Guide · AI digital twin
An AI digital twin is an AI-run copy of something real, and the word after "of" tells you which kind. A machine or system twin uses sensor data to predict failures. A person twin is a language model that writes and answers as you would. A company twin models how work and cash move through a firm, with AI agents playing the people, so you can test a decision before paying for it.
Take the 5-question quizSynthetic customer panelsCompare the threeHow a company twin worksWhy the replay test
Search "AI digital twin" and the results disagree about what you asked. Engineering vendors explain twins of turbines and factories, Read AI sells a twin of you that answers your email, and research labs publish studies on twins of whole survey panels. Each is right about its own kind. In one line each, with the colour that follows each kind down this page:
The original meaning. NASA tested physical replicas of spacecraft on the ground; Michael Grieves proposed the virtual counterpart in 2002 and NASA's John Vickers named it "digital twin" in 2010 (IBM). AI now runs inside these twins: Ansys TwinAI, for example, combines physics models with field data and uses machine learning to model the gap between the two (Ansys). At factory scale, BMW twins more than 30 plants (numbers below).
The newest meaning. Nielsen Norman Group defines it as a generative model, typically built on a large language model, that acts as a proxy for a specific person (NN/g). Read AI launched Ada on 25 February 2026: you copy it on an email and it schedules meetings or answers questions from what you have shared with Read AI (Read AI).
Gartner's name for it is the digital twin of an organization (DTO), and it published a Magic Quadrant for DTO platforms on 27 July 2026 (Gartner; also cited with that date by GBTEC). The AI version adds agents that play clients, partners and staff while a calculation engine keeps the numbers honest. Our DTO guide covers the method in full.
| Machine or system | Person | Company | |
|---|---|---|---|
| What it copies | A physical asset or network | One person's knowledge, tone and habits | How demand turns into delivered work and cash |
| Data in | Sensors, telemetry, engineering models | Emails, meetings, documents, talks | CRM, time or mission tracking, invoices, interviews |
| What the AI does | Predicts failures; fills the gap between the physics model and the readings | Writes and answers in the person's place | Plays human decisions (sign, leave, delegate); a flow engine computes the numbers |
| Decision out | Maintain, tune, redesign | A reply, an answer, a booked meeting | Go or no-go on a hire, a price or a new service, with a range |
| Who buys it | Engineers, plant and asset managers | Individuals, creators, sales teams | Founders, managing partners, operations leaders |
| Public entry price | Platform licence plus an integration project, mostly on quote (see vendors) | Free (Read AI's Ada) to $299 a month (Delphi Scaler) | Enterprise DTO platforms on quote; our firm twin builds from $55,555 |
| How you check it | Compare its predictions with sensor readings | Read what it sends; set approval rules | Make it replay the last 12 months before you trust it |
| Main risk | The model drifts away from the real asset | It speaks for you badly, or without consent | False precision from agents that guess numbers |
Same eight criteria, same order, in each card.
The quick test: if the thing you want to copy has sensors, you need kind 1. If it has a voice, kind 2. If it has a payroll, kind 3.
The quick test above names the kind. This quiz goes one step further: it weighs the decision you face against the data you hold, your budget, your deadline and the people who will use the twin, then gives you a match, its entry price and a first step you can take this month. It also covers a fourth option the quick test misses: a synthetic panel of your customers. Nothing you click leaves your browser.
The four outcomes, with entry price and first step
A digital twin does not need AI. IBM sorts classic twins into four types, from the smallest to the broadest: component twins (one valve), asset twins (a whole machine), system or unit twins (machines working together) and process twins (a whole production facility or supply chain) (IBM). Most of them ran on physics equations and business rules long before language models existed.
AI adds three abilities, and each one opened a new kind of twin:
The third ability is the most powerful and the least reliable, which is why the next two sections matter.
Accurate enough to act for you when you check its work. Not accurate enough to predict what a specific person will decide in a new situation. The research says both.
The best result so far comes from Stanford and Google DeepMind: agents built from two-hour interviews reproduced their person's survey answers at 83% of the consistency people showed with themselves two weeks later (Park et al., 2024).
The sobering one comes from Columbia Business School, which built twins from long questionnaires (Twin-2K-500) and tested them in pre-registered studies. They barely beat a plain language model with no personal data, and their answers correlated weakly with the real people's (average r = 0.20), with stereotyping and hyper-rationality among the distortions (Columbia, 2025). NN/g adds consent and misrepresentation risks (NN/g).
So the products that work treat the twin as a delegate you supervise. Read AI's Ada checks with you before it answers anything other than scheduling, and it is free for Read AI users (Read AI). Delphi lets experts publish a "digital mind" their audience can chat with or call, from a free plan up to paid tiers (Delphi).
The same research became a product category: panels of person twins that stand in for a customer base. Simile, whose CEO Joon Sung Park led the Stanford study above, sells simulated users for marketing and product research; it closed its second funding round of 2026 in July (TechCrunch) and publishes no price list (Simile). Synthetic Users publishes annual plans from $12,500 a year (Synthetic Users).
The Columbia result applies here too: a panel is a set of person twins. Before you trust one with a new question, rerun a survey you already fielded and compare. Use the panel where it matched, and real interviews for the rest.
It has two layers, and they never swap jobs.
The flow engine does the arithmetic. It moves units of work (a mandate, an order, a file) through the steps of the firm, each with its capacity, queue and delay, and it counts hours, invoices and cash. It applies laws that hold in any firm. Little's law says work in progress equals throughput times lead time: accept 20% more work with the same capacity and every file simply waits longer. The engine runs the year hundreds of times with different random draws, so it returns a range (a pessimistic, a median and an optimistic case) instead of one number that looks precise and is not.
AI agents stand in for human decisions. Wherever a person's choice changes the flow, an agent plays that person: the client who signs or walks away at a new price, the partner who reviews every file or delegates, the senior who takes on a sixth mandate. Classroom business simulation games script these choices in advance; a company twin calibrates them on the firm's own history. Given what the Columbia study found, these agents work under four rules:
To keep it fast and cheap, agents deliberate on a sample of a few hundred decisions, and their answers become response curves (the chance of signing at each price, for example) that the engine reads thousands of times. You get the judgement of a language model at the speed of a statistical model.
| Job inside the twin | Done by | Why this one |
|---|---|---|
| Count files, hours, invoices and cash | Flow engine | Arithmetic must be exact and repeatable |
| Decide whether a client signs at a higher fee | Client agent, calibrated on past proposals | A formula misses context, and no real client can be asked 10,000 times |
| Decide whether the founder delegates review | Partner agent, calibrated on interviews and history | Habits matter more than org charts |
| Choose what to test, and what to do after | You | The twin advises; it does not run the firm |
In owner-led firms of 5 to 50 people, the first suspect for a bottleneck is usually the founder's own review time. You can watch that queue form in our free Twin Lite simulator, and see a full decision played out in the Maison Delaunay case (a fictional firm). Once a decision is taken and the work itself needs agents, that part belongs to AI Jungle, the hub this site is part of.
Because a twin that cannot reproduce last year has no business forecasting next year. The replay test starts the twin where the firm stood twelve months ago, feeds it what really came in (the requests, the hires, the departures) and compares its output with what really happened.
Our tolerance: within ±10% of real revenue and throughput, and within ±15% of real lead times. A made-up example to show the arithmetic: a firm that billed $2.4 million last year needs a replay between $2.16 million and $2.64 million; if its average file took 60 days, the replay must land between 51 and 69 days.
The test does three jobs:
Ask any vendor of a company twin, ours included, for the replay result before you look at a single scenario.
At industrial scale it looks like this. BMW Group links building, equipment, logistics and vehicle data into digital twins of more than 30 production sites, simulated in real time on NVIDIA Omniverse. A collision check before a new model enters a plant used to need almost four weeks of real testing; in the twin it takes three days. BMW projects that the Virtual Factory will cut production planning costs by up to 30%, a forecast rather than a measured result, as it brings more than 40 new or updated models into production by 2027 (BMW Group, June 2025).
Few firms start at that scale. Start with one asset and one failure mode you can measure. Make sure the sensors exist and their data reaches you, then pick a platform from our ranking of digital twin software by job. Our 2026 guide to digital twin vendors groups them by what they twin.
Choose by job. For email and scheduling, Read AI's Ada works by being copied on threads. To publish your expertise for an audience, Delphi's free plan is enough to test. Feed it your own material only, and write down what it may send without asking you.
Six phases: frame the decision, connect read-only to your CRM, time tracking and invoicing, discover the real flow, model it, pass the replay test, then play scenarios. For a firm of up to 50 people, the first five phases take six to ten weeks. Before any of that, ten numbers in the free simulator show you whether the questions are worth a full twin.
It is a model of something real, kept in sync with data, in which AI does part of the work: predicting failures in a machine twin, writing in a person's place in a person twin, or playing clients and partners in a company twin. A digital twin can exist without AI; the AI is what lets it learn from data and imitate people.
IBM lists component twins, asset twins, system or unit twins and process twins, from a single part to a whole production facility (IBM). Twins of people and twins of whole organizations sit outside that engineering ladder.
Twins of a person are cheap: Read AI's Ada is free for Read AI users, and Delphi runs from a free plan to $79 and $299 a month, with custom pricing for public figures (Delphi). The cost is your time feeding and supervising it. Synthetic customer panels are sold by the year: Synthetic Users from $12,500 a year (Synthetic Users), Simile on quote.
Decide what you are copying first. For a machine, instrument one asset and choose a platform. For yourself, sign up to a person-twin tool and feed it your own material. For your company, frame one decision, connect your CRM, time tracking and invoicing, and insist on a replay of the last 12 months before trusting it.
It depends on the job. Read AI's Ada suits email and scheduling because it works inside your threads; Delphi suits experts who want an audience to talk to their knowledge. Whichever you choose, check what it sends in your name before you let it act alone.
For machines and buildings, platform licences are public only at the entry level and full projects are quoted case by case; our digital twin company list gathers the public prices. For a company twin of an owner-led service firm, our builds start at $55,555.
Start free: Twin Lite runs your next 52 weeks 300 times in your browser from ten numbers and shows your likely bottleneck, revenue range and cash floor. When a real decision is on the table, a 30-minute scoping call fixes the decision and the data you need.