A Turnkey AI Project: When a Fixed Price Works and How to Hire a Firm to Take Your Pilot to Production

A stalled AI pilot can be finished as a turnkey job, with a working system handed over at the end. Start with a short fixed-price assessment that measures the pilot on your documents and sets the score it must reach. Raising the score to that target is work nobody can price in advance, so buy it as time and materials: hourly rates with a cap on the total. Fix the price of the rest, the work you can describe on paper, phase by phase.

What does “turnkey” mean for an AI system?

For an AI system, turnkey means the firm hands over a system that is ready to use, and the contract says what “ready” means. For example, the system runs inside your perimeter, on servers or cloud accounts your company controls. It reaches the target score on your documents and sends uncertain cases to a person. Your staff can run it after the firm leaves.

In construction, a turnkey contractor designs and builds the whole facility and hands it over ready to use. FIDIC, the International Federation of Consulting Engineers, publishes a standard contract for such projects, known as the Silver Book.

FIDIC's page for the 2017 edition names three cases where the Silver Book does not fit:

• Bidders lack the time or information to check what the client requires, do their own design and risk studies, and estimate the cost
• Much of the work is underground or in other places bidders cannot inspect, unless the contract covers conditions nobody could foresee
• The client plans to supervise or control the contractor's work closely, or to review most of the drawings

For these cases FIDIC recommends a different standard contract, its Yellow Book. A stalled AI pilot matches the first two cases until a firm has run your documents through it, because nobody has inspected the ground. The phased contract described below matches the third, since it keeps you in close control of the work. So read “turnkey” as the handover of a working system, and price the work phase by phase.

Why is a fixed price hard for AI work?

Many parts of a software project can be described before anyone builds them. A report, or a connection to another system, fits on paper, and a firm can price what is written.

A system that reads documents with a language model has one part nobody can describe in advance: how often it reads each field correctly on your documents. A field is one piece of data, such as a date or an amount. The answer depends on your scans and templates, and on cases nobody remembers until they appear. Nobody knows how much work the target score will take until someone measures the system on your documents.

The US government's purchasing rules make this distinction too. Part 16 of the Federal Acquisition Regulation (FAR) covers types of contracts. It describes a time-and-materials contract as one paid at fixed hourly rates for the hours worked, plus the actual cost of materials. It allows such a contract only when nobody can estimate the work accurately at signing, or predict its cost “with any reasonable degree of confidence”.

Under the FAR, a fixed-price contract suits work with clear specifications, where a fair price can be set at the start. Part 16 also tells contracting officers to avoid long use of time and materials once experience provides a basis for firmer pricing.

So a fixed price for the whole job, quoted before anyone has run your documents, tends to carry a large margin for the unknown. The alternative is a list of exclusions, things the price does not cover, and each one hands a risk back to you. Look for both in the proposal.

Which stages suit a fixed price, and which suit time and materials?

This is not legal advice. Your lawyer writes the contract, and this article lists what it should cover. Split the work, and price each part by what is known when it starts:

• Start with an assessment at a fixed price. The firm reads the pilot's code and runs it on a sample of your real documents inside your perimeter, then writes down why it stalled. You receive a draft acceptance set, a group of your real documents with the correct answer for each, and a priced plan for the next phase. Both are yours whether or not you continue
• The middle phase, raising the score, is the uncertain one. It suits time and materials with a ceiling, a cap on the total the firm may bill. The firm tries different prompts (the written instructions the model follows) and models against the acceptance set and shows you the scores every week. You can stop when progress stalls
• Once the target score is reached, the rest is ordinary engineering, priced as fixed phases. It covers installing the system inside your perimeter, alerts when quality drops, the review queue where staff check uncertain cases, and moving your team onto the new system. Each fixed phase ends with an acceptance run, a test that scores the system on the acceptance set and decides whether the phase is paid
• After launch, running the system fits a monthly fee with a written list of what it covers, because someone has to watch quality and handle model upgrades

The FAR also requires a time-and-materials contract to include “a ceiling price that the contractor exceeds at its own risk”. Put a ceiling in your contract too. It caps what you pay, but under the FAR's own payment clause the firm does not have to keep working once the ceiling is reached, unless you raise it. Agree in advance what happens then: you stop, raise the ceiling or re-plan the work.

A firm may offer the whole sequence as one contract and call it turnkey. That can work when every phase has its own price and acceptance run, and you may stop between phases.

What should “done” mean in the contract?

Done means the system passes its acceptance run, scored against a pass mark agreed before the price is fixed. The contract should name:

• Which documents are in the acceptance set, how many of each type, and who wrote down the correct answers. The set stays locked once the phase is priced
• The pass mark for each field and each document type, which together make up the target score, and the fields that always go to a person whatever the score
• Where the acceptance run happens and who does it. It happens inside your perimeter, on the exact version the firm will hand over. Your staff carry it out, or watch the firm do it
• What happens if the system fails its acceptance run. The firm gets a set period to fix it at its own cost. After that, you may stop and keep what was built

Keep part of the acceptance set from the firm until the acceptance run. If the firm has seen every test document, it can adjust the system to those exact documents, and the score will overstate how well it reads new ones. The article on AI pilots that never reached production, linked above, explains how to build the set.

What can break a fixed-price AI contract later?

Agree in advance how these events change the price:

• New document types or templates that are not in the acceptance set. Treat them as changes, priced the way the phases are
• The provider retiring the model the system uses. OpenAI's deprecations page, for example, sets at least six months' notice for a model in general release and at least three months for special versions of one. Preview models can be retired at much shorter notice, and safety or compliance concerns can shorten any of these periods. Say who pays to test the replacement on the acceptance set and to switch to it
• Delays on your side, such as slow approval of the firm's access to your systems, or late delivery of the documents and correct answers the firm needs. A fixed price assumes your part arrives on time, so list what your side must deliver, and by when
• A change in your own rules, for example your risk team deciding that a class of documents may no longer go to an outside model. Those documents then have to be processed inside your perimeter, which changes the work and its price

Running costs sit outside the build price unless the contract says otherwise. Model usage fees, or the cost of the servers that run the model, grow with the number of documents. Agree who pays them during the build and after launch, and how many documents a month the price assumes.

How do we tie payments to results?

Tie each fixed-price payment to the system passing its acceptance run inside your perimeter, and do not make any of those payments due on a date alone. For the time-and-materials phase, you pay for hours up to the ceiling, and the weekly scores tell you whether to go on.

Hold back part of the final payment until the system has run in production for an agreed period without falling below the target score. Agree the share and the period before you sign.

Keep the right to stop after any phase and take everything built so far, including the code, the prompts, the acceptance set and the scripts that score it. The article linked below covers how code ownership works in a contract.

How do we check a firm that offers to finish our pilot?

Write a one-page brief that says what the pilot does, which documents it reads and how many a day, where the data may go, and the score you need. Send it to every firm on your shortlist with the same questions:

• What do you need to see before you can fix a price, and for which phase would you fix it?
• Show us how a past contract for similar work was structured, with prices and the client's name removed
• What happens if an acceptance run fails at the end of a phase?
• Who on your team has taken an AI system to production, and will that person work on our project?
• Can your engineers work inside our perimeter, under our access rules?

The article on outsourcing AI development without your data leaving, linked above, covers the last question in detail.

What in a proposal should make us wary?

• A price for the whole job arrives before the firm has seen a single document of yours
• The proposal promises a percentage of correct answers but does not say on which documents it will be measured
• The word “turnkey” appears with no acceptance run and no right to stop between phases
• Running costs are missing, or appear only as “to be agreed”
• The draft contract lists the prompts or the acceptance set as the firm's property

Where does amBrain fit?

The one language model project amBrain describes in public is this: “We have taken an LLM integration to production inside a client's FinTech perimeter: extracting and normalising unstructured broker and venue notices — corporate actions, instrument and margin changes — into structured records the trading system consumes.”

This article is not a case study. It does not say that the project above began as another team's stalled pilot. The payment models above, turnkey and fixed price included, are general advice, and the article does not say whether amBrain offers any of them. It gives no amBrain prices or timelines.

amBrain takes over projects that stalled with another team and brings them to production.

amBrain has been building software since 2019. On ownership, amBrain's line is: “The client keeps full ownership of the product and the code, except our reusable components.” Ask amBrain for the list of those components by name before you sign, as you would ask any other firm.

If your pilot has stalled, send amBrain the same brief and questions you send every other firm, and compare the answers side by side.

Common questions

• Our pilot already has an accuracy figure. Why pay for an assessment? If the pilot was tested on made-up data or on documents someone cleaned by hand, its figure says little about your real documents. The assessment measures the pilot on a sample of them inside your perimeter, and the draft acceptance set it produces is yours either way
• Should we hire our own ML team instead? If AI is the product you sell, probably yes, and the article linked above on hiring an ML team covers that choice. If the system is a tool inside your operations, a firm can finish the pilot while you hire the people who will run it
• Do we have to show our documents to every firm on the shortlist? No. The brief describes your documents without including them. Only the firm you pick for the assessment sees real documents, and it works inside your perimeter