The Digital Twin Is for Users, Not Factories

TL;DR

• Digital twins have generated massive value in manufacturing, aviation, and infrastructure. The same concept, a persistent and evolving model of a complex system, applies to individual users.
• A user digital twin (self-model) is a structured representation of an individual’s beliefs, goals, preferences, and trajectory that evolves with every interaction.
• This shifts personalization from reactive pattern matching to predictive understanding.

The digital twin concept originated at NASA during the Apollo program [1], where engineers built simulation models of spacecraft to diagnose problems and test solutions in real time. Since then, the approach has expanded across manufacturing, energy, and logistics. The global digital twin market is projected to reach $150 billion by 2030 [2], growing at nearly 48% annually. Yet almost none of that investment models the people who use software products. A user digital twin is a persistent, evolving representation of an individual’s beliefs, goals, preferences, and trajectory that updates with every interaction. This post covers the four properties that distinguish digital twins from user profiles, how predictive personalization replaces reactive pattern matching, and the compounding value that makes user twins an irreplicable moat.

What Makes a Digital Twin

A digital twin is not a database record. It is not a user profile. It is not a behavioral segment. According to an IEEE survey on digital twin definitions and characteristics [3], a digital twin is a virtual replica of a physical system that maintains a bidirectional data connection with its real-world counterpart. Four properties distinguish it from simpler representations:

Persistent state. The twin maintains a current representation of the system it models. For a jet engine, that includes temperature, vibration, fuel efficiency, and wear patterns. For a user, it includes beliefs, preferences, goals, constraints, and communication style.

Continuous update. The twin evolves as new data arrives. It does not just store snapshots. It integrates new observations into a coherent, evolving model. When sensor data shows increased vibration, the engine twin updates its wear model. When a user expresses a new goal, the user twin updates its goal graph.

Predictive capability. Because the twin models the system’s dynamics, it can predict future states. The engine twin can forecast when a component will fail. The user twin can forecast when a user will need help, when they are likely to upgrade, or when they are at risk of churning.

Decision support. The twin informs decisions about the system. The engine twin recommends specific maintenance actions for specific engines. The user twin recommends specific interactions for specific users.

Why Product Teams Do Not Build User Twins

The concept of a user digital twin is not technically complex. The reason it is rare is not engineering difficulty. It is conceptual inertia.

Product teams have been trained to think about users through the lens of behavioral analytics: events, funnels, segments, cohorts. These tools answer “what are users doing?” They do not answer “what does this specific user believe, want, and need?”

The mental model matters. When you think about users as event streams, you build event-processing infrastructure. When you think about users as complex systems that can be modeled, you build modeling infrastructure. The technology follows the mental model.

Industrial companies made this conceptual leap decades ago. They stopped treating individual machines as instances of a type and started modeling each one as a unique system with unique characteristics. The U.S. Department of Energy found [4] that predictive maintenance programs, powered by digital twin models, reduce maintenance costs by 25 to 30 percent and cut downtime by 35 to 45 percent compared to reactive approaches.

Product companies are still treating users as instances of segments. The conceptual leap to modeling each user as a unique system is the same leap, and the gains are analogous.

Anatomy of a User Digital Twin

A user digital twin, what we call a self-model, has four layers. The Nielsen Norman Group describes digital twins of people as generative models that attempt to act as proxies for particular individuals [5], trained on personal data like demographics, behavioral history, and past interactions. The architecture below extends that concept with structured belief and goal layers:

Layer 1: Belief Graph. The foundational layer captures what the user believes about the world, about your product, about their own needs, and about the alternatives. Each belief has a confidence score and an evidence chain. Beliefs are connected. Changing one can affect others.

Layer 2: Goal Hierarchy. Active goals with priority ordering, progress tracking, and dependency relationships. A user’s goals provide the context for interpreting their behavior. Someone exploring your API documentation because they want to evaluate your product has different needs than someone exploring it because they are implementing an integration.

Layer 3: Preference Profile. Communication style, learning preferences, interaction patterns, and environmental constraints. This layer shapes how the product interacts with the user, independent of what it provides.

Layer 4: Trajectory. How the user’s beliefs, goals, and preferences are evolving over time. This is the predictive layer, the trajectory of change that enables anticipation rather than reaction.

1// Layer 1: Belief Graph← What the user believes
2const beliefs = model.getBeliefs();
3// 'AI must prove ROI < 90 days' (confidence: 0.84)
4// 'Current tooling has data quality issues' (confidence: 0.79)
5
6// Layer 2: Goal Hierarchy← What the user wants
7const goals = model.getActiveGoals();
8// Primary: 'Evaluate personalization infrastructure'
9// Sub-goal: 'Build business case for CFO'
10
11// Layer 3: Preferences← How the user works
12const prefs = model.getBeliefs({ context: 'communication' });
13// 'Prefers data-driven arguments over narrative' (0.88)
14
15// Layer 4: Trajectory← Where the user is heading
16const trajectory = model.getTrajectory('purchase_readiness');
17// Trending: evaluation → business_case → procurement

From Reactive to Predictive

The transformative power of digital twins in industrial applications is prediction. A factory does not wait for an engine to fail. The digital twin predicts the failure and triggers preventive maintenance. NASA used exactly this approach with the James Webb Space Telescope [6], building digital twins to monitor a sunshield deployment that had 344 different ways to fail.

User digital twins enable the same shift. Instead of reacting to user behavior after it happens, the product can anticipate needs based on the user’s current state and trajectory.

Reactive personalization (current standard): User visits pricing page three times. System flags as high intent. System serves pricing-focused content.

Predictive personalization (digital twin): User’s belief graph shows they need to justify ROI to a CFO. Their goal hierarchy shows “build business case” as an active sub-goal. Their trajectory shows they are moving from evaluation to procurement readiness. System proactively provides an ROI calculator, a CFO-ready slide deck, and a comparison framework, before the user asks for any of it.

The difference is not just speed. Reactive personalization responds to symptoms (pricing page visits). Predictive personalization addresses causes (the user needs to build an internal business case). The interventions are qualitatively different.

The Business Model Implications

User digital twins do not just improve personalization. They change the economics of your product.

Network effects within the individual. Traditional network effects come from adding more users to a platform. User digital twins create a different kind of network effect: the more one user interacts, the more valuable the product becomes to that user specifically. This is a personal network effect that creates per-user lock-in independent of total user count.

Increased switching costs. Research published in Electronic Markets [7] found that personalized recommendations significantly increase perceived switching costs, mediated through customer satisfaction and habit formation. When a product deeply understands each user through a persistent model, starting over with a new product that knows nothing becomes a meaningful loss. This is a form of retention that behavioral-only products cannot achieve.

Expansion revenue. A digital twin that understands a user’s evolving goals can identify expansion opportunities organically. When the twin detects that a user’s needs have outgrown their current plan, the upgrade recommendation is not a sales pitch. It is a natural extension of the product’s understanding.

Reduced acquisition cost. Users with well-developed digital twins become advocates. The experience of being genuinely understood by a product is rare enough that users share it. Word-of-mouth acquisition from deeply satisfied users is the most capital-efficient growth channel.

The Compounding Value of User Twins

Industrial digital twins become more valuable over time as they accumulate data and refine their models. The same is true for user digital twins, but the compounding is even more powerful because the user feels the improvement.

Month one: the product knows your communication style and basic preferences. It is helpful but generic.

Month three: the product understands your goals, your constraints, and the trajectory of your needs. It starts anticipating what you need before you articulate it.

Month six: the product has a rich model of how you think about problems, what evidence you find persuasive, and how your priorities have evolved. It feels like working with a colleague who deeply understands you.

Month twelve: the digital twin has tracked your trajectory through a full cycle. Quarterly planning, team changes, strategic shifts. It can contextualize your current needs against a year of evolving understanding.

No competing product can replicate this. A competitor would need twelve months of observations to build equivalent understanding. Even if their technology is superior, they cannot accelerate the accumulation of individual understanding. Time and interaction are the irreducible inputs.

This compounding creates a moat. The longer a user stays, the more valuable the product becomes to them specifically. Unlike feature-based moats that can be replicated by competitors, understanding-based moats are unique to each user and cannot be transferred. As researchers studying consumer digital twins [8] have documented, each interaction adds new layers of consumer motivations, preferences, and emotional drivers, creating iterative learning that deepens over time.

Trade-offs and Limitations

User digital twins are powerful but not universally appropriate.

Modeling complexity. Users are genuinely more complex than machines. A jet engine has well-understood physics that constrain its behavior. Users have beliefs, emotions, and contexts that are harder to model reliably. As the Nielsen Norman Group points out [9], current AI simulations tend to fail to capture the messy, nuanced nature of real human behavior. The self-model must be designed with appropriate humility about the limits of its understanding.

Privacy considerations. Detailed models of individual users create significant privacy responsibilities. Users must have visibility into their twin, control over what is modeled, and the ability to delete the twin entirely. Transparency is not optional. It is a design requirement.

Cold start challenge. A new user digital twin starts empty. The product needs enough interactions to build a meaningful model. This creates a chicken-and-egg problem: the product needs to be valuable before the twin is useful, and the twin is not useful until the product has enough interactions.

Maintenance overhead. Unlike industrial twins that operate in relatively stable physical environments, users exist in rapidly changing contexts. Career changes, company changes, role changes, and life changes can make an entire twin stale overnight. The system needs mechanisms to detect and adapt to discontinuous change.

What to Do Next

1. Model one user manually. Pick your most engaged user and create a written digital twin: their beliefs about your product, their active goals, their communication preferences, and how these have evolved since they started. If this exercise reveals understanding you do not currently capture in any system, you have found the value gap.

2. Identify your prediction opportunities. List the top five user behaviors you would like to predict: churn risk, upgrade readiness, feature adoption, support needs, expansion potential. For each, identify the user beliefs and goals that would make prediction possible. These are the layers your user twin needs.

3. Build the first layer. Start with the belief graph, the most fundamental layer of a user digital twin. Use explicit user signals (stated preferences, feedback, support interactions) to populate initial beliefs. Layer in behavioral inference as confidence allows. See how Clarity models users as digital twins to understand the architecture.

You model your servers. You model your pipelines. It is time to model your users. Self-models are digital twins for the people who matter most. Start modeling.

References

1. originated at NASA during the Apollo program
2. global digital twin market is projected to reach $150 billion by 2030
3. an IEEE survey on digital twin definitions and characteristics
4. U.S. Department of Energy found
5. generative models that attempt to act as proxies for particular individuals
6. James Webb Space Telescope
7. Electronic Markets
8. researchers studying consumer digital twins
9. points out

Related

• The Personalization Stack Is Broken: Here’s the Missing Layer
• Why Your LLM Sounds Generic (And How Self-Models Fix It)
• Why Your AI Forgets Users After 90 Days