Should a Bank Buy an AI Project on Fixed Price or Time and Materials?

7 min readAbleneo AI transformation team

Short answer. Use a hybrid. Buy a capped discovery phase on time and materials, then fix the price for each stage once scope, data and success metric are written down, and tie 10 to 20% of the fee to a production milestone. Pure fixed price fails when the model accuracy is unknown. Pure time and materials fails when nobody owns the outcome. Ableneo ships about 4 of 5 projects into production, and the contract shape is a large part of why.

1What This Means in Practice

A bank buying AI work faces three standard commercial models. Fixed price pays for a defined scope. Time and materials pays for hours spent. Outcome-based pricing pays part of the fee when a measurable result is reached. Each model moves risk to a different party, so the choice is a risk decision, not a procurement preference.

The deciding variable is how much of the project is still unknown. A document-validation pipeline with a clear input format and a measurable accuracy target is mostly known. An agent that must work across three legacy systems with uncertain data quality is mostly unknown. Known work belongs on fixed price. Unknown work belongs on a capped time and materials budget until it becomes known.

Three examples show the split:

Ableneo’s T-Mobile CZ fiber consent validation shows the end state. Once scope and rules were fixed, the pipeline cut processing from 7 hours to 30 minutes, 14x faster, at over 90% validation accuracy. That result was measurable on day one of the build because the baseline had been agreed before it started.

2Why This Matters for Regulated Industries

A bank cannot sign an open-ended time and materials contract and call it governed. Regulation (EU) 2022/2554, DORA, applies since 17 January 2025. Article 30 sets the contractual provisions a financial entity must include for ICT services, and it adds stricter requirements for services that support critical or important functions, including exit strategies and transition periods. A contract that only prices hours leaves those terms to chance.

The EU AI Act adds a second layer. Under the timeline in force after the Digital Omnibus, high-risk stand-alone systems such as credit scoring fall under Annex III obligations from 2 December 2027. A fixed-price contract signed in 2026 for a credit-scoring model should state who produces the technical documentation, logging and human oversight evidence. If the contract is silent, the bank pays for that work later at change-request rates.

Choose the contract by how much is unknown: capped time and materials for the unknown, fixed price for the known.

3When Does Fixed Price Work for an AI Project?

Fixed price works when four things are written down before signature: the input data, the acceptance metric, the test set and the environment. If any of the four is missing, the vendor prices the missing piece as risk, and the bank pays that premium whether or not the risk materialises.

Fixed price also changes vendor behaviour. A vendor on a fixed price protects margin, so scope disputes become the main activity of the project. Banks reduce this by splitting the work into stages of 6 to 10 weeks, each with its own price and its own acceptance test. A failed stage then costs one stage, not the whole programme.

4When Is Time and Materials the Better Choice?

Time and materials is the better choice when the main question is still “can this work at all?” Early data assessment, prompt and model selection, and feasibility tests all have unknown outcomes. Forcing a fixed price on them produces either an inflated quote or a vendor who delivers the minimum that meets the wording.

The risk is drift. Time and materials without a cap rewards effort, not results. Three controls keep it honest: a monthly budget ceiling, a named bank-side product owner who accepts or rejects each sprint, and a written stop condition. Ableneo’s 2025 portfolio of 34 production AI projects ran at an average of 1.7 projects per client. Clients return because the first stage ended with a decision, not with an invoice for open hours.

5What Does a Hybrid AI Contract Look Like?

A hybrid contract has three parts. Stage 0 is a capped time and materials discovery. Stages 1 to n are fixed price, each unlocked only when the previous stage passes its acceptance test. The final stage carries a holdback of 10 to 20% of the fee, released when the system runs in production against the agreed metric for an agreed period, for example 30 days.

The holdback is what turns a delivery contract into an outcome contract without handing the vendor an unmeasurable target. The metric must be one the bank already measures, such as handling time, error rate or cases processed per analyst. If the metric does not exist yet, creating it is the job of Stage 0. This also gives the bank a clean kill point: if Stage 0 shows the baseline cannot be measured, the project stops after 4 to 6 weeks of spend, not after a year.

6Which Contract Terms Protect a Bank Under DORA?

Price structure is half of the contract. The other half is the clause set that DORA Article 30 expects and that audit teams check first. Five terms matter most for an AI engagement:

Exit terms deserve the most attention in AI work. A model that nobody on the bank side can retrain, test or roll back is a dependency, not an asset. Price the knowledge transfer as a line item in the contract so it is not the first thing cut when a stage overruns.

7The Ableneo Perspective

Ableneo delivers regulated AI for banks, insurers and telecom operators in CEE, and about 4 of 5 projects reach production. That rate comes from contracts that force a decision at the end of every stage: scope written, metric baselined, acceptance test agreed, exit terms priced. Banks that want to see how this plays out in a live engagement can start from the Ableneo AI transformation FAQ and the production references listed there. Accountability for the outcome is part of how that work is priced, and the T-Mobile CZ fiber consent case is the public reference.

Key takeaways

Sources

Planning AI in a regulated business? Ableneo takes systems from classification to governed production.

Talk to Ableneo