Why fixed-price programs de-risk your first AI project

Fixed price AI programs exist because the first AI project in a company carries a specific kind of risk: nobody yet knows whether the idea is feasible, what it will cost, or whether the data is good enough. A fixed-price discovery answers those three questions for a known fee that is credited to the build, so the first step never costs twice. The programs that follow each have a fixed scope, price and end date, which puts the risk of estimating the work onto the vendor and gives you a clean decision point after every stage.

This article explains where the risk in a first AI project actually sits, how a staged fixed-price structure contains it, what fixed price does not protect you from, and how to tell a genuine fixed-price program from an hourly engagement with a cap.

Where AI project risk comes from

Software projects fail on scope and estimation. AI projects add three failure modes that are specific to the technology. The data may not support the use case: documents are inconsistent, labels are missing, or the signal is weaker than everyone assumed. The model may not reach the accuracy the workflow needs, and no amount of prompting will close the gap. And the organisation may not be able to act on the output: the system produces a correct recommendation nobody is allowed to use.

Time-and-materials engagements discover these problems slowly and expensively, because the incentive is to keep working. Fixed-price programs are designed to surface them early, at a known cost, with an explicit decision at the end of each stage. The reasons AI pilots never reach production are almost all visible by the end of a properly run discovery.

How a staged fixed-price structure contains risk

The discovery fee is credited, so it never costs twice

The most common objection to a paid discovery is "we are paying to be sold to". Crediting the fee to the build removes that: if you proceed, discovery was part of the build price; if you do not, you paid a modest fixed sum for a feasibility answer, a data assessment and an architecture you can take to any vendor. Either way the money bought something. The AI discovery sprint describes what the ten days contain.

Fixed dates matter as much as fixed prices

A fixed price with an open-ended timeline is only half a commitment. A fixed date forces scope discipline on both sides: the vendor has to say what will not be in the build, and the client has to make decisions on time. Six weeks for Launch 6 is not a target; it is the deadline the price is built on.

What fixed price moves onto the vendor

Estimation risk. If the vendor has misjudged the effort, that is the vendor's problem, not a change request. This only works when the vendor has done the shape of work many times before, which is why fixed-price programs tend to have standard shapes: a retrieval assistant over documents, a support agent on a helpdesk, a copilot in a SaaS product, a voice agent on a phone line. A vendor offering a fixed price for something they have never built is guessing, and you will pay for the guess in quality.

Fixed price also moves the incentive. Under hourly billing, a problem discovered late is revenue. Under fixed price, a problem discovered late is loss, so the vendor's tooling is built to find problems early: evaluation sets before prompts, data sampling in week one, shadow mode before autonomy. The engineering practices and the commercial model reinforce each other.

What fixed price does not protect you from

Honesty requires listing the limits. Fixed price does not protect you from choosing the wrong use case, which is why discovery exists. It does not protect you from your own organisation failing to adopt the result, which is a change-management problem no vendor can fix alone. It does not eliminate running costs: inference, hosting and support continue after launch and should be forecast during discovery. And it does not remove the need for your own people to attend demos and make decisions on time; a fixed date is a commitment from both sides.

It also does not mean scope is infinite. A fixed-price program has a scope lock: the deliverables are agreed in writing at the start and changes are traded, not added. Clients sometimes read this as inflexibility. It is the opposite: it is the mechanism that makes the price and date possible. Scope lock explains how trades work in practice.

Spotting a real fixed-price program

• The scope is written down as deliverables, not as a list of activities
• The price is a number, not a range that resolves to hourly billing
• The end date is a date, and there is a plan for what happens if it slips
• Evaluation criteria for "done" are agreed at the start
• Change requests are traded against existing scope rather than billed
• Everything built is owned by you on payment
• The vendor can describe the last three times they ran the same program

Compare this with a "fixed price" that is really a capped estimate with hourly billing beneath it, where the cap is reached and the conversation turns to extensions. The fixed price vs time and materials comparison goes deeper on the difference.

A worked example

A direct-to-consumer brand wanted personalised recommendations and a WhatsApp agent for order queries. Their leadership was split: the marketing lead was confident, the finance lead had seen a previous AI pilot consume a budget without shipping. They ran Sprint Zero. Discovery found that the event data needed for personalisation was patchy but fixable, that the WhatsApp agent had a well-defined set of intents with clean order data behind them, and that the two should be sequenced rather than built together. The finance lead approved the WhatsApp agent as a Launch 6 with a fixed date, with the discovery fee credited, and the personalisation work as a second program once the event pipeline was repaired. The personalisation and WhatsApp agent case study describes the result. What made the approval possible was not confidence in AI; it was that each decision had a known price and a known end.

Team and timeline

A first AI project usually runs as Sprint Zero followed by either ProofRun or Launch 6, depending on whether the hardest technical question is already answered. Sprint Zero is staffed by a lead engineer and a solutions architect for ten working days and needs two to three hours a week from your side. Launch 6 adds one or two applied-AI engineers and, where there is a user interface, a designer, with a weekly demo you attend. The Sprint Zero page and the pricing page list current figures in USD and INR. To scope a first project, contact us with a paragraph on the problem and who would use the result.

Before you start: a checklist

• Write the use case as a sentence: who does what, with which data, to what end
• Name the decision-maker who will approve or stop the project after discovery
• Gather a sample of real data for the discovery team, with access agreed
• Agree in advance what accuracy or outcome would make a proof of concept a pass
• Ask the vendor for the deliverables list and the scope-lock process in writing
• Budget for running costs and a care plan, not only the build
• Set aside time for weekly demos from your side
• Confirm ownership terms before the first invoice

Glossary

• Fixed-price program: a defined scope, price and end date, with changes traded rather than billed
• Credited discovery: a discovery fee deducted from the price of the build that follows it
• Scope lock: the agreed deliverables list that governs what a program contains
• Feasibility answer: a written statement of whether the use case can work, with the evidence
• Data assessment: a review of the real data's quality, coverage and access during discovery
• Exit point: the end of a stage, where the client may stop with a complete deliverable in hand

Related reading

See fixed-price AI development: how it works and when it fits, the AI proof of concept vs demo distinction and our AI strategy services. For a general treatment of why estimation fails on novel software work, Joel Spolsky's writing on evidence-based scheduling remains a clear explanation.

Pay a known fee for a straight answer, then a known price for a dated build; that is how a first AI project stops being a gamble.