AI in a work contract and the statement of work

A work contract to deploy an AI system is not only deadlines and certificates. The statement of work and the contract need to close, in advance, where data is stored and processed, who else sees prompts and logs, how an incident is notified, whether a human is in the decision loop, and who pays for a model error. Law 243-FZ raises what the client expects on transparency and control — without these clauses, a dispute after a failure turns into “that is how the industry does it.”

This material is for orientation, not legal advice. Requirements for a particular system depend on the industry and the data involved — check 243-FZ, the personal-data rules, and industry standards on the date the contract is signed.

Before you sign a work contract for AI, build a draft with blocks on data and liability, or check the contractor’s template:

Draft a contract Check for risks

Why a statement of work for AI is wider than a classic IT work contract

In ordinary software the result is more predictable: a bug can be reproduced. A model’s answer drifts and depends on training data, the prompt, and external APIs. The client needs to understand the boundaries: what counts as a defect, what is an acceptable margin of error, and what is a security incident. More on the legal background is in AI law and contracts in 2026.

In practice the statement of work and the contract work as a pair: the statement of work describes system behavior and metrics, and the contract describes roles, data, subcontractors, and money when there is a failure.

Blocks worth putting in the contract and the annexes

1. Data and localization

Categories of data: personal data, trade secrets, logs, training samples
Place of storage and processing (including a requirement to locate it in the Russian Federation, if that applies)
A ban on using the client’s data to fine-tune someone else’s model without consent
The period and method for deleting or returning data after the work ends

2. Sub-processors and external models

A list of clouds, APIs, and LLMs that the traffic passes through
The procedure for approving a replacement sub-processor
Who contracts with the model provider, and in whose name the license stands
The contractor’s liability for the acts of persons it engages

3. Incidents, human control, model errors

The notice period for a leak, unavailability, or mass generation failures
Where a human-in-the-loop is mandatory (decisions with legal or financial consequences)
Criteria for a defect vs the model’s “probabilistic behavior”
A liability cap, exclusions, and insurance (if any)

Practical tip

In an “Architecture and data” annex, record the flow diagram: who the sender is, which API, where the logs are, and the retention period. When the model or the data-center region changes, update the annex in writing — otherwise “we have always hosted it this way” cannot be proved.

What to put in the statement of work

The purpose of the system and prohibited use scenarios
Quality metrics and the acceptance procedure (a test set, thresholds)
Requirements for logging requests and responses, and for the client’s access to logs
Escalation scenarios when the model hallucinates in critical functions
A rollback plan: switching off the AI module without stopping the main process

If the contractor sent a new version of the contract or the statement of work, compare it with the agreed version before you sign:

Compare two versions Check for risks

One-sided wording to look at first

Contractor templates often contain imbalances:

“Result as is” with no acceptance metrics — the client will not be able to refuse the certificate
A right to change sub-processors and the storage region without notice
A zero notice period for an incident, or only a “reasonable period”
A full disclaimer of liability for the model’s output on critical decisions
A right for the contractor to train its own models on the client’s data “to improve the service”

Look for these clauses in the text before the certificate: “as is” with no metrics, a change of storage region without notice, training a model on the client’s data. For a review of the file — check the contract.

A 10-minute checklist

The contract or statement of work states the data categories and the place of processing
There is a list of sub-processors and a procedure for replacing them
Fine-tuning third-party models on the client’s data without consent is prohibited
An incident notice period is written in
Human oversight is defined for critical decisions
The boundary of a model defect and the liability cap are clear

Short conclusion

A work contract for AI without blocks on data, sub-processors, incidents, and human control is a postponed dispute. Fix the architecture and the liability in the annexes before work starts: after an incident, “we will agree fairly” almost never works.

Build or check a work contract for AI in DocumentCHEK

Upload the contract and the statement-of-work annexes — you will get a review of one-sided terms on data, liability, and acceptance.

Check the contract Draft a contract

Need to record the differences between two versions of a statement of work? Compare versions online.

AI in a work contract: data, localization, and liability in the statement of work