The central finding of the Information Commissioner’s Office (ICO) in AdTech 3.0: Online advertising in the age of LLMs and agentic AI, published on 13 November, is that agentic AI and large language models (LLMs) “don’t create a new set of data protection requirements for online advertising”. That line led our news report on the 51-page document. Read past it, though, and the report asks more of firms than the headline suggests. Under seven themes, from accountability to risk assessment, the regulator works through evidence an advertiser might point to for an AI-run campaign, and says each piece may fall short on its own: a performance dashboard, a supplier’s “aggregated” label, a purpose written as campaign optimisation, a cohort signal, an impact assessment signed off before launch. The question for UK advertisers, agencies and publishers is whether they can obtain that evidence when they increasingly depend on suppliers to tell them how those systems work.
Strategic Insight: The report says plainly that it “doesn’t create new legal requirements”. In our view, that makes its worked examples more useful, not less: they show how the regulator already reads PECR (the Privacy and Electronic Communications Regulations) and the UK GDPR against agentic systems, and its findings will feed the ICO’s coming guidance on automated decision-making (ADM) and agentic AI.
Why did the ICO write a report that changes no rules?
Organisations asked for it. Firms told the ICO that being clearer about how to meet the existing rules would help them take up these technologies and business models responsibly. The regulator commissioned outside research on the market and spoke to advertisers, publishers, agencies and technology providers.
It uses “AdTech 3.0” for a direction of travel, not a product or a standard. The defining shift, in its account, is not that ad systems use AI, since machine learning has been used in online advertising for a while. It is that people now talk to advertising systems in natural language, and organisations can hand tasks to systems that act with some autonomy. The report is also clear that these tools often sit on top of existing programmatic infrastructure rather than replacing it.
The real story is where the evidence sits
The ICO’s 2019 update report into adtech and real-time bidding raised worries over consent and lawful basis, transparency, special category data and tangled supply chains. The new report says agentic AI does not make those problems go away. What it changes, on the ICO’s account, is who holds the information needed to deal with them. When an agent turns a goal into operational decisions, talks to other systems or pulls from several data sources, separate organisations may each hold information about a different part of the processing.
The executive summary states the consequence. Firms now lean more heavily on suppliers to tell them how processing happens, and need evidence on where data comes from, what it is used for, who is responsible and what controls apply, “rather than relying on labels or contract-based assurances alone”.
| What a firm might rely on | What the ICO says | Report section |
|---|---|---|
| A dashboard of campaign performance | Watching performance alone is not meaningful oversight of personal data use | 4.1 |
| A person who signs off the system’s output | Some human involvement does not by itself stop a decision being solely automated | 4.1 |
| A supplier label such as “aggregated” or “privacy-preserving” | Labels do not explain what personal data a system uses, its origin or how decisions are made | 4.5 |
| A contract naming each party’s role | Controller or processor status follows what actually happens to the data | 4.5 |
| A stated purpose of “campaign optimisation” | Broad wording may not adequately describe the processing | 4.6 |
| A cohort-level output | Whether it is proportionate turns on the whole chain of processing | 4.6 |
| An impact assessment done before the system goes live | It may reflect only one point in the system’s life | 4.7 |
Strategic Reality: Accountability cannot be passed to the software. “An organisation can’t transfer its responsibility by having an agentic AI system handle campaign activities or optimisation,” the report says. Nor does it fall automatically on the advertiser: the ICO warns parties not to treat the advertiser as answerable for all the processing just because it set the campaign objective.
What does an AI-run campaign look like in the ICO’s examples?
The report examines five settings, among them these three.
In chatbot advertising, the LLM answers a prompt and, in the uses the ICO examined, hands signals about the request, such as its apparent topic or what the person seems to want, to a separate ad system, which picks the ad. The report treats the line between an ad placed beside a response and one built into the response as significant. In the examples it saw, advertisers could buy the slot next to an answer but could not pay to change the generated answer itself. For background, see our earlier coverage of advertising in ChatGPT.
In industry protocols, AdCP (Ad Context Protocol) lets an advertiser set out what a campaign should achieve in plain language whilst agents query publishers, weigh up opportunities, set up buying arrangements and make adjustments. The ICO cites one early AdCP campaign on connected TV (CTV), which it says was not evidence that the approach has become standard. A second model, ARTF (Agentic Real Time Framework), places an advertiser’s agent inside a seller’s technical environment. When the research was done it was still a proposal that no organisation had used commercially.
In walled gardens, the large platforms that control their own services, ad systems and user data, the advertiser sets an outcome and a budget and the platform’s system decides on audiences, placements and spend. Some describe this as “supervised autonomy”. The advertiser watches dashboards of aggregated performance data, and in some cases platforms have removed manual controls that used to be available.
Critical Context: In our reading, this is where the report’s two strands meet. In some walled gardens the advertiser’s view is a dashboard of aggregated campaign performance. Elsewhere, the same report says that “simply monitoring campaign performance won’t provide meaningful oversight” of how personal information is used. The report does not address what an advertiser should do if a platform declines to share more.
Where conversation changes profiling
Profiling is an established feature of adtech. The report argues that a conversation is “a related but distinct development”: one exchange or a chat history can expose what someone is going through right now, what worries them or what they plan to do, and let the system read meaning beyond the words used. A prompt might mention a health condition, a religious belief or a political opinion outright. Or the system may infer one without the person ever stating it.
The test the ICO applies is intent and effect, not accuracy. An inference is generally special category data if the organisation aims to infer a special category trait, or treats someone differently because of it, “regardless of how confident it is that the inference is correct”. In its example, a person asks a chatbot about symptoms, the service infers a possible health condition and lets that shape the ads shown. The processing involves special category data. The report repeats the 2019 position that using special category data for online advertising requires the person’s explicit consent.
⚠️ Warning: Not instructing an agent to target by health, religion or politics does not settle the question. In the ICO’s audience-expansion example, patterns a system finds could be read as proxies for those characteristics. What counts, the report says, is whether the advertiser uses the system to infer special category information or to handle people differently on that basis.
Consent has to travel down the chain
Where an ad service stores or reads information on a person’s device, PECR’s regulation 6 applies, and consent must meet the UK GDPR standard, chatbot or not. The ICO’s guidance on storage and access technologies, cited in the report, also says firms need effective ways to pass consent withdrawals to relevant third parties so no one keeps using the data. The report found that more automation could make consent and withdrawal signals harder to pass on and honour. It also sees the reverse possibility: agentic workflows could carry those signals in tidier, more uniform formats. AdCP’s labelling scheme for data handling might offer a structure for that, though the ICO says how well it works would turn on whether participants implement and respect it.
Who is affected?
The report’s examples fall unevenly across the ad supply chain.
| Group | What the report means for them | What is still unknown |
|---|---|---|
| Advertisers and agencies | Record their role, what the agent was authorised to do and what is known about its actions | How much detail walled-garden platforms will share |
| Chatbot and walled-garden operators | Let people understand whether their interactions shape the ads they see, and keep paid ads distinct from generated content | What form those explanations will take |
| Ad tech suppliers and protocol builders | Their customers need evidence of data sources, purposes and controls, not labels alone | Whether data labels in protocols will be accurate and respected |
For people using chatbots, the report notes that a conversation may feel like advice rather than a place where ads are sold, and the two-system design behind a chatbot ad may not be apparent. A general notice that a service “uses AI to personalise advertising”, the ICO says, does not tell people which of their information shapes the ad or which organisations take part in choosing it.
What this means for advertisers buying agentic tools
The ICO sets out what checks on a supplier might cover for AdTech 3.0: where a system gets its data, what processing it carries out, how far an agent is allowed to act and what limits apply to data passed between systems. It also points to its AI audit framework as a way for firms to work out and record who is controller and who is processor in a supply chain built on AI.
Purposes need the same precision. The report notes that an agent “optimising” a campaign might shift budget between publishers, adjust an existing audience or generate new inferences about people, and that each may draw on different data and serve a different purpose. Connecting an agent to a new source of consumer data part-way through a campaign is its example of “purpose creep”.
Implementation Note: The report’s DPIA example covers a similar case. An organisation assesses an ad system that selects audiences from a defined dataset, then adds an agentic capability with access to new consumer data. The original assessment may still cover parts of the processing, but revisiting it lets the organisation check it still reflects the risks.
The open challenges
Visibility inside closed platforms
The report places accountability with the organisations that decide why and how data is processed, and expects them to understand and supervise whatever systems they develop, purchase or run. It also describes advertisers in some walled gardens seeing aggregated dashboards whilst the platform makes operational decisions. It does not address what an advertiser should do if a platform declines to share more.
Words that mean different things
Engagement with industry turned up inconsistent terms for these systems, which the ICO says makes it harder to describe processing and divide responsibility in contracts. One “agentic” tool may apply a brief inside a single ad environment; another may deal with several suppliers and data sources. The ICO’s answer is to look at what the system does rather than what it is called. In our view, that puts the burden of finding out on the buyer.
Data that hides where it came from
The report describes “data washing”: combining or transforming personal information so its origin and conditions of use become less apparent, then presenting the result as an audience insight or aggregated signal. Its position is that reshaping personal data does not excuse a firm from looking at how it was first collected. A cohort-level output may disclose less to the advertiser, but building the cohort may still have involved processing personal data.
Assessments for systems that keep changing
New permissions, data sources, capabilities or links to other systems can change what an agent does with data and how risky it is. The report points to the UK GDPR’s article 35(11), which requires a DPIA to be reviewed where the risk changes, and frames review as running through the system’s lifecycle. In our view, for agents that learn and adapt, that makes it a continuing task, not a launch gate.
Reality Check: Much of the market the report describes is early. ARTF had not been used commercially, consumer shopping agents were not yet established in UK advertising, and fully agentic buying was not common across CTV when the ICO wrote the report. The regulator is setting out its reading before several of these models reach scale.
What to watch
Four signposts will show whether the report’s examples harden into expectations firms are held to:
- The ICO’s agentic AI and ADM guidance and its code of practice on AI and ADM. The report says its findings will inform both. If the guidance carries forward the point that performance monitoring is not oversight, or that “campaign optimisation” is too broad a purpose, advertisers will have firmer ground to ask platforms for more.
- How chatbot services explain their ads. The ICO suggests a conversational interface could let a person explore what influenced an ad. Whether operators build that, or stop at a general AI notice, will show how far the transparency expectations land.
- Consent signals in agent protocols. The report sees AdCP’s labelling taxonomy as a possible way to carry consent and restrictions between systems, if participants implement and respect it. Uptake of those signals by participating ad systems would be a practical test.
- Joint working with other regulators. The report says chatbot and agentic ads may also raise questions for regulators of ad standards, consumer law, competition and online safety, and that the ICO will share evidence where remits overlap. Joint statements on chatbot ads would show whether disclosure rules are converging.
Firms looking for a new rulebook will not find one here. The report’s value lies in its examples, which take the familiar duties of the real-time bidding era and show where they bite when the buyer is software. The organisations best placed are those that can say what their agents were allowed to do, what data they touched and why, without relying on a supplier’s dashboard or label to say it for them.
Source: Information Commissioner’s Office, AdTech 3.0: Online advertising in the age of LLMs and agentic AI, 13 November 2026. Also referred to: the ICO’s Update report into adtech and real time bidding (2019) and its technology guidance pipeline.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching whether the ICO’s forthcoming agentic AI guidance turns these examples into expectations advertisers can cite to their platforms. Read more analysis at Insights, or get in touch.