Fixed price vs time and materials for AI projects

Fixed price vs time and materials is a choice about who carries the risk of the estimate being wrong. Fixed price puts it on the vendor and works when the scope is locked; time and materials puts it on the buyer and works when nobody can lock the scope yet. AI projects need both, in sequence: a short fixed-price sprint to discover or prove, then a fixed-price build against the scope that sprint produced. The failure mode is a fixed price agreed before anyone knows what the model can do, which ends in either a padded quote or a dispute. This article sets out when each model fits and how to write an AI project contract that avoids the trap.

The two software pricing models, and the third

Time and materials bills hours at a rate; you pay for effort and control the direction week by week. Fixed price bills a sum for a defined outcome; the vendor absorbs overruns and you give up the freedom to change scope without a change order. The third model, which is what most good AI work uses, is staged fixed price: each stage is small enough to estimate honestly, each produces the information needed to price the next, and you can stop after any stage. That is how our programs are structured.

Why AI makes plain fixed price dangerous

Conventional software can be specified before it is built: the screen does this, the report shows that. An AI feature cannot be fully specified in advance because its most important property, accuracy on your data, is unknown until it is measured. A vendor asked to fix a price for "an agent that resolves support tickets" before seeing the tickets has two options: pad the quote to cover the unknown, or quote low and argue about what "resolves" means later. Neither serves you. The scope lock for an AI project has to include the evaluation method and a measured baseline, and those come from a sprint, not from a proposal.

Why T&M drifts on AI projects

Time and materials fits research because the direction changes as you learn. It fails on AI product work because learning never stops: there is always another prompt to tune, another model to try, another edge case. Without a fixed outcome and date, an AI project on T&M becomes a standing team with a standing invoice and a demo that is nearly ready. The tell is a project that has been "two weeks from production" for a quarter. The difference between a proof of concept and a demo is described in AI proof of concept vs demo.

The staged model in practice

Stage one: discovery, fixed price, ten days

A Sprint Zero discovery at $3,250 / ₹2,00,000 looks at the real data, the real systems and the real users, and produces a scope, an integration list, an evaluation plan and a fixed quote for the build. The fee is credited to the build, so it costs nothing extra if you proceed, and it costs very little if you learn the project should not proceed.

Stage two: proof, fixed price, three weeks

Where feasibility is genuinely uncertain, a ProofRun POC sprint at $6,250–10,500 builds a working slice against a graded evaluation set from your data and reports the measured accuracy and cost per request. That measurement becomes the acceptance target in the build contract. This is the stage most vendors skip, and the stage that makes the fixed-price build honest.

Stage three: build, fixed price, fixed date

With the scope and the measured target in hand, the build is a normal fixed-price contract: Launch 6 at $26,500–45,500 over six weeks, or a service scope from the pricing page. Changes go through change orders, but because discovery did its job, they are rare. How this works end to end is set out in fixed-price AI development.

What a good AI project contract contains

• A scope written as user outcomes and system boundaries, not as a feature list
• An evaluation method: the test set, who grades it, and the metric
• An acceptance target set from a measured baseline, not a vendor claim
• A fixed date and what happens if it slips
• Shadow-mode and go-live criteria, including which actions stay human
• A running-cost estimate and who pays for inference during the build
• Ownership: code, prompts, evals, models, infrastructure and documentation transfer to you
• A change-order process with prices, so scope changes are decisions rather than arguments

When time and materials is still right

Genuine research, where the question is whether something is possible at all and the answer changes the direction weekly. A standing team embedded with yours on a roadmap you control. Maintenance and small changes after launch, which is what a care plan is: a monthly retainer with a fixed number of hours and a response time, so you get T&M flexibility inside a fixed budget. Martin Fowler's note on fixed-price contracts makes the general case for why fixed scope and fixed price sit uneasily together in software; staged fixed price is the practical answer for AI work.

A worked example

A hospital network wanted a voice agent to handle appointment calls in several Indian languages. A plain fixed price would have required promising recognition accuracy in Kannada and Tamil on hospital phone lines before anyone had heard a call. Instead, discovery mapped the call types and the systems, a POC sprint measured transcription and intent accuracy per language on recorded calls, and the build was then priced fixed with acceptance targets per language taken from that measurement, and with certain call types routed to staff regardless. The contract had a date, a target and a definition of done that both sides had already seen evidence for. The multilingual voice agent case study describes the deployment.

Team and timeline

The staged model needs, on your side, a decision-maker who can approve or stop at each gate within a few days, and a subject-matter owner who helps grade the evaluation set. On ours, a solutions lead through discovery, then the build team sized to the scope. Elapsed time from first meeting to production for a typical agent or copilot is discovery in ten working days, an optional three-week POC, and a six-week build, followed by shadow mode. Every stage is fixed price and the total is known before the build starts. Care Plans from $1,000 / ₹68,000 a month take over after launch. All prices are on the pricing page.

Before you start: a checklist

• Decide whether your project is research or a build; only research suits open T&M
• Insist on a discovery stage before any fixed build price
• Ask how accuracy will be measured and who provides the test data
• Agree the acceptance target only after a measured baseline exists
• Get the date, the change-order price and the ownership terms in writing
• Ask for the running-cost estimate as part of the quote
• Confirm what happens at each stage gate if you decide to stop
• Check the care plan terms before signing the build, not after

Questions clients ask

• What if the POC shows the target is not reachable? You stop, having spent the sprint fee rather than the build budget. That is the point of staging; a negative result early is a good outcome.
• Can we change scope during a fixed-price build? Yes, through a priced change order. Because discovery locked the scope against real data, changes are usually small and rare.
• Does fixed price mean the vendor cuts corners to protect margin? It can, which is why the contract fixes the evaluation method and the acceptance target. A vendor who agrees to be measured has little room to cut.
• Who pays for model inference during the build? Agree it up front; we include it within the sprint and build prices and hand over a measured running-cost figure at the end.
• Is the total cost known before we start? After discovery, yes. Before discovery, only the discovery fee is fixed, which is the honest position.

Glossary

• Time and materials (T&M): billing by hours at a rate; the buyer carries estimate risk
• Fixed price: a set fee for a defined scope and date; the vendor carries estimate risk
• Staged fixed price: a sequence of small fixed-price stages, each informing the next
• Scope lock: the agreed boundary of the work, including the evaluation method
• Change order: a priced, agreed change to a locked scope
• Acceptance target: the measured accuracy or outcome the build must reach
• Stage gate: a decision point where the buyer may proceed, change or stop

Related reading

See fixed-price AI development, AI proof of concept vs demo and what a six-week AI MVP actually contains. Martin Fowler's FixedPrice note is the classic statement of the underlying tension.

Fix the price once the scope and the measurement are known, keep T&M for research and care, and never sign a fixed price for accuracy nobody has measured.