Fixed-price AI development: how it works and when it fits

Fixed price AI development means a defined scope, a fixed date and a fixed fee, agreed before the work starts, with anything outside that scope going to a separately priced backlog. It is the pricing model most buyers want and most AI vendors avoid, because the uncertainty in AI work sits with whoever carries the price. It can be done well, but only under conditions that this article sets out plainly: short programs, scope locked early, thresholds agreed in writing and a change process that both sides respect.

Here is how it works in practice, where it fits, where it does not, and what to ask a vendor who offers it.

Why an AI development pricing model matters more than usual

With ordinary software, the main uncertainty is how long a known thing takes to build. With AI, there is a second uncertainty: whether the thing can be built to the required standard at all, because model behaviour on your data is not known until tested. Time and materials pushes both uncertainties onto the buyer, who pays for every week of discovering them. Fixed price pushes them onto the builder, who then has an incentive either to manage the uncertainty properly or to pad the price and narrow the scope until the risk disappears.

The honest version of fixed cost software for AI does the former. It uses short, staged programs so that the uncertainty is retired step by step, and it prices each step only when the previous one has reduced the unknowns enough to price the next.

Fixed price, time and materials, and staged fixed price compared

How fixed price AI work is made to hold

Stages that retire uncertainty in order

A discovery sprint benchmarks models on your data and produces a fixed price for a proof of concept. The POC proves the system clears agreed thresholds and produces a fixed price for the MVP. The MVP delivers a working product to a fixed date. Each price is set with the knowledge the previous stage produced, so no one is guessing at feasibility when they quote a build. Skipping stages is how fixed-price AI projects fail: a six-week build priced before anyone has tested a model on the data is a guess with a signature on it.

Scope locked early, in writing

For an MVP the scope is one user segment, one workflow and one AI capability, written down and signed off by day five. Everything else, however good the idea, goes to the backlog. This sounds harsh and is the reason the date holds. The full argument is in Scope lock: the discipline that makes fast MVPs possible.

Thresholds, not adjectives

"Accurate" is not a deliverable. "Field-level extraction accuracy at or above the agreed threshold on the evaluation sample" is. Fixed price only works if done is defined by measurements both sides can run. The evaluation suite built in discovery and the POC becomes the acceptance test for the build.

A priced backlog instead of change orders

Change requests are inevitable and are not the enemy. The enemy is unpriced change absorbed into a fixed scope until the date slips. In our programs every new request goes to a backlog with a rough price against it, reviewed weekly with the sponsor, who can pull an item in by pushing one out or fund it as a v1.1. The backlog becomes the plan for the next phase rather than a source of friction.

When fixed price fits, and when it does not

It fits discovery, proof-of-concept and MVP work because those are short, bounded and end in a decision. It fits bounded modernisation where the existing system can be assessed before pricing. It fits any project where a board or an investor needs a number and a date they can rely on.

It fits poorly when the scope genuinely cannot be known: open-ended research, a product whose requirements are still being discovered with users, or a long-running team that will work on whatever is most valuable each month. For those, time and materials or a retained team is more honest, and pretending otherwise leads to a fixed price with a very wide scope clause. It also fits poorly when the buyer cannot commit a decision-maker to weekly scope reviews; a fixed price with no one to say no on the client side becomes a slow argument.

The trade-off between the two models is compared in detail in Fixed price vs time and materials for AI projects.

What to ask a vendor offering a fixed price

• What has to be true before you will quote a fixed price, and what do you do to make it true?
• How is done defined, and can we run the acceptance test ourselves?
• What happens to a request that is not in scope, and who decides its priority?
• Who owns the code, prompts, models and evaluation data at the end?
• What is excluded, in writing, and what are the assumptions the price rests on?
• What happens if the model does not clear the threshold: who carries that, and what is the exit?

Ask, too, how the vendor handles model changes during the program. Providers deprecate versions and release new ones on their own schedule, and a fixed-price build that depends on one specific model version is exposed. The answer you want is that prompts and evals sit behind a routing layer, so a model change during the build is a regression run rather than a renegotiation.

A vendor with good answers has done this before. A vendor whose fixed price has no stages, no thresholds and no backlog mechanism is offering a time-and-materials engagement with a cap, which is a different thing.

A worked example

A university wanted to modernise a student-records ERP that several generations of staff had built on. Two proposals quoted a fixed price for a rewrite, each with a scope clause wide enough to make the price meaningless. A third approach began with a paid assessment of the existing system, which found that most of it could be kept and wrapped, and that the part causing pain was a bounded set of workflows. That assessment produced a fixed price and a date for modernising those workflows, with a priced backlog for the rest. The work delivered on the date, and the backlog became a second phase the university chose to fund on the evidence of the first. The engagement is described in the university ERP modernisation case study.

Team and timeline

All of our programs are fixed price and fixed date. Sprint Zero, the ten-day discovery sprint, is $3,250 or ₹2,00,000, credited to the next build. ProofRun, the three-week POC, is $6,250–10,500 or from ₹4,00,000. Launch 6, the six-week AI-accelerated MVP, is $26,500–45,500 or from ₹17,60,000. ReCore, for legacy-to-AI modernisation over eight to sixteen weeks, is $31,500–105,000 or more depending on the assessment. Each is staffed by a lead engineer, one or two AI or product engineers and an architect, and each requires a named client sponsor for the weekly backlog review. After launch, a Care Plan from $1,000 a month covers maintenance. All current figures, in USD and INR with GST, are on the pricing page, and the broader cost picture is in How much does AI development cost in 2026?.

Before you start: a checklist

• Confirm the vendor will benchmark on your data before quoting a build
• Get the scope, exclusions and assumptions in writing, one page if possible
• Agree the acceptance thresholds and confirm you can run the evaluation yourself
• Name a sponsor with authority to say no at the weekly backlog review
• Agree the backlog mechanism: how requests are priced and pulled in or deferred
• Confirm ownership of code, prompts, models, infrastructure and documentation
• Ask what happens on a no-go at each stage and what you keep
• Check invoicing currency and tax treatment match your finance team's needs

Glossary

• Fixed price, fixed date: an agreed fee and delivery date for a defined scope
• Staged fixed price: a sequence of short fixed-price programs, each priced with the previous one's results
• Scope lock: the point, early in a program, after which scope changes go to the backlog rather than the plan
• Priced backlog: the list of out-of-scope requests, each with a rough cost, reviewed weekly
• Acceptance threshold: a measurable score that defines done
• Change order: the traditional, slower alternative to a priced backlog

Related reading

What a six-week AI MVP actually contains shows what a fixed-price MVP scope looks like in practice, and How to choose an AI development company: a 12-point checklist covers the wider vendor decision. Joel Spolsky's essay on evidence-based scheduling remains a clear primary source on why estimates need data behind them.

Fixed price is not a promise to absorb any change; it is a discipline that both sides keep, and when it is kept the buyer gets what no other pricing model offers: a date and a number that hold.