The six documents your clients will start asking you for

The same sequence comes up with some regularity. Ahead of a renewal, a long-standing client sends over a compliance questionnaire prepared by its adviser or procurement team: the system’s technical documentation, access to logs, an incident-notification commitment, a document explaining how data is processed, a liability clause and instructions for use. Six documents. In most cases the supplier has none of them in presentable form, gives an interim response, and a negotiation that should have taken two weeks stretches to two months.

It is worth understanding why this is happening. It is not a whim of the client’s adviser, and it will not go away.

Why the request is arriving now

Regulation (EU) 2024/1689, the AI Act, does not impose the same obligations on everyone. The most demanding part concerns high-risk systems. Regulation (EU) 2026/1744 moved its application to 2 December 2027 for the systems listed in Annex III and to 2 August 2028 for systems that are safety components of regulated products. The postponement has persuaded many businesses that there is time. The opposite is true, for a purely contractual reason: agreements signed today will still be in force after those dates, and a client signing now does not want to find itself, two years from now, tied to a supplier that cannot support its compliance.

Public-sector buyers are moving first, and not out of virtue. Article 30 of Italian Legislative Decree no. 36 of 31 March 2023, the Public Contracts Code, already requires a contracting authority purchasing automated solutions to secure access to the source code and the documentation needed to understand their logic. The Agency for Digital Italy, AgID, began the process for guidelines on the development and procurement of AI systems in the Italian public administration with determination no. 43 of 10 March 2026. The public consultation closed on 11 April 2026, and the text is awaiting opinions from the Unified Conference, the Italian Data Protection Authority and the Ministry of Enterprises and Made in Italy. Once adopted, those guidelines will tell authorities what to ask you for. In the private sector, the pressure comes from Article 26 of the AI Act. From 2 December 2027, deployers of high-risk systems will have to entrust human oversight to competent people, retain logs for at least six months and, in the workplace, inform workers’ representatives in advance. Clients cannot meet those duties on their own: they need things that sit on your side of the contract.

The six documents

They do not all carry the same weight. That is the first point to make clear when the questionnaire arrives: some are already legal duties; others are matters for negotiation. Confusing them weakens your position on both fronts.

1. Technical documentation. Architecture, the models used including third-party models, the nature and provenance of training data, expected performance, known limitations and version history: this is the content described by Article 11 and Annex IV of the AI Act. It is required by law only for high-risk systems and from 2 December 2027. Even so, it remains the document that an Italian public-sector client must require under Article 30 of the Public Contracts Code, and the one on which everything else is built.
2. Logs. Automatic event recording and client access in a readable, exportable format. For high-risk systems, Article 12 of the AI Act requires the system to record events throughout its lifetime, while Article 19 places retention on the provider, but only to the extent that the logs are under its control. That limit creates the client’s more awkward question: if the logs remain on your servers, how can the client retain them for six months as Article 26 of the AI Act requires? This is an architecture issue, not a matter of drafting clauses, and it is better resolved before it appears in a tender.
3. Incident notification. There are two separate tracks. If you process personal data on the client’s behalf, you are a processor, and Article 33(2) of Regulation (EU) 2016/679, the GDPR, has required you since 25 May 2018 to notify the controller without undue delay. Better-prepared clients will specify a period in hours because they have 72 hours to notify the Italian Data Protection Authority. A malfunction of the AI system is a different matter. It is regulated as a serious incident by Article 73 of the AI Act from 2 December 2027 and only for high-risk systems. For now it is contractual territory, and clients will want it to cover systematic output errors, not only service outages.
4. The document explaining how data is used. This is the data processing agreement required by Article 28 of the GDPR: purposes, categories of data, an up-to-date list of sub-processors, processing locations and transfers outside the Union, security measures and audit rights. It is the only one of the six that is already required in full, and its absence is not a negotiating point: it is non-compliance. Then there is a question now asked almost every time, which standard contracts do not address because they were written before it mattered: are users’ data and conversations used to train or improve your models or third-party models?
5. Assumption of liability. As phrased in questionnaires, this means nothing. You should break it down rather than simply accept it. There are three separate matters: a warranty that the system complies with current law, together with a commitment to update it at your own cost as duties arise during the contract; an indemnity for third-party claims and penalties caused by defects attributable to you; and advance notice of material changes, such as switching the underlying model or a sub-supplier. No rule mandates these terms. They are the price of entry for certain clients and should be treated as a commercial decision, with a liability cap proportionate to the contract value.
6. The user guide. These are the instructions for use under Article 13 of the AI Act: what the system can and cannot do, excluded uses, how human oversight is exercised and how unreliable outputs can be recognised. This obligation also starts on 2 December 2027 for high-risk systems, but clients need the guide now for two of their own duties that are already in force: staff training under Article 4 of the AI Act since 2 February 2025, and transparency under Article 50 since 2 August 2026. The division of responsibility for the latter must be put in writing because it is asymmetric: making it clear to a person that they are interacting with an AI system is your design obligation; disclosing artificially generated text published to inform the public falls to the client.

The point almost nobody checks

One question comes before all six: in the language of the AI Act, what are you?

Article 25(1) provides that a distributor, importer, deployer or any other third party is considered the provider of a high-risk system, with all the resulting obligations, in three cases: where it places its name or trade mark on a high-risk system already on the market — point (a) preserves contractual arrangements that allocate the obligations differently, making the clause something to negotiate rather than simply accept —; where it makes a substantial modification that leaves the system high-risk; or where it changes the system’s intended purpose so that a system that was not high-risk becomes high-risk.

That description matches a sizeable part of the Italian market: a company takes a third-party general-purpose model, integrates it into a product bearing its own brand and specialises it for a use case. The company sees itself as an integrator, but if the use case falls within Annex III it is the provider. The resulting duties are not three extra documents. They are the provider obligations listed in Article 16 of the AI Act: the quality management system under Article 17, conformity assessment under Article 43, the EU declaration of conformity under Article 47 and CE marking under Article 48 of the same Regulation. That is a difference in scale, and it needs to be checked now.

There is an orderly way out, and it must be documented. Article 6(3) of the AI Act allows a system listed in Annex III to be treated as not high-risk where it performs a narrow procedural task, improves the result of a previously completed human activity, detects decision-making patterns without replacing human assessment, or performs preparatory tasks. The exception never applies to systems that perform profiling. Anyone relying on it must document the assessment before placing the system on the market and keep it available to the authorities, as paragraph 4 of the same Article requires. That paragraph also makes the system subject to the registration requirement in Article 49(2): the provider and the system still go into the EU database. The way out is orderly, not silent. In practice, that documented assessment is a seventh document, and it is the first one an informed client will ask to see.

If, instead, you supply components that become part of someone else’s high-risk system, Article 25(4) requires a written agreement specifying the information and technical access that provider needs in order to comply. This, too, is not negotiable: it is the law.

Where to start

The costliest mistake is producing all six documents for one client while a tender is under way. Written once as a stable part of the product, they save weeks of work. Reconstructed under deadline, they cost several times as much and are poorer. The sensible order follows legal certainty, not difficulty: first, the Article 28 GDPR agreement, which is already required, and the written answer on the use of data for training; next, classification under Article 25 and, where necessary, the documented assessment under Article 6(3), because everything else depends on them; finally, the technical documentation and instructions for use, which are the most demanding but also the only ones that can be reused across every contract once they have been written.

To understand the form the requests will take, look at the model contractual clauses for the procurement of artificial intelligence published by the European Commission’s Public Buyers Community and updated in March 2025. There is a full version for high-risk systems and a lighter one for other systems. They are voluntary and intended for public-sector buyers, but they are the text your clients’ advisers are reading. It is better to read them before they appear as an annex to a tender specification.

The other side of this note — what to ask the supplier when your company is buying the AI — is covered in the note on contracts with AI vendors.