Ten commitments, one conspicuous absence
The ICO's report sets out, in more operational detail than most regulators have offered so far, how it expects UK data protection law to apply to the way AI models are built and trained. Attached to the report are commitments from ten of the largest AI developers: Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI, and Stability AI all agreed to specific changes in how they handle personal data used in training and deployment.
X.AI is not on that list. The ICO confirmed it has opened a formal investigation into the company's Grok system rather than securing a voluntary commitment, which is a meaningfully different regulatory posture. Voluntary commitments from the other ten developers let the ICO claim progress without litigation. An investigation signals the regulator believes voluntary engagement was not a realistic path with this particular company, a distinction worth watching for what it predicts about enforcement priorities going forward.
What the commitments actually cover
The ICO organized the commitments around four areas: identifying a lawful basis for processing personal data during training, providing meaningful transparency about that processing, enabling people to exercise their data rights, and demonstrating that safeguards materially reduce risk rather than existing on paper. The emphasis on demonstrating effectiveness is the sharper requirement here. Plenty of AI governance programs can point to a safeguard's existence. Far fewer can show evidence that the safeguard changes an outcome.
Where a developer relies on legitimate interests as its lawful basis, a common choice for training data processing, the ICO now expects a documented legitimate interests assessment, not an after the fact justification produced when asked. The regulator was equally direct that special category data embedded inside large training sets, health information, and other sensitive data that ends up in scraped or licensed corpora, is hard to manage but that difficulty does not reduce the legal obligation to manage it.
The line that matters for agentic AI
ICO director of technology regulation Richard Nevinson put the agentic AI question in terms that should end a specific industry argument: the fact that an AI agent acts with autonomy is not an excuse for poor compliance. That sentence is aimed directly at the defense some AI vendors have offered when an agent takes an unexpected action, that the system operated independently and therefore responsibility is diffuse. The ICO's position is that autonomy changes the technical story, not the legal one.
The regulator's six week call for evidence, running to November 20, 2026, is gathering input specifically on how data protection risk changes as AI systems gain more autonomy. The ICO cited documented cases of training data extraction, including sensitive items like API keys and passwords surfacing from models, and of AI agents bypassing intended protections, using unauthorized channels, or reaching external systems beyond their intended scope. Those are not hypothetical risks the ICO is pre empting. They are incidents the regulator is already aware of.
Why a six week comment period is worth your attention
A call for evidence sounds like a procedural step, and in isolation it is. What makes this one consequential is where it feeds: the ICO says the submissions will inform a forthcoming statutory code of practice on AI and automated decision making. A code of practice under UK data protection law is not merely advisory once finalized, it becomes the reference standard regulators and courts use to judge whether an organization's practices were reasonable.
Companies deploying agentic AI systems that touch UK personal data, directly or through a vendor, have a narrow and genuinely useful window here. Submitting evidence or at minimum reading what other organizations submit gives visibility into which agentic AI risk scenarios the regulator is taking seriously before the code of practice locks in a specific framework. Waiting until the code publishes to engage means responding to settled language instead of helping shape it.
The enquiries already underway
Beyond the public commitments and the call for evidence, the ICO disclosed it has already made direct enquiries with OpenAI, Anthropic, Meta, and the UK AI Security Institute about recent agentic AI testing and deployment. That detail suggests the regulator is not waiting for the six week evidence window to close before examining specific deployments, it is running informal fact finding in parallel with the formal process. Regulators that move this way typically intend the formal process to validate conclusions they are already forming, rather than to start from a blank slate, which is worth factoring into how much weight to put on the eventual code of practice landing close to what these four organizations are already being asked about.
For any enterprise running agentic AI features that process UK user or customer data, that combination, public commitments from major model providers, an open investigation into one holdout, and direct enquiries into specific deployments, adds up to a regulator that is actively building a enforcement capability rather than issuing a report and moving on. Treat this as the groundwork phase of a code of practice that will eventually be used against specific companies, not as a one time publication to skim and file away.
What this means for your data governance roadmap
If your company processes UK personal data through any AI feature, agentic or otherwise, the ICO's four commitment areas, lawful basis, transparency, rights enablement, and demonstrable safeguard effectiveness, are a reasonable internal checklist to run against your own systems now, independent of whether you are one of the ten companies named. The legitimate interests assessment requirement in particular is a concrete, low cost document most companies can produce quickly and should have on file before anyone asks for it.
The agentic AI call for evidence is the part to actually act on rather than file away. If your organization operates AI agents with any meaningful autonomy over UK personal data, consider submitting evidence of your own safeguards before November 20, both to shape the resulting code of practice and to create a documented record of good faith engagement that will matter if that code is ever used to assess your practices retroactively.
