Digital Twins for Business: Beyond the Hype

TL;DR

• Digital twins originated in manufacturing and aerospace — living models of physical systems that update in real time with sensor data
• The software industry has borrowed the term but not the substance: most “digital twins” in business are static dashboards or CRM records, not real-time adaptive models
• Self-models are digital twins for users — structured, evolving representations of what each user believes, prefers, and needs, updated from every interaction
• This matters because the gap between AI adoption and AI value (only 17% of companies report 5%+ EBIT impact from GenAI, per McKinsey 2025) is primarily a personalization gap

The term “digital twin” has been circulating in enterprise software for years now, mostly as a buzzword attached to IoT dashboards and 3D visualization tools. But the original concept — a living computational model that mirrors a real-world entity, updates continuously from real data, and enables prediction and simulation — is one of the most powerful ideas in engineering.

The problem is that digital twins have been applied to factories, turbines, and supply chains, but almost never to the entities that matter most for software companies: the people using the products.

What Digital Twins Actually Are

NASA coined the concept in the early 2000s for spacecraft maintenance. The idea was straightforward: create a computational model of the physical spacecraft that receives the same sensor data as the real thing. Run simulations on the model. Predict failures before they happen. Test interventions before deploying them.

The three properties that make a digital twin useful — as opposed to just a database record — are:

1. Continuous updating: The model receives a stream of real-world data and adjusts itself accordingly. It is not a snapshot; it is a living system.
2. Predictive capability: Because the model mirrors the real system, you can run it forward in time to predict future states. “If nothing changes, this bearing will fail in 72 hours.”
3. Simulation capability: You can test interventions on the model before applying them to the real system. “If we reduce the load by 15%, the bearing lasts another 400 hours.”

GE uses digital twins of jet engines to predict maintenance needs across their fleet. Siemens uses them for factory optimization. The concept works because physical systems follow laws of physics — the twin’s predictions are grounded in well-understood dynamics.

Why Business “Digital Twins” Are Usually Just Dashboards

When the enterprise software industry adopted the term, it kept the name but dropped the substance. Most things called “digital twins” in business contexts are one of three things:

A dashboard — A visualization of current-state data pulled from sensors, databases, or APIs. This is observation, not modeling. A dashboard tells you what happened. A digital twin tells you what will happen.

A CRM record — A structured data record about a customer: company size, industry, purchase history, support tickets. This is useful data, but it has no predictive dynamics. It does not update its own structure based on new interactions. It does not simulate how the customer will respond to a price change.

A 3D visualization — A rendered model of a physical space or system. This is the most visually impressive and the least useful. A pretty picture of a building is not a digital twin of the building. A digital twin would predict HVAC load, occupancy patterns, and energy consumption.

None of these satisfy the three properties above. They do not update continuously, they do not predict, and they cannot simulate.

The Missing Application: Digital Twins of Users

Here is the gap: the industrial world has digital twins of jet engines, power plants, and supply chains. The software world has nothing comparable for the humans using its products.

This is a strange omission. Software companies interact with their users thousands of times per day. Every click, every search, every conversation with a chatbot, every support ticket, every abandoned cart is a data point. The volume of user interaction data dwarfs what a jet engine sensor produces.

And yet, most software products treat each interaction as if it were the first. The user’s history exists in logs and databases, but it is not synthesized into a coherent model that predicts behavior, explains preferences, or simulates responses to product changes.

The consequences show up in the failure statistics. BCG’s 2025 research found that 74% of organizations struggle to scale AI value. McKinsey’s 2025 State of AI report shows that while 78% of companies use AI in at least one function, only 17% report 5% or more EBIT impact from generative AI. These numbers describe products that are technically functional but not useful in the ways that move business outcomes — because they do not understand their users with enough depth to personalize effectively.

Self-Models as Digital Twins for Users

A self-model is what a digital twin looks like when the entity being modeled is a person rather than a machine.

Instead of sensor data from temperature gauges and vibration monitors, a self-model ingests interaction data: what the user asked, what they clicked, what they ignored, how they responded to recommendations, what language they used to describe their problems.

Instead of predicting mechanical failure, a self-model predicts user needs: what information they will want next, how they prefer to receive it, what their current goals are, and how those goals have shifted over time.

And instead of simulating load changes on a turbine, a self-model simulates user responses to product changes: “If we change the onboarding flow, how will this user segment respond? If we adjust the recommendation algorithm, will this user engage more or disengage?”

The key architectural difference from traditional user analytics is that a self-model is not just a collection of data points — it is a structured representation of the user’s beliefs, preferences, and goals, with confidence scores that update based on evidence. Each piece of user understanding is tracked, explainable, and independently updateable.

1// A self-model is a digital twin of user understanding← Not a CRM record — a living model
2interface UserDigitalTwin {
3  beliefs: Belief[];      // What the user thinks and values
4  goals: Goal[];          // What they are trying to accomplish
5  preferences: Pref[];    // How they prefer to interact
6  confidence: number;     // How well we know this user
7  lastUpdated: Date;      // Continuously, not quarterly
8}
9
10// Every interaction updates the twin
11function observe(twin: UserDigitalTwin, event: InteractionEvent) {
12  // Update beliefs based on new evidence← Bayesian belief updating
13  twin.beliefs = updateBeliefs(twin.beliefs, event);
14  // Adjust confidence scores
15  twin.confidence = recalculate(twin);
16  twin.lastUpdated = new Date();
17}

Why Personas Are Not Digital Twins

Product teams often argue that they already have user representations in the form of personas. But personas violate all three digital twin properties.

Personas do not update continuously. They are created during a research sprint, printed on a poster, and taped to a wall where they decay for months until someone runs another research sprint. The user they describe on day 1 is not the same user on day 90.

Personas do not predict. They describe a generalized archetype — “Sarah the Series B CTO who cares about scalability” — but they cannot tell you what Sarah will do when she encounters your new pricing page. They are descriptive, not predictive.

Personas cannot simulate. You cannot run a persona forward to test product decisions. “How will Sarah respond if we add a mandatory onboarding quiz?” You can guess, but you are guessing based on a static document, not running a simulation on a calibrated model.

The gap between personas and digital twins is the gap between a photograph and a weather model. Both represent something real. Only one can tell you what happens next.

The Compounding Advantage

The reason digital twins matter for AI products specifically — not just software in general — is that AI products have a unique property: they can use the twin directly.

When an AI product has a self-model of the user, every response can be personalized without explicit prompting. The AI knows the user’s context, their goals, their communication preferences, and their history — not because the user told it these things in the current session, but because the digital twin has been accumulating this understanding across every interaction.

This creates a compounding advantage that static products cannot match:

Each interaction makes the twin more accurate, which makes the AI more useful, which increases engagement, which produces more interaction data, which makes the twin more accurate. This is the flywheel that separates AI products that retain users from AI products that lose them.

The Mystica case study illustrates this directly. Mystica, a spiritual wellness app (formerly “Relationship Psychics”), saw a 60% revenue increase within 90 days of integrating self-models. The AI went from treating every conversation as the first to maintaining a continuous understanding of each user’s journey. Users stayed because the AI remembered who they were.

From Hype to Architecture

Digital twins for users are not a future concept. The architectural pattern exists today. The question for product and engineering teams is whether they will build this capability or continue to ship AI products that forget their users between sessions.

The 80%+ AI project failure rate (RAND Corporation, 2024) is not just a technology problem. It is a user understanding problem. Products that cannot model their users cannot personalize for them. Products that cannot personalize will join the 42% of organizations (S&P Global, 2025) that abandon their AI initiatives.

The industrial world figured this out decades ago: you cannot optimize a system you do not model. The same principle applies to users. You cannot serve a user you do not understand. And you cannot understand a user from a quarterly survey and a CRM record.

At Clarity, we build self-models — digital twins for users that give AI products persistent, evolving understanding of every person they serve. If your AI product still treats every interaction as the first, we should talk.