Solutions · Customer digital twins
Customer digital twins for better commercial decisions
Test offers, messages and products on a simulated version of your own customers before you spend the budget.
In build
Definition
What is a customer digital twin?
A customer digital twin is a simulated respondent that stands in for a group of your real customers. It is built from what that group did — what they bought, how often, through which channel and for how long — and it answers a proposal in words: whether it would take an offer, what it likes about it, and what it objects to. CrowdOS builds one twin for each weighted cohort of customers, not one for each person.
Introducing customer digital twins
From a score to a reaction
Customer analytics has moved in steps. First came risk scores and segments. Then propensity models and next-best-action engines that predict a single outcome from a single input. Each step made better use of the data a business already holds, and each stopped short of the question commercial teams ask most: if we changed this offer, this message or this price, how would our customers react?
Large language models make that question answerable in a new way. Combined with a business's own behavioural and transactional data, they can represent groups of customers as virtual respondents that react to a proposal in words, not just in a score. CrowdOS customer digital twins are built this way: from your customer file, grouped into weighted cohorts, each represented by a persistent archetype that can be asked, challenged, and put in a room with the others.
A customer digital twin is not a prediction engine for individuals, and it is not a replacement for talking to customers. It is a fast, inexpensive way to rank options and find objections before a campaign, so that the research and media budget goes to the ideas most likely to work.
Better commercial decisions
What can customer digital twins be used for?
01
Offer and incentive testing
Put candidate offers in front of every cohort before launch.Offers ranked by cohort, with the objections each cohort raises.02
Message and creative testing
Try framings, headlines and creative against the cohorts you need to move.Which framing moves which cohort, and which leaves it unmoved.03
Product and proposition testing
Pressure-test a new product with the customers most and least likely to adopt it.Who it is for, who it is not for, and why, in their words.04
Pitch preparation
Give sales and service teams something better than a segment description.Cohort-by-cohort talking points drawn from what the twins pushed back on.05
Campaign learning
Upload results after a campaign and compare them with what the twins said.Where the twins ranked offers correctly, and where they did not.
The design
Why cohort twins and not individual twins?
One approach builds a twin for every customer. Published research on language-model personas points the other way: they approximate the distribution of a group's answers far better than the answers of any one person. CrowdOS is built on that finding rather than against it.
Your customer file is grouped into between twenty and two hundred cohorts by behaviour. Each cohort is weighted by its share of your customers and represented by one archetype. The result is an audience you can read at a glance, that costs a fraction of simulating every customer, and that never needs a named individual's record to run.
How your data comes together
From a customer file to a twin
Demographics, tenure, product holdings, transactions, engagement and service history are reduced to cohort-level features when the file is uploaded. Direct identifiers — names, emails, phone numbers, national ID numbers, full addresses — are dropped before anything is stored.
Each cohort becomes a profile written from its distinguishing features by a fixed template. No language model writes the profile, and no personality is inferred from purchases: where you have survey data that states attitudes, it is added; where you do not, the twin starts from a population prior.
- Demographics
- Tenure
- Product holdings
- Transactions
- Engagement
- Service history
Cohort features
One twin per cohort
Cohort 7 · Lapsing card holders
IllustrativeNo annual fee for a year if they move their spending back to the card
“A year without the fee is fair, but I moved my spending because the other card earns more back. Fix that and I'd come back.”
- Rewards rate, not the fee
- Doubts the fee stays waived
- Switching cards is a chore
Directional
What sets CrowdOS twins apart
Four things that set CrowdOS twins apart
01
Built from behaviour, not stated preferences
Cohorts are derived from what customers did. Survey answers are added only when you have them, and are never invented.In build
02
Deliberation you can watch
Twins can be put in a room to argue, so you see how a position forms and which objections survive the discussion.Live
03
A ledger that shows the misses
Every benchmark run we publish lists its questions one by one, including the ones the crowd got most wrong. With your permission, your own campaign results would form a private ledger for your organisation.Live
04
Grounded in Southeast Asia
The market crowds for Singapore, Malaysia, Indonesia, Thailand, the Philippines and Vietnam are built country by country from national population figures — age, sex and region, and for Singapore, Malaysia and Indonesia ethnicity and religion — rather than from one global profile.Live
Compared
How do cohort twins compare with other ways to ask customers?
| Cohort twins | Per-customer twins | Synthetic respondents | Human panels | |
|---|---|---|---|---|
| Built from | Your customers' behaviour, grouped into weighted cohorts | Each customer's own record | Census or written profiles of a general population | Recruited people |
| Unit of answer | A cohort, weighted by its share of customers | One customer | A sample of personas | A sample of people |
| Aims to predict one named person | No | Yes | No | No; it measures the people it reaches |
| Needs your customer data | Yes, a file export | Yes, row level | No | No, a sampling frame |
| Time to a first result | Minutes, once cohorts are built | Minutes, once twins are built | Minutes | Days to weeks |
| Best used for | Ranking options and finding objections, by cohort | Individual next-best-action | Early reads on a general market | Precise population estimates and depth |
Works with what you already have
What data do I need to start?
Start with a CSV or spreadsheet export of your customer file: one row per customer, with tenure, products held, recent purchases, frequency, spend and last channel if you have them. You map its columns to a small set of feature groups, and direct identifiers are dropped on upload. CRM connectors are planned.
Every engagement is covered by a data-processing addendum, which sets the purpose, the retention period and the deletion terms for your data.
In build
What we don't claim
What customer digital twins do not do
Twins do not predict what a named individual will do. They do not replace probability-based research when you need a precise estimate for a population, and they do not replace talking to customers when you need to understand why.
Language-model respondents approximate groups better than individuals, and even then they miss. Pew Research Center's 2026 study of AI stand-ins for survey takers found large errors; our own ledger publishes ours, question by question.
We are building reliability tiers so that thin data, and questions the method handles poorly, are flagged on a result. Until they ship, treat every result as directional and check it against your next campaign.
From Atelier by CrowdOS
Digital Twins powered by CrowdOS
Atelier by CrowdOS is our bespoke services arm for industry, and Digital Twins powered by CrowdOS is its first solution. It is for businesses that want a model of their own: a private cohort crowd, a private calibration layer and, where your data supports it, a small model trained only on your customers and your outcomes, hosted in isolation and deleted when you leave.
By application
Questions
What is a customer digital twin?
A simulated respondent that stands in for a group of your real customers. It is built from that group's behaviour and answers a proposal in words: whether it would take it, what it likes and what it objects to. CrowdOS builds one twin per weighted cohort, not one per person.
How is a cohort twin different from a persona?
A marketing persona is a portrait someone writes. A cohort twin is generated from the measured behaviour of a real group of your customers, weighted by how many customers it stands for, and it can be asked questions and put in a room with other twins.
What data do I need to start?
A customer file export with several thousand rows and at least six usable features, such as tenure, products held, recency, frequency, spend and last channel. Names, emails and other direct identifiers are not needed and are dropped on upload. Our guide to preparing a customer file lists the columns.
Is my customer data used to train AI models?
Your data is used to build your cohorts and run your studies. Whether it may be used for anything else, including calibration on your own campaign outcomes, is set out in the data-processing addendum that covers every engagement. That addendum is being drafted with counsel; we will publish its terms here.
How accurate are the results?
We do not publish an accuracy figure for customer twins. The crowd's public benchmark runs, with every miss, are on our ledger. For your own twins the measure that matters is your campaigns: after a campaign, upload the results and see where the twins ranked offers correctly and where they did not.
Can twins replace a survey or focus group?
No. They are a fast way to narrow a set of options and find objections before you spend on research or media. When you need a precise estimate for a population, use a probability-based survey; when you need to understand why, talk to customers.
Which countries are supported?
Your customer file can come from any country, because the cohorts are built from your own data. The CrowdOS market crowds cover 24 markets, including Singapore, Malaysia, Indonesia, Thailand, the Philippines and Vietnam.
What is Atelier by CrowdOS, and what is Digital Twins powered by CrowdOS?
Atelier by CrowdOS is our bespoke services arm for industry. Digital Twins powered by CrowdOS is its first solution: a private twin model built for one business, delivered as an engagement with a success measure agreed before the work begins.
Contact
Tell us what you want to test
We are working with a small number of design partners. Tell us the decision you have coming up and how your customer data is held today, and we will reply within two business days.