If AI Talks to the Client, Who Owns the Advice?

Expert commentary — 7 min read

Client-facing AI changes the governance question.

Internal AI use can often be governed through model inventories, access controls, value measurement, human oversight and operational risk. Client-facing AI adds another layer: the client's perception.

That is why this topic matters.

A message may begin as a draft. It may then be personalised. It may be reviewed by an adviser. It may be sent through a private channel. It may explain why a product or strategy could fit a client's situation.

At that point, the internal story — "the model only generated text" — is not enough.

The external question becomes: what did the client reasonably understand?

ESMA's supervisory briefing is important because it focuses on presentation. It states that a recommendation presented as suitable or based on the client's circumstances can amount to investment advice if the other tests are met. It also says that email correspondence, websites, apps, private messages and interactive software systems can be channels through which advice is provided.

This is the practical board insight.

AI client communication is not a single category.

It is a ladder.

At the bottom, AI drafts wording for human-controlled use. In the middle, AI explains products, risks or process steps. Higher up, AI personalises recommendations. At the top, AI participates in client-facing advice.

The governance requirement increases at each step.

A generic product explanation may require accuracy, clarity and disclosure. A personalised recommendation requires suitability. A client-facing advice interaction requires the institution to evidence that the communication was appropriate, fair, reviewed and reconstructable.

That is why classification matters.

The board should not ask only: "Do we use AI in client communication?"

The better question is:

Where on the communication ladder is each use case?

This is not semantic hygiene. It is the operating model.

If a message is Draft, the key control is review before release.

If a message is Explain, the key control is fair, clear and not misleading communication.

If a message is Recommend, the key control is suitability.

If a message is Advise, the key control is full advice governance and reconstruction.

The same model can produce outputs at several levels. A language model can draft a neutral explanation in one context and create a personalised recommendation in another. Therefore, model classification alone is insufficient. The communication must be classified by use, channel, personalisation, wording and client context.

This is also why a named owner matters.

Technology may provide the tool. Compliance may define the boundary. Legal may advise on classification. The adviser may review the message. But the business must own the client outcome.

A client does not experience a model, a policy or a committee.

A client experiences a message.

The institution must therefore be able to answer four questions for every material AI-touched client communication:

• Who owned it?
• What was it classified as?
• Why was it suitable, fair and not misleading?
• Can we reconstruct exactly what the client saw?

If those answers are visible, AI can support client service safely and professionally.

If they are not visible, the institution may have a communication channel without a reliable accountability chain.

The objective is not to prevent AI from assisting client-facing functions.

The objective is to ensure that AI-assisted communication remains governed communication.

AI can generate the words.

It cannot own the advice.