published rates, you set a spend cap per phase, and the supplier owes you a written warning at 70% of the cap with a reason attached. You keep the visibility and the exit. Finance gets a ceiling.
Two things make the cap real. Rates you can verify, which is why we publish current day rates by role and seniority, and a phase short enough to learn from. Six weeks works. Six months is a fixed price in disguise.
When time and materials is the wrong answer
If nobody on your side can give two to three hours a week to prioritisation and acceptance, time and materials will disappoint you. The model prices flexibility, and flexibility needs someone to exercise it. Buy a fixed-scope piece of work instead, or buy a managed outcome and accept the loading; the estimation-risk trade between managed services and augmentation is this same decision in a different wrapper.
Skip it too if your procurement rules demand a fixed total before a purchase order can exist. In that case split the work: a small fixed-price discovery that produces the specification, then capped time and materials to build against it.