Digital Twin.
Rehearse the campaign before reality runs it.
Every send is an experiment on real customers. The twin lets you run it on the model first.
The digital twin takes the behavior model and runs it forward: your campaign played against every member — and, in parallel, the counterfactual of deliberately doing nothing. Member by member, rolled up into an honest forecast of the difference, calibrated against held-out reality. Before a single message goes out.
Your draft campaign
audience · offer · timing
Simulated: send it
the model plays every member forward under the campaign
Simulated: do nothing
the same members, played forward untouched
The expected difference
with confidence · calibrated against held-out reality
Why simulate
The most expensive way to test a campaign is to send it.
A/B tests spend real members
Live experiments are the gold standard — but every arm is real members getting a real treatment, and a bad variant has a real cost. Simulation triages first, so live tests are spent on the questions worth testing.
“Compared to what?” goes unanswered
A campaign that “made $50K” might have made $45K by itself. The twin always runs the doing-nothing counterfactual, so the number you look at is the difference — the only number that matters.
Intuition doesn't scale to segments
You can guess how one member reacts. Nobody can guess how fifty thousand different members each react to the same offer. The model plays them individually and adds it up honestly.
Under the hood
How a rehearsal actually runs.
A rehearsal isn't one big average — it's your member base played forward one member at a time. The behavior model already knows each member's rhythm; the twin asks it, member by member: with this campaign, what does the next stretch of behavior look like? And without it?
The two futures are aggregated honestly — segments too small to trust are suppressed rather than averaged into fiction — and the difference between the arms is the forecast. Change a lever and run it again: the audience, the offer, the channel, the timing, or the decision to send at all.
The levers a rehearsal can vary — including the one most tools forget: not sending.
audience, offer, channel, timing — a real draft, not a hypothetical
one arm gets the campaign; a mirror arm gets deliberate silence
the model plays every member's next stretch of behavior, in both arms
roll-ups carry confidence; too-small segments are suppressed, not smoothed
campaign minus silence — the only number worth deciding on
What it does
Three ways to ask the twin “what if?”
Pre-flight a campaign
Before sending, play the draft against the model of your base and read the expected difference versus silence. Send it, fix it, or skip it — informed, not hopeful.
Compare the options
Two offers, three audiences, this week or next — run the variants through the same simulated base and see which is worth a live test at all. The twin narrows; the live experiment decides.
Ask your research twins
De-identified member twins, grounded in real first-party profiles, that you can put questions to: concept reactions, survey pre-tests, message framing — an always-on research panel that never fatigues. Results come back aggregated with confidence labels, never as individual voices.
Calibration
A rehearsal you can trust — because it's graded against reality.
A simulation is only useful if it's honest about how good it is. The twin's forecasts are continuously checked against held-out reality — real outcomes the model never saw — and every number it shows carries that calibration with it.
Where your data can't support an answer — a segment too small, a behavior too new — the twin says so and stays silent. It will tell you not enough signal before it tells you a story. And a simulation never replaces a live experiment; it decides which experiments are worth running.
The model it runs forward is the learned half of the world model — Behavior Model →
Research twins
Early accessA research panel that never fatigues — grounded in your real members.
Anyone can ask a chatbot to “pretend to be a 34-year-old shopper.” That's a persona — a stereotype with a name. A research twinis different: it's grounded in a de-identified profile of a real member of yourprogram — real purchase rhythm, real category mix, real engagement pattern — so its answers reflect your base, not the internet's average customer.
Put questions to the panel the way you'd brief a research agency: reaction to a concept, appeal of an offer, framing of a message, a survey you want to pre-test before spending real members' goodwill on it. Answers come back in minutes, aggregated across the panel — a rehearsal for research, the way the simulator is a rehearsal for campaigns.
The rails it runs on
De-identified twice
twins are built only from an allow-listed, de-identified slice of the profile — identity never enters the panel
Aggregate only
results are reported across the panel with suppression below sample thresholds — never as an individual member's voice
Confidence ceilings
concept-level questions are capped at low confidence by design — the panel informs judgment, it doesn't replace validation
Collapse & sycophancy checks
the panel is continuously tested for collapsing into one voice or telling the asker what they want to hear
Where it fits
The twin triages. The experiment decides. The holdout proves.
Simulation doesn't replace live measurement — it makes live measurement affordable by spending it only where it matters. Three instruments, three jobs:
| Digital Twin | Live A/B test | Holdout group |
|---|---|---|
| Before the send. Costs nothing real; narrows many options to few. | During the send. Spends real members; decides between the finalists. | After the send. Untouched members prove the lift was real. |
| Answers: which ideas are worth testing at all? | Answers: which variant actually wins? | Answers: did the campaign beat doing nothing — for real? |
Flash runs all three — Experiments & A/B →
What the twin will never tell you
- A guaranteed revenue number. Forecasts carry confidence, and confidence is part of the answer.
- Anything about a segment too small to support an honest estimate — those are suppressed, not smoothed.
- An individual member's voice or future. Twins and rollouts report only in aggregate.
- That you can skip measurement. A rehearsal picks what's worth testing; reality still gets the final word.
If a simulation tool promises you certainty, it's describing its marketing, not its math.
Pre-flight forecasts surface where campaigns are built — with the counterfactual always in view.
Related reading
Keep exploring the pages most related to this one.
Intelligent Decision
A next-best-action: who, which offer, which channel, when.
Read more CapabilityDashboard
Real-time scan-to-redemption data as KPI trends, funnel diagnostics, and a rule + LLM next-best-action panel — plus retention curves by signup cohort and a 14-day revenue outlook that's honest about being a projection, not a commitment.
Read more CapabilitySoTag for Slack
The Slack agent that hands governed, audited numbers to Claude in-thread: daily anomaly scan + human-approved coupon-send. Early access.
Read more CapabilitySoClaw — Hire your AI employees
Hire AI employees for your loyalty program — autonomous agents that each hold a real job: a Loyalty Manager that wins back at-risk customers, an Order Manager that clears the receipt and order review queue, a Risk Manager that protects points from fraud, and a Marketing Manager that turns a goal into a ready campaign. Each works one case at a time within limits you set and logs every decision; you manage them like a team with goals, budgets, tasks and feedback. Off by default, capped, instantly pausable. Early access.
Read more CapabilityAI Governance & Approvals
Every AI action runs on a per-action authorization gear — automatic, human review, or forbidden — with an approvals console, a complete decision log, one-click undo, and automatic demotion to human review when an action starts misbehaving.
Read more CapabilityBusiness Ontology
The concept behind the governed AI stack: a world model of your business — objects, properties and relationships plus governed actions — as a metadata layer mapped over the data you already have, so AI knows what a customer is, what has happened, and what it may do about it.
Read more