What has been approved when every customer sees a different message?

With AI personalisation, a financial services campaign can begin with one proposition and end with many different communications. One customer sees convenience emphasised, another sees potential returns and a third receives reassurance about flexibility. Each version may draw on the same product facts. The impression created by their arrangement can still differ.

That is where AI personalisation becomes an approval question. Reviewing a representative message may establish that the campaign can produce something acceptable. It does not establish that every permitted variation will preserve the balance on which that judgement depended.

Zango AI’s report launch on 7 October 2026 provides a useful sector perspective. The company describes The New Era of Marketing in Financial Services, a collection of eleven essays from financial services and related leaders in Portugal examining AI, communications and customer trust. It is a collection of professional arguments, rather than an empirical study of campaign outcomes or a new regulatory standard. The question it brings into focus is practical: what should approval mean when the communication is generated rather than selected from a fixed library?

A correct fact can receive a different emphasis

In financial marketing, the problem cannot be reduced to checking whether a number or product statement is true. A benefit can be accurate and still dominate a qualification that materially changes its meaning. A communication can retain all the required words while making the less appealing information harder to notice.

Imagine an illustrative savings campaign in which an AI system adapts the message to a customer’s stated goal. It might lead with a headline rate for someone seeking growth and with access flexibility for someone building a reserve. If the rate is conditional or access has restrictions, the system needs to preserve those relationships as it changes emphasis. The product facts have not changed, but the reasons a customer might act on them have.

For firms within the UK Consumer Duty’s scope, the FCA’s consumer understanding rules require communications to support informed decisions and be clear, fair and not misleading. They also address tailoring to customer characteristics and, where appropriate, testing and monitoring. Those requirements supply the UK context for this argument; they should not be presented as conclusions derived from a Portuguese essay collection.

Personalisation can serve those aims when it makes relevant information easier to understand. The risk arises when the same adaptation process optimises persuasion without preserving the information needed to judge the offer. A system should not assume that a higher response rate establishes better understanding.

Approval can govern the range without approving every output

If every version requires the same manual review as a new campaign, much of personalisation’s operational benefit disappears. The opposite approach, treating approval of the generation system as permission for any resulting message, leaves a wide gap between intention and behaviour.

A more useful design is to specify the range of communications the system may produce. That range could include verified claims, product facts and the contexts in which they may be used. It would also define what the system may change, which relationships it must preserve and what circumstances require escalation. This is an operating proposal that needs appropriate validation; it is not an automatic safe harbour under financial-promotions rules.

The distinction between a reusable claim and a reusable permission matters. A claim approved in a detailed product explanation may need a different presentation in a short notification. A customer characteristic may justify a clearer explanation without justifying the omission of a risk. An approved range must therefore describe relevant relationships, rather than providing a bag of authorised sentences.

There is a related lesson in the missing layer in agentic marketing: production capability has advanced faster than the decisions surrounding it. A workflow needs to supply the system with the organisation’s accepted reasoning, not just ask it to produce compliant copy at the end.

The customer context also has boundaries

Personalisation depends on information about the recipient. More information can make a message more specific while raising separate questions about whether it should be used for that purpose. A marketing approval is not a blanket permission to draw on everything an organisation knows about the customer.

An illustrative system might need a customer’s selected language or preferred explanation format to improve understanding. It would require a different assessment before turning sensitive support correspondence into persuasive audience signals. The personalisation brief should state which context is available, what it is intended to do and where the system must stop.

The review also needs to consider the interaction between content and placement. Approving an advert in a private conversational context explores a related challenge: the organisation needs evidence that its controls worked even when a reviewer cannot inspect every individual encounter. A personalised campaign similarly requires a record of the conditions used to generate the communication, with access and retention proportionate to the information involved.

That record need not become an invitation to retain every personal detail indefinitely. Its purpose is to support accountability for the communication and its governing conditions. The design should identify the evidence required for that purpose, rather than equating auditability with unlimited storage.

Assurance has to observe the range in use

Testing a generation system should examine variations that might change the customer’s interpretation: an unusually short format, a different channel, a product limitation or a combination of customer characteristics. A set of typical examples can help establish expected performance, but boundary cases reveal whether the controls preserve the conditions of approval.

After launch, teams need a way to inspect what the system actually produces and connect problems to the instructions or sources behind them. If one class of outputs repeatedly weakens a qualification, the response should address that pattern rather than correcting isolated messages. If the product changes, the organisation should know which generation rules and live communications depend on the old facts.

The Consumer Duty monitoring argument adds a further test: whether communications are helping customers reach informed decisions. Output inspection and customer evidence answer different questions. A technically accurate sample cannot, on its own, show that customers understand the proposition.

The Zango launch is useful because it makes scale visible. The harder challenge is to preserve judgement as scale grows. A personalised communication should be able to change its expression without losing the conditions that make the proposition fair. Approval becomes meaningful when the organisation can explain that continuity, even though the exact words vary from customer to customer.