Product teams are increasingly using artificial intelligence to act like a customer before a real person sees a new feature, app, campaign or support flow. These AI synthetic users can review a prototype quickly, compare product ideas, identify possible friction points and produce interview-style feedback.
The approach is attractive to teams working with limited time and research budgets. But synthetic customers are best treated as tools for generating hypotheses, not proof that people will behave in a particular way. An AI system can explain why it might buy a product or abandon a checkout page. It does not have a household budget, physical constraints, personal history or real consequences attached to its choices.
The key question is not whether a model can imitate customer language. It is whether the simulation has been validated well enough for the decision it is being used to support.
What AI synthetic users are
Terms such as synthetic customers, simulated users and AI personas are often used interchangeably, although they can describe different systems.
At the simplest level, an AI persona is a prompt given to a large language model. A researcher might ask it to respond as a budget-conscious parent, a first-time business owner or an older adult unfamiliar with mobile banking. This can help teams brainstorm questions, identify assumptions and consider alternative viewpoints. However, the response is heavily shaped by the prompt and the model’s training data.
More structured systems combine a model with selected attributes, such as location, income range, device access, stated preferences and past interactions. Some tools draw on survey results, customer-service records or product analytics to create segments that can be queried repeatedly. Others use behavioral modeling, including rules or statistical patterns intended to simulate choices over time.
The most ambitious systems are sometimes described as digital twins of consumers. That label should be used carefully. A consumer model may represent selected patterns from available data, but it is not a complete representation of a person or of a human life.
How simulated users can support product development
AI personas can be useful at several stages of product development. Early in the process, teams can use them to compare concepts, list possible objections or surface questions an unfamiliar audience may ask. Designers may use them to review the wording of onboarding screens, help content and error messages.
In operational settings, simulated users can help test customer-support scenarios, examine whether an automated agent follows a defined policy or explore unusual sequences of requests. Marketing teams may compare messages before investing in production. Product managers may use AI product testing to reduce a long list of ideas to a smaller set that deserves research with people.
These uses are most credible when they assess internal consistency, expose obvious omissions or create scenarios quickly. They are less suitable for predicting demand, willingness to pay or emotional responses in real-world settings.
Where simulation can help
- Generating interview guides, research hypotheses and edge cases before a study begins.
- Testing many support or workflow scenarios that would be difficult to script manually.
- Checking whether a prototype handles a defined set of constraints consistently.
- Exploring rare but important situations, such as a lost device, account lockout or complex service complaint.
- Helping researchers decide which concepts should be tested with real participants.
Used this way, AI synthetic users are a filter and rehearsal space rather than the final judge of customer needs.
Why the approach is appealing
Recruiting participants takes time, incentives cost money and research teams cannot interview every possible audience before each product change. Simulated users are available on demand and can be asked to address the same scenario repeatedly. That repeatability can help when a team wants to compare versions of a flow under a consistent set of assumptions.
They may also make early-stage research more accessible inside an organization. A product manager who is not trained to run a formal study can use a carefully designed simulation to challenge an idea before it reaches engineering. This can encourage useful questions earlier in the process.
However, speed can create false confidence. A detailed and fluent answer is not necessarily a reliable answer, particularly when the system is asked to predict behavior rather than comment on a clearly defined scenario.
Plausible feedback is not evidence of human behavior
Large language models generate likely text. They do not experience a product or hold stable preferences in the way a person does. When asked about a subscription price, a model can produce a persuasive rationale based on patterns in its training data. It does not have to pay the bill.
This matters because many important product decisions depend on observed behavior rather than stated opinion. People may say they value privacy yet accept permissions when rushed. They may prefer simple pricing in principle but choose a familiar brand. They may abandon a form because of an interruption, poor connectivity, fatigue or a small point of confusion that the design team did not anticipate.
Research evaluating language models as simulated survey participants suggests that outputs can sometimes resemble broad patterns in existing data under particular conditions. But results can change with the prompt, model, population and topic. Agreement with an average response does not show that a model represents the variation within a real population.
For user research, the standard should be higher. Usability studies observe what people can do, where they hesitate and what they misunderstand. A simulated user can describe an accessibility problem, but it cannot reliably demonstrate how a person using a screen reader, a slow connection or an unfamiliar language will experience a service.
A synthetic customer can model an assumption. It is not automatically an independent witness.
Where AI personas fail to represent people
Every model has boundaries, and those boundaries matter when the subject is human behavior. Training data may give greater visibility to people who are active online, more likely to answer surveys or better served by mainstream digital products. A system can therefore produce a polished representation of a typical user while poorly representing people outside that pattern.
Cultural and linguistic differences are difficult to reduce to profile fields. A direct English-language prompt may not reflect how trust, politeness, family decision-making or financial risk are understood in another context. Age is also more than a birth year: technology familiarity, confidence, cognitive load and access to support can change how someone uses a product.
Accessibility is a further limitation. A model can be asked to role-play a blind user or a person with motor impairments, but it does not use assistive technology or experience the cumulative burden of inaccessible design. Accessibility specialists and people with relevant lived experience remain essential when testing services in real conditions.
Socioeconomic constraints can be flattened in the same way. Limited mobile data, a shared phone, no credit card, unstable connectivity or concern about an unexpected fee can alter an entire customer journey. These are not simply preferences that can be added to a prompt.
Synthetic data is not automatically private
Some systems use synthetic data: artificial records designed to preserve selected statistical properties of an original dataset without directly reproducing individual records. This can be useful for software testing, experimentation and controlled analysis, especially when it reduces unnecessary exposure of raw customer information.
Yet synthetic does not automatically mean anonymous or safe. Privacy risk depends on how the data was created, the source information used, access controls and whether outputs could be combined with other information to identify someone. A dataset may also retain sensitive patterns or expose information about small groups if safeguards are weak.
Data-protection obligations can still apply when organizations use customer information to build personas or behavioral models. Under laws such as the European Union’s General Data Protection Regulation, organizations must assess whether processing involves personal data or can be linked to identifiable people. Relabeling, aggregating or generating customer information does not by itself remove obligations around lawful processing, data minimisation, security and governance.
The risk of circular validation
A major risk is circular validation. A company may have historical data that mainly reflects its existing and relatively well-served customers. It then builds AI personas from that data, tests a new pricing model with those personas and receives positive feedback. The outcome may appear to be evidence, while actually confirming assumptions already embedded in the company’s data and strategy.
This can turn AI bias into a product roadmap. Teams may optimize for users they already understand while overlooking people who were excluded, dissatisfied or unable to use the service. A model’s coherent language can make these gaps harder to notice.
Human research is not free from bias. Its value is that teams can recruit beyond their usual customer base, observe disagreement, test competing explanations and encounter findings that do not fit the original plan.
A responsible role for AI in user research
A credible approach is layered. Synthetic users can help teams develop questions, identify edge cases and prioritize experiments. Real participants should validate whether a need exists, whether a product is understandable and accessible, whether people will use it over time and whether the service creates unexpected harms.
A practical process can include:
- Use AI personas to make assumptions explicit and generate alternative scenarios.
- Document the model, data sources, prompts, intended population and known exclusions.
- Test high-risk assumptions with recruited participants from relevant and underserved groups.
- Use usability sessions, observational research, surveys and controlled experiments where they fit the question.
- Record where human findings contradict the simulation, then revise or discard the model when necessary.
This treats AI as a research assistant rather than a replacement for contact with customers. It also makes uncertainty visible: documentation of what a system cannot represent may be more useful than a dashboard of confident synthetic responses.
Hybrid research, not a customer-free future
Synthetic customers may become a routine layer in product development, particularly where teams need to examine many scenarios quickly. Their strongest role is likely to remain early exploration, quality checks and scenario testing.
But human-centered design still requires contact with people affected by a product. Real users provide evidence about lived experience, changing norms, practical constraints and consequences that a model cannot possess.
The question is not whether AI can sound like a customer. It is whether the system has been validated for the decision at hand, whether its blind spots are understood and whether people affected by the product still have a meaningful place in the research process.
Image by zhuyongbo on Pixabay.