What an AI Policy for a Preclinical CRO Has to Cover in the US and Germany

The binding part is shorter than most people expect

When I am asked what an AI policy for a preclinical CRO has to cover, I start by separating what the law requires from what clients and auditors will ask for. On the EU side, a company that only deploys general tools such as Microsoft 365 Copilot carries a light load under the AI Act. Article 4 on AI literacy has applied since 2 February 2025, and the Digital Omnibus, which Lewis Silkin reports entered into force on 27 July 2026, softened it to a duty to take measures that support literacy. In practice that means role-based training and a record that it happened.

The Article 5 prohibitions apply to every deployer as well, and the one that matters at work is the ban on emotion recognition in the workplace. Duties for the general-purpose models themselves sit with Microsoft and the model providers. As long as you stay out of the high-risk categories, the AI Act asks your policy for a list of prohibited uses and evidence of training, and little else.

Draw the GxP line first

The part of the policy that needs the most care is the boundary around regulated work. FDA's GLP rule covers nonclinical laboratory studies that test articles for safety, and the definition in 21 CFR 58.3 excludes basic exploratory studies. So most discovery, kinase profiling and efficacy pharmacology falls outside Part 58, while GLP toxicology sits inside it, and under section 58.33 the study director stays responsible for the conduct, interpretation and reporting of the study.

Part 11 comes into play when software creates or changes electronic records that a predicate rule requires, and FDA's 2003 scope and application guidance reads that narrowly. A Copilot draft of an internal email is outside it, while an AI step that writes values into a GLP report would be inside. The policy line I recommend allows no AI generation or alteration of GLP or GMP raw data, study records or reports unless it runs through a validated and documented process. The paperwork around the bench is a different matter, and I covered where AI helps with quotes, protocols and reports in an earlier post.

Organise the policy by data, then by tool

My view is that an AI policy written tool by tool goes stale within a quarter, because the tools and their names keep changing. A policy written around classes of data lasts much longer. A CRO usually needs public, internal, confidential and client restricted classes, with GxP records and personal data flagged separately, and for each class the policy names the approved AI tools that may touch it.

Client restricted is where most of the effort goes. Sponsor contracts increasingly carry AI clauses, such as a ban on training with their data, limits on retention, disclosure of AI use or a flat exclusion of certain uses. If each class maps to a sensitivity label in Microsoft 365, the policy and the technical control describe the same thing, and you can show an auditor both. Data exchanged with sponsors adds its own complications, which I described in a post on sponsor and CRO data handoffs.

Keep HR uses out of scope

The AI Act treats several employment uses as high-risk under Annex III point 4: recruitment and candidate screening, decisions on promotion or termination, task allocation based on personal traits, and monitoring or evaluating performance. After the Omnibus, obligations for these stand-alone systems apply from 2 December 2027, and a deployer must inform workers' representatives before use under Article 26(7). The cleanest policy line is that general AI tools are not used for any of these purposes, and any proposal to do so goes through a separate review first.

The same line helps in Germany, where the works council will ask exactly this question. On the US side, AI in hiring can fall under state and city rules, so if anyone wants to screen CVs with AI, take it to employment counsel before it starts.

The German steps set the timeline

Two German processes usually decide when a rollout can begin. The German data protection conference says in its May 2024 guidance on AI and data protection that a data protection impact assessment will often be needed for AI applications, and it asks employers to provide company accounts, switch off training on inputs and involve the data protection officer. For Copilot, plan on a DPIA by default.

The works council has a co-determination right under section 87(1) No. 6 of the Works Constitution Act over technical systems capable of monitoring staff, and the Federal Labour Court applied it to Office 365 in its decision of 8 March 2022. Treat the works agreement as a gating item for any pilot at a German site. One caveat matters for groups that bought their German labs: a works council normally covers its own employer, so two German companies can mean two councils and two agreements, and German counsel should confirm the structure before you set dates.

Borrow structure from the voluntary frameworks

None of the voluntary frameworks is mandatory for a CRO, but they save drafting time. NIST's AI Risk Management Framework organises the work into govern, map, measure and manage, and its generative AI profile, NIST AI 600-1, published in July 2024, lists 12 risks that make a ready-made risk register US sponsors will recognise. ISO/IEC 42001:2023 is a certifiable management system standard, and you can borrow its structure for the policy without committing to certification.

Where QueryNow fits

We help GxP teams write an AI policy and then put it to work: the data classes, the matching labels and Copilot settings, the AI inventory and the evidence a sponsor's auditor asks for. We start with a short discovery and then give a fixed price for the rest, and our AI governance for life sciences work draws on 12 years as technology partner to Arcondis, a life-science compliance consultancy. We work from Plano, Texas and Munich, and our preclinical CRO page shows the workflows we build once the policy is in place.