AI Services Contract Scope, Data Security and Outcomes

An AI services contract should define exactly what the provider will build or operate, what data it may use and how it must protect that data, and how the parties will measure results. It should also set out responsibilities, testing and acceptance, change control, ownership, incident response, and what happens when the service ends.

Define the service and its boundaries

“AI services” can mean a custom model, a chatbot built on a third-party model, document processing, recommendations, or ongoing model operations. A contract should describe the actual service rather than rely on a broad label. Identify the intended users, business process, supported channels, and environments, such as development, testing, and production.

Explain what is included and excluded. For example, a project might include a support-ticket classifier and integration with a specified ticketing system, but exclude writing customer responses, migrating old records, or providing 24-hour human support. State whether the provider is responsible for model selection, prompt and workflow design, integrations, hosting, monitoring, maintenance, and user training.

Useful scope details include:

Deliverables and their format, such as source code, configuration, documentation, evaluation reports, or a deployed service.
Milestones, dependencies, and the party responsible for each task.
Supported languages, data types, user volumes, interfaces, and operating environments.
Known limitations, excluded use cases, and assumptions about the customer’s systems or data.
Service hours, support channels, and maintenance responsibilities after launch.

Attach a statement of work or technical schedule when the details are too extensive for the main agreement. The documents should identify which one controls if their terms conflict.

Specify data use and security

Describe the information the provider may access, including whether it contains personal, confidential, regulated, or commercially sensitive data. Specify permitted purposes: for example, processing submitted records to provide the contracted classification service. Do not assume that permission to process customer data also permits the provider to use it for unrelated analytics, product development, or model training.

Address the full data lifecycle: collection, transfer, storage, access, use, retention, deletion, and return. The contract should say whether prompts, outputs, logs, and feedback are stored; where they are stored; how long they are retained; and whether they are shared with model or hosting providers. Name or describe subprocessors and require appropriate controls when they handle data. Include a process for notifying the customer of changes to those providers.

Security terms should be concrete and proportionate to the information and service. They commonly cover access controls, authentication, encryption in transit and at rest, staff confidentiality and training, vulnerability management, backups, and security assessments. Require the provider to notify the customer promptly of a suspected or confirmed incident affecting contract data, provide relevant information, cooperate with investigation and response, and preserve appropriate records. Set a notification deadline that can realistically be met and does not conflict with applicable obligations.

For sensitive data, define restrictions on copying it into public tools or unapproved services. Clarify who handles customer notices, user requests, and regulatory inquiries if personal information is involved. Contract language should be reviewed against the parties’ actual roles and applicable law; a generic promise to “comply with all laws” does not explain operational responsibilities.

Set measurable outcomes and acceptance tests

Agree on what success means before work begins. A statement such as “improve productivity” is not a usable acceptance standard on its own. Translate desired outcomes into tests and metrics tied to the intended use, such as classification accuracy on an agreed test set, response time under specified conditions, successful completion of defined workflows, or a reduction in manual review effort measured using an agreed baseline.

Define how measurement works: dataset and sample size, evaluation method, test environment, thresholds, exclusions, and who supplies the data. For a document-extraction system, the parties might test a representative set of forms and assess whether specified fields are extracted correctly. They should state how errors are counted, including whether a missing field, an incorrect value, and a low-confidence result are treated differently.

Include a review period, a process for reporting defects, and a reasonable remedy such as correction, retesting, or an agreed change request. Distinguish acceptance of a deliverable from a guarantee of business results. Outcomes can depend on data quality, user adoption, changing input, and third-party services, so a provider may not be able to promise a particular revenue gain or error-free performance. Make any commitments explicit rather than implying them through aspirational language.

Allocate responsibilities, ownership, and change

List customer responsibilities, including providing lawful access to data, designating decision-makers, validating outputs, and making required systems available. Identify provider responsibilities, such as documenting assumptions, securing access, testing integrations, and reporting material limitations. Assign responsibility for human review when a decision has meaningful consequences. An AI output should not silently become the final decision if the agreed workflow requires a person to verify it.

Clarify rights in customer data, deliverables, pre-existing tools, and service-generated outputs. The agreement should distinguish customer materials from the provider’s reusable software, templates, and know-how, and explain what rights each party receives. Address whether the customer can retain or export outputs and configuration at termination. Avoid assuming that the customer automatically owns every output or that the provider can freely reuse everything involved.

AI systems and their dependencies change. Set a change-control process for new features, altered models, changes to data sources, and material updates to third-party terms. Require notice of changes that could affect performance, security, or compatibility, and specify when re-testing or approval is needed. A process for evaluating model changes is especially useful where stable behavior matters.

Plan for risk, continuity, and exit

Contracts should address service availability and recovery in terms appropriate to the use case. If the service is important to operations, define support response expectations, escalation contacts, backup or recovery arrangements, and any agreed service levels. State what happens if an external model or hosting service becomes unavailable, changes its interface, or withdraws access. A provider cannot control every third-party dependency, but it can explain its fallback and communication duties.

Set an exit process that covers notice, transition assistance, data export in a usable format, deletion of remaining copies, and confirmation of deletion where appropriate. Include any retention exceptions that apply, such as backups or records the provider must keep, and explain when those copies are removed. Consider whether the customer needs continued access to documentation, logs, or configuration to operate a replacement service.

Common contract problems include vague scope, success metrics without a test method, undefined training use, unlisted subprocessors, and no workable exit plan. Another misconception is that a highly accurate benchmark guarantees safe or reliable operation in every real-world setting. Performance depends on the evaluation data and operating conditions, so testing should reflect the intended workflow and continue after launch where risk warrants it.

Frequently Asked Questions

Should an AI provider be allowed to train on customer data?
Only if the contract clearly permits it and defines the data, purpose, safeguards, retention, and any limits on reuse. If training is not required for the service, state that customer data may not be used for it.

What should an AI acceptance test measure?
Measure performance against representative, agreed data and conditions. Define the metric, threshold, error categories, test period, and process for fixing and retesting failures.

Who is responsible for checking AI-generated output?
The contract should assign review duties based on the use case. It should identify when a qualified or authorized person must verify an output before it is acted on.

Does the customer own AI-generated outputs?
Not automatically in every arrangement. The contract should say what rights apply to outputs, customer inputs, provider tools, and any materials created during the work.

What happens if the underlying model changes?
Require notice of material changes and establish whether the provider must assess impact, rerun tests, seek approval, or offer a fallback if performance or compatibility is affected.

What should happen to data when the contract ends?
Specify return or export, deletion timelines, any justified retention exceptions, and how the provider confirms completion. Include transition support if the service is operationally important.

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.

More cards in this drawer

The whole Numbers and Powers drawer →