Consider a relatively simple customer:
They purchase four times a year, spend £60 per transaction, generate £90 contribution, have a 15% probability of lapsing and show relatively high price sensitivity.
The business is considering giving them a £10 benefit.
What we want to know is not: “What would this customer say about the offer?”
Nor even: “What would an AI agent pretending to be this customer do?”
We want to estimate:
What is the probability distribution of this customer’s behaviour with the intervention compared with what would probably have happened without it?
Perhaps frequency rises from 4.0 to an expected 4.3. Perhaps average order value falls slightly because the customer brings forward a purchase. Perhaps churn probability falls from 15% to 12%. Perhaps £6 of the £10 benefit simply subsidises behaviour that would have happened anyway. And perhaps the expected incremental contribution is only £3.
That is a Digital Customer Twin problem.
It is fundamentally a problem of counterfactual estimation.
A Digital Customer Twin does not need beliefs, intentions, language or agency.
It does not need generative AI.
At its simplest, it can be represented mathematically:
Customer state + intervention > probability distribution of future customer states
The customer state might contain frequency, spend, tenure, category behaviour, channel behaviour, engagement, price sensitivity, service history and churn risk.
The intervention might be a discount, loyalty benefit, price change, service improvement, communication, subscription, new product or change in returns policy.
The model estimates what changes.
That can be achieved using causal inference, experimental evidence, response coefficients, propensity models, nearest-neighbour comparisons, survival models, Bayesian updating, uplift modelling, Monte Carlo simulation or combinations of these techniques.
None requires an autonomous agent.
The important question is not whether the customer representation can act.
It is whether the model can estimate response.
This distinction becomes clearer if we separate two questions.
A conventional predictive model might ask: What is this customer likely to do next?
A customer with a 70% probability of purchasing next month might also have a 70% probability of purchasing if we give them a £20 voucher.
A conventional predictive model may correctly identify them as highly likely to purchase.
But that tells us nothing about whether the voucher caused the purchase.
If their probability of purchasing is 70% with or without the voucher, the expected incremental effect of the intervention is zero. We have simply spent £20 rewarding behaviour that was likely to happen anyway.
The Digital Customer Twin is trying to answer the different question:
How does the customer’s expected behaviour change because we intervene?
Agents can still be extremely useful.
But they sit somewhere else in the architecture.
An agent could search historical campaigns and experiments for evidence of how customers actually responded to previous interventions. It could search academic research for relevant behavioural effects, identify analogous interventions, surface candidate response coefficients and assess the strength of the evidence behind them. Another agent could interrogate customer data to identify comparable populations. Another could configure and run simulations. Another could test alternative strategies against commercial and operational constraints. As new experiments are completed, agents could help incorporate that evidence back into the model and recalibrate its response estimates.
An agent could therefore ask:
“What evidence do we have for the likely frequency response to this intervention?”
The Digital Customer Twin uses that evidence to estimate:
“Here is the distribution of likely outcomes if we intervene.”
And an agent can then explore:
“What happens if we change the audience, benefit, threshold or timing?”
The twin runs another counterfactual.
Agents can therefore be enormously valuable in building, interrogating, calibrating and operating a Digital Customer Twin.
But that is very different from making the agent the customer.
The distinction is particularly important because generative AI has made synthetic research remarkably easy.
We can create thousands of personas and ask them:
Would you buy this? Which proposition do you prefer? Would this make you more loyal? How would you react to a 10% price increase? This can be useful for generating hypotheses, exploring language or identifying possible reactions.
But there is an obvious problem.
Customers themselves are often poor predictors of their future behaviour. Creating an AI that convincingly imitates what a customer might say therefore does not necessarily tell us what the customer will do.
A Digital Customer Twin should ultimately be calibrated against observed behaviour.
If a simulation predicts that a strategy increases frequency by 8%, and repeated controlled interventions show an increase of 2%, the twin should learn.
The objective is not psychological realism.
It is predictive and causal calibration.
There is another misconception worth addressing.
If a Digital Customer Twin is sufficiently sophisticated, surely it should be able to predict completely novel customer strategies? No model can reliably infer behavioural effects for which there is no relevant evidence.
The useful question is therefore not whether an identical strategy has been tried before. It is whether its components have behavioural analogues.
A new loyalty proposition may never have existed before.
But discounts have. Status has. Recognition has. Subscriptions have. Free delivery has. Threshold rewards have. Cashback has. Credit-card partnerships have. Loss aversion has. Progress mechanics have.
An agent can help find and evaluate that evidence. The twin can use it to construct an initial estimate from the closest available behavioural analogues and attach uncertainty to that estimate. The further the proposed intervention moves from observed evidence, the wider that uncertainty should become.
And once the strategy is tested, the new evidence updates the model.
That is considerably more defensible than asking an LLM-powered agent to imagine what a customer would do and treating its answer as behavioural evidence.
This exposes the deeper problem with the idea of an “agentic customer twin”. We already face uncertainty about how a real customer will respond to an intervention. If we substitute an artificial customer, we add another chain of assumptions.
We assume the LLM can infer their unobserved preferences and motivations. We assume the resulting synthetic persona accurately represents the individual. We assume its reasoning resembles theirs. We assume what the agent says it would do corresponds to what the real customer will actually do. And then we use that synthetic response to predict the economics of the business decision.
We are piling assumption upon assumption.
That is almost the polar opposite of what a Digital Customer Twin should be trying to achieve.
Apply Occam’s razor. If we have evidence showing how comparable customers responded to comparable interventions, why insert a synthetic personality between the evidence and the prediction?
The shorter chain is:
Observed customer state > intervention > empirically grounded response > predicted outcome
The individual customer’s data tells us where they are now. Evidence tells us how customers in that state tend to respond. Probability represents what we don’t know.
We don’t need to simulate the person.
We need to simulate the effect of the decision on their behaviour.
The exciting development in Digital Customer Twins is therefore not that AI can create millions of autonomous simulated consumers.
It is that businesses can begin to build a continuously learning model connecting:
customers > interventions > behavioural responses > economics
Every campaign, experiment, proposition change and customer interaction can provide additional evidence about those relationships. Over time, the organisation develops something much more valuable than a collection of personas.
It develops institutional knowledge about what happens when it does things.
Agents could make this dramatically more powerful.
They can find evidence, identify analogues, propose coefficients, interrogate uncertainty, configure simulations and help the model learn from new interventions.
But they do not replace the evidence.
And they do not become the customer.
The defining characteristic of the Digital Customer Twin remains the counterfactual: What is likely to happen to this customer, or customers like them, if we intervene – compared with what would otherwise have happened?
That is not an agent problem.
It is a decision-modelling problem.
Agents can help build the twin. They can interrogate the twin. They can operate the twin.
They are not the twin.
We, and our partners (such as Google), use optional tools including tracking tags and cookies to tailor your online experience, improve our products and enhance services for the organisations we work with.
To find out more visit our Cookie Policy page. Click ‘Accept All’ to enable them, ‘Reject All’ to decline or ‘Cookie Settings’ to select your preferences.