The customer you will never meet

Imagine a renewal that no human ever reads. A customer mentions to their assistant, in passing, that the motor policy is due next month, and asks it to sort out something like last year for less if it can find it. The assistant goes away, queries a dozen providers, checks the cover against what it understands the customer to need, completes the forms, accepts the terms, authorises a payment capped at that single purchase, and reports back in one line. The customer never sees your website, never opens your policy wording, never speaks to anyone. The transaction, from quote to bound cover, happens between your systems and a piece of software acting on someone else’s behalf. For a growing list of purchases, that is no longer a thought experiment.

This is not science fiction

The plumbing is already being laid. In September 2025, OpenAI launched Instant Checkout in ChatGPT, built on an open standard it calls the Agentic Commerce Protocol and developed with Stripe, designed to let AI agents, people and businesses work together to complete purchases. In OpenAI’s own description, the assistant acts as the user’s AI agent, securely passing information between user and merchant, just like a digital personal shopper, and payment is made with tokens that are only authorised for specific amounts and specific merchants with the user’s permission. This is the architecture of agent-mediated purchase, in production, today.

The analysts have noticed. Gartner predicts that by 2028, 90 per cent of business-to-business buying will be intermediated by AI agents, pushing more than fifteen trillion dollars of spend through what it calls agent exchanges. Tellingly, Gartner adds that in this new world verifiable operational data becomes a currency, and that digital trust frameworks and verifiability are prerequisites for participation. Hold onto that phrase. It is most of the argument of this piece in eight words.

A note of realism, because the reader is rightly sceptical of hype. Consumers are not yet rushing to hand over their wallets. Gartner’s own 2026 research found that shoppers want AI to help them shop more than they want it to buy for them, and the highest-stakes, most regulated products will be the last they delegate. Insurance is a considered purchase, not an impulse one. But last is not never, and the direction of travel is not in doubt. The firms that wait for the channel to arrive before preparing for it will prepare too late.

What actually breaks

Insurance regulation assumes a human on the other side of the transaction. Agent-mediated purchase quietly removes that assumption, and several things that worked yesterday stop working.

Start with consent. Today the consent chain runs from a person to your firm. Insert an agent and it runs from a person, to the agent, to you. Who consented, to what, and with what authority? Did the customer authorise this specific purchase, or hand the agent a standing mandate to transact within limits? Did they consent to your processing of their data, or only to the agent’s? Under the GDPR, consent must be freely given, specific, informed and unambiguous, with lawful basis governed by Article 6 and the conditions for consent set out in Article 7. A blanket instruction to an assistant to handle my insurance is neither specific nor informed, and you may never have spoken to the person who supposedly gave it.

Then the record of processing. Article 30 of the GDPR requires you to maintain a record of processing activities listing, among other things, the categories of data subjects, the recipients of the data, and the controllers involved. Where, in that template, does the agent sit? Is the assistant a recipient, a processor acting on your behalf, or a separate controller acting for the customer? The honest answer for most firms is that their ROPA has no row for it, because the category did not exist when the template was written.

The controller boundary is harder still. The provider of the agent is making decisions about personal data: which insurers to approach, what to disclose, what to accept. That looks a great deal like controllership. Your firm is also a controller. You may be joint controllers under Article 26, with a duty to agree and make transparent who is responsible for what, or you may be two independent controllers in a chain. Most firms have never mapped this, and the answer changes your transparency obligations under Articles 13 and 14, because the party you must inform is not the party you are dealing with.

Finally, automated decision-making. Article 22 of the GDPR gives a person the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects, together with the right to obtain human intervention. The irony of agent-mediated insurance is that the buyer is now automated too. When software sells to software, the safeguards built around a human reading a screen do not fire on their own, and the obligation to provide for human intervention does not disappear simply because no human was present.

The regulatory reality

None of this sits in a vacuum. The regulatory load is already here, and it is not waiting for the channel to mature.

Under the EU AI Act, Regulation (EU) 2024/1689, AI systems used for risk assessment and pricing in relation to natural persons in the case of life and health insurance are classified as high-risk under Annex III, as are systems used to evaluate the creditworthiness of natural persons. High-risk status brings the full weight of the Act, including the data and data governance requirements of Article 10, and for these Annex III systems the obligations apply from 2 August 2026. If your own pricing engine is already high-risk, then a transaction in which an automated buyer meets your automated underwriter is not the softer case, it is the harder one.

DORA, the Digital Operational Resilience Act, Regulation (EU) 2022/2554, has been in application since 17 January 2025 and covers insurers and reinsurers among its twenty categories of financial entity. It requires firms to maintain a register of their information and communications technology third-party arrangements, and to manage the concentration and dependency risk those arrangements create. An external agent platform that becomes the front door to a meaningful share of your new business is precisely the kind of dependency DORA is concerned with, and most firms have not yet asked whether it belongs in the register at all.

Distribution rules add a further twist. The Insurance Distribution Directive requires, in Article 20, that any contract proposed is consistent with the customer’s demands and needs, established on the basis of information obtained from that customer. When that information arrives through an agent that has interpreted a casual instruction, how do you discharge the duty, and how do you evidence that you did? Supervisors are alert to the shift. In August 2025, EIOPA published an opinion on AI governance and risk management, setting out how existing expectations on fairness, data governance, record-keeping and human oversight apply to AI in insurance. The expectations are not waiting for new law to catch up.

The reframe

Here is the point most easily missed. Every one of these problems is a governance problem before it is an AI problem. The consent chain, the ROPA, the controller boundary, the lawful basis, the third-party register, the demands-and-needs evidence: each depends on having a single, trusted, consistent view of your data, your processing and your obligations, available in real time. An agent transacting on a customer’s behalf touches all of them at once, in the moment, with no human to paper over the seams. You cannot expose a fragmented, internally contradictory governance estate to an autonomous counterparty and expect a defensible outcome.

This is why the cost-reduction framing of governance consolidation is a trap. Treated as housekeeping, as a way to retire a few overlapping licences and trim some headcount, the work gets done slowly, defensively, and to the minimum standard. Treated correctly, it is the infrastructure that decides whether you can transact in the next channel at all. Recall Gartner’s phrase: verifiable data becomes the currency, and trust frameworks are the prerequisite for participation. The firms that can present a clean, consistent, machine-readable account of who they are and what they are permitted to do will be the ones agents can deal with safely. The firms that cannot will be quietly passed over, not by a competitor’s better marketing, but by their own inability to answer a machine’s questions at a machine’s speed.

Governance is not the brake on your AI ambition. It is the thing that makes the ambition deliverable.

What readiness actually requires

The questions worth asking now are not about which platform to buy. They are about whether your foundation can bear the weight.

Can you produce, on demand, a single trusted view of a customer, a policy and the consents attached to them, or does the answer differ depending on which system you ask? Does your record of processing have a defensible entry for an agent-mediated transaction, and have you decided whether the agent’s provider is a processor, a controller, or a joint controller with you? Do you know which of your systems would be in scope as high-risk under the AI Act, and can you evidence the data governance behind them? Is your third-party register complete enough that a new agent dependency would actually surface in it? And when a supervisor asks how you met the demands-and-needs duty for a purchase that no human at your firm ever witnessed, do you have an answer, or a silence?

If those questions are uncomfortable, that is the finding. The discomfort is not a reason to slow down on AI. It is the map of the work that has to come first.

Where this leaves you

We have watched the agentic shift coming, and we have helped organisations build the governance foundation that makes new channels possible. We help privacy, risk and governance teams do three things: define what they actually need, select the right platform on the right criteria, and build the operating model around it. The insurers who win the agentic channel will not be the ones with the boldest AI slides. They will be the ones whose governance was ready. We help you be ready.

Sources

• OpenAI, “Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol,” 29 September 2025. openai.com/index/buy-it-in-chatgpt
• Gartner, “Gartner Unveils Top Predictions for IT Organizations and Users in 2026 and Beyond,” 21 October 2025. gartner.com
• Gartner, “Gartner Survey Finds Consumers Want AI Shopping Help, But Not AI Purchase Decisions,” 27 May 2026. businesswire.com
• EU Artificial Intelligence Act (Regulation (EU) 2024/1689), Annex III, point 5(c) (risk assessment and pricing in life and health insurance) and point 5(b) (creditworthiness); Article 10 (Data and Data Governance), applicable for Annex III systems from 2 August 2026; Article 113 (application timeline). artificialintelligenceact.eu/annex/3; /article/10; /article/113
• Digital Operational Resilience Act (DORA), Regulation (EU) 2022/2554, in application since 17 January 2025; register of information on ICT third-party arrangements. European Insurance and Occupational Pensions Authority (EIOPA), eiopa.europa.eu
• General Data Protection Regulation (Regulation (EU) 2016/679), Article 6 (lawfulness of processing), Article 7 (conditions for consent), Article 22 (automated individual decision-making), Article 26 (joint controllers), Articles 13 and 14 (information to be provided), and Article 30 (records of processing activities). gdpr-info.eu
• Insurance Distribution Directive (Directive (EU) 2016/97), Article 20 (demands-and-needs test and personalised recommendation). eur-lex.europa.eu
• EIOPA, “Opinion on Artificial Intelligence governance and risk management” (EIOPA-BoS-25-360), 6 August 2025. eiopa.europa.eu