Skip to content

Business and implementations20:06

AI implementation contract: eight clauses missing from standard IT agreements

The guide outlines what to include in a contract with an AI vendor: 30-day data export, auto-upgrade ban, 5% hallucination threshold, and liability capped at 12 monthly fees.

AI-generated image.

Listen to the article summary

Synthetic voice.
0:00
0:00

Advisory guide written from the contractor's perspective. It describes typical solutions and does not determine any individual case. In pre-implementation meetings at companies with 50–250 employees, one claim appears regularly: a standard IT implementation contract plus a GDPR addendum is enough for an AI rollout. This sounds reasonable. A standard IT contract regulates deadlines, acceptance, penalties, code ownership, and data processing. If a chatbot or document automation workflow is just "another system", why draft new provisions. In addition, the AI Act does not impose a separate contract format for low-risk systems.

The strongest counterargument

This argument has a solid foundation. A standard IT contract typically includes an acceptance schedule, contractual penalties, assignment of economic copyright to the code, a bug reporting procedure, and a data processing agreement under Article 28 of the GDPR. For an ERP or e-commerce implementation, this was usually sufficient. AI system vendors also offer templates with general provisions on security and restrictions on training models on client data. GDPR fines reach 4% of annual turnover, so the presence of a data processing addendum is often treated as adequate protection. Moreover, AI Act penalties, up to 15,000,000 EUR or 3% of global turnover, mainly apply to providers of high-risk systems, not a company deploying a customer order chatbot.

AI-generated image.

What this argument overlooks

A standard IT contract leaves four critical AI risk factors unresolved. First, prompts and configurations: whether they qualify as a protected work under copyright law depends on their creative nature, yet the template usually stays silent on who holds the economic copyright. As a result, the vendor may reuse the exact same prompt layer for a competitor after the contract ends. Second, model versioning: OpenAI and Anthropic deprecate models with a few weeks or months of notice, as detailed in their deprecation documentation, and a standard contract does not grant the client the right to block an auto-upgrade before acceptance testing. Third, data export: conversation logs and vector embeddings are not standard SQL tables, rendering the clause "vendor will return data" effectively meaningless. Fourth, hallucinations: liability for an "incorrect output" requires a defined threshold; otherwise, every inaccurate response turns into a dispute over quality. On top of that, the AI Act distinguishes between provider and deployer roles with distinct obligations, which standard IT contracts rarely address; starting 2.08.2026, chatbots must disclose AI interaction and label generated content.

Our position

In our practice, a production AI implementation contract includes eight clauses absent from a standard IT contract: rights to prompts and configuration (transfer of economic copyrights or an irrevocable, perpetual license with the right to modify and sublicense), 60-day notice of model deprecation and a ban on auto-upgrades without acceptance testing, data export within 30 days in JSON/CSV/JSONL format, backup deletion within 60 days, hallucination threshold ≤5% on the test set, vendor liability cap at 12 months of fees, escrow or notary deposit of prompts and code, and 3 months of post-exit assistance at the standard rate. In our case, during an email assistant deployment on a 500-message sample, we first received a model card with metrics, and acceptance tests on that sample determined the final version. Without these provisions, the client gets a working system, but no exit path and no liability threshold for what the system generates.

AI-generated image.

Recommended actions for this quarter

This quarter, before signing, check if the vendor will even discuss data export within 30 days and the ≤5% hallucination test. Refusing both does not kill the project, but clarifies which risks remain with the client. We would change our position under one condition: if the vendor provides a local open-source model instance with an audited model card and full access to weights, external API deprecation clauses become irrelevant. Until then, treat an AI implementation contract as a standalone agreement, not an addendum to an IT template.

Source materials

Sources consulted during research. The text above is our own; these third parties are not responsible for its content and have not authorised it.

No-obligation call

Schedule a workshop with Paul

Thirty minutes or a full workshop, choose what suits you. Book a slot directly in the calendar and get instant confirmation.

Paul Lazniak

HEXART Founder