AI can help organisations develop and adapt digital experiences more quickly. It can generate options, synthesise research and tailor journeys to a person's context. None of those activities proves that customers are better served or that the organisation has made a better product investment. Speed matters when it helps people reach the right outcome with less avoidable effort while preserving comprehension, control, privacy and a credible route to correction. That outcome is the test, not the volume of generated work or the apparent smoothness of personalisation.

Speed is an input, not the experience

Generative tools can shorten parts of design work. Organisations can explore interface alternatives, synthesise customer research, draft content and build prototypes more quickly. That can widen the options considered and lower the cost of testing an idea. It can also accelerate a large amount of plausible work before the underlying customer need, behaviour or consequence is understood.

The commercial distinction is between producing design work faster and improving the experience people receive. A polished prototype does not establish that the product will be understandable, useful or inclusive. If output moves faster than customer evidence, product judgement and technical validation, the organisation accumulates rework and implementation risk rather than advantage.

Research with creative and UX practitioners has documented both the benefits of generative AI and the difficulty of judging where it belongs. Organisations need enough information about model capabilities and limits to make responsible product choices. The implication is not to reject acceleration. It is to ensure that faster production does not bypass customer research, model scrutiny or technical validation.

Define what better means for the customer

Efficiency is often described as fewer clicks, shorter handling time or less manual work. Those measures can be useful, but they are incomplete. A recommendation that saves two minutes and causes a costly correction has not improved the task. A personalised route that removes options may feel effortless until the user needs an alternative the system has hidden.

Start with the outcome the person is trying to reach and the conditions that make it acceptable. Include effort, but also completion quality, understanding, confidence, accessibility, recoverability and consequence. The relative importance of each condition depends on the work. A low-risk creative suggestion can tolerate different uncertainty from an eligibility decision or a change to an authoritative record.

This reframes the product decision. AI is not successful because it appears in the workflow or reduces visible interaction. It is successful when the complete experience helps the intended users achieve the defined outcome more dependably than a credible alternative.

Personalisation changes the customer contract

AI can adapt content, sequence, assistance and recommendations from declared preferences, current context or observed behaviour. That can remove irrelevant work and help a product respond to needs that a fixed journey cannot anticipate. It also changes the experience over time and can make two users see different choices without understanding why.

Microsoft's human-AI interaction guidelines recommend learning from behaviour, adapting cautiously, supporting granular feedback, explaining the consequence of user actions and providing global controls. These principles show that personalisation is not only a ranking technique. It changes the relationship between the organisation and the person using the product: what may be inferred, how the experience may change and how that person can inspect or alter the result.

When personal data is involved, this is also a governance concern. The Information Commissioner's Office states that organisations should consider fairness, purpose, transparency and the overall impact of AI processing on people. Its guidance is under review following changes in UK law, so the exact compliance position needs specialist verification. The durable design implication is that personalisation should be purposeful, legible and proportionate rather than treated as an invisible default.

Personalisation must preserve customer control

An adaptive interface can reduce effort by bringing forward relevant choices or generating controls for the task at hand. A 2025 study of dynamic prompt controls for comprehension tasks found that participants preferred the adaptive approach in the study and reported greater control, while still finding it difficult to reason about how generated controls affected the result. The study involved a small, defined task and does not prove that dynamic interfaces are universally better. It demonstrates the central tension: adaptation can help while making system behaviour harder to predict.

Important states and commitments therefore need to remain stable. The experience should show what has changed, which information shaped the change and what will happen if the user corrects it. People should be able to recover a known route, express a preference directly and turn off adaptation where that is meaningful. They should not have to make repeated corrections merely to teach the product how not to obstruct them.

Good personalisation reduces irrelevant effort without making the user reconstruct the interface or negotiate with an unexplained profile. The system may adapt the route, but it should not silently rewrite the product's promise, authority boundary or available recourse.

Test the real behaviour, not an ideal response

A scripted demonstration can make an AI experience appear unusually coherent. Real use introduces incomplete context, conflicting preferences, unexpected language, latency, model variation and change over time. The design therefore needs to be tested with the behaviour it will actually contain, including errors and adaptation that the user did not request.

Canvil research used a design tool as a probe with 12 designers in an initial study and 17 participants across six groups in a later study. The researchers found that design requirements and model behaviour need to shape each other. The implication for product approval is clear: the interface cannot be assessed separately from the model, context, controls and evaluation that determine how it behaves in use.

Ask for evidence from representative people, tasks and information. Vary the responses, expose likely failure modes and observe what people understand, accept, correct or abandon. Compare the AI route with the current experience and with a simpler non-AI improvement. This shows whether AI is creating a better outcome rather than benefiting from novelty or an undemanding comparison.

Measure customer outcomes and hidden friction

User satisfaction matters, but a single rating cannot explain why an experience worked. People may prefer a fluent response that is incomplete, dislike a necessary verification step or report enthusiasm before they have lived with adaptation over time. Behavioural evidence can also mislead when the product makes one route easier to select than the alternatives.

A 2025 personalised productivity-agent study combined a survey of 363 participants with a comparison involving 40 participants. The authors emphasised user-centred design, adaptability and the balance between personalisation and privacy. Its context is bounded and provider-affiliated, but it reinforces the need to evaluate personalisation as a relationship among usefulness, preference and information use rather than as a feature in isolation.

A credible business case should measure task completion, time and repeated use alongside correction, reversal, abandonment, accessibility, unwanted adaptation and the consequence of accepted errors. Results should be examined across different people and contexts rather than hidden inside one average. The evidence can then support a decision about the experience, model, context, controls or whether AI should remain in that part of the task.

Use an outcome test before scaling AI

Before approving a wider rollout, apply six questions to one important customer or employee journey. The questions are practical guidance, not a universal scoring method. Their purpose is to keep delivery speed, user benefit and operating responsibility inside the same investment decision.

A credible case should explain the intended outcome, the source of adaptation, the evidence of improvement and the route for correction in plain language. If it can only show that more options were generated or fewer steps were displayed, it demonstrates activity or interface compression, not a better experience.

  • Outcome: What is the person trying to achieve, and what makes the result acceptable?
  • Alternative: Is AI better than the current journey and a simpler non-AI improvement?
  • Adaptation: What changes for the person, which information shapes it and why is that change appropriate?
  • Control: How can the person understand, correct, decline or reverse the adaptation?
  • Evidence: Which behavioural, qualitative and operational signals show improvement rather than novelty?
  • Learning: Who reviews unequal outcomes, failures and unwanted change, and what can they alter?

Sources and further reading

  1. Microsoft ResearchDesignerly Understanding: Information Needs for Model Transparency to Support Design Ideation for AI-Powered User Experience
  2. Microsoft ResearchCanvil: Designerly Adaptation for LLM-Powered User Experiences
  3. Microsoft ResearchEvolving Roles and Workflows of Creative Practitioners in the Age of Generative AI
  4. Microsoft ResearchGuidelines for Human-AI Interaction
  5. Microsoft ResearchDynamic Prompt Middleware: Contextual Prompt Refinement Controls for Comprehension Tasks
  6. Microsoft ResearchFrom User Surveys to Telemetry-Driven Agents: Exploring the Potential of Personalized Productivity Solutions
  7. Information Commissioner's OfficeHow do we ensure fairness in AI?