Pilot proposal
HilbertHilbert decides who gets which offer, and proves what each one earns.
Signal38% of the growth budget goes to discounts.
~120 · Promotions live per weekDiscounts are set by category, not by what each shopper is worth.
The minimum discount per shopper tier, and the budget it frees.
Most new shoppers stop before their third order.
A weekly list of new shoppers to nudge, before the drop.
Every lapsed shopper gets the same discount, whether they come back for good or once.
Win-back sends sized by the margin each shopper brings back.
Store teams order from last year's pattern.
All categories
At least TBD: X% more net contribution margin in the Hilbert group than under today's rules.
The minimum discount per shopper tier, and the budget it frees.
A weekly list of new shoppers to nudge, before the drop.
Win-back sends sized by the margin each shopper brings back.
Same shoppers, same budget, same period. The result is measured in your own A/B tool.
Chosen by Hilbert, at the moment Hilbert recommends.
Today's rules, unchanged, with the same budget.
No action. Shows what happens without any change.
What each shopper is worth over time.
Who is slowing down, before they leave.
Which lapsed shoppers will return.
What each shopper buys next.
For every shopper, recomputed daily, stored in Orchard Market's own data.
At least TBD: X% more net contribution margin in the Hilbert group than under today's rules. Read against a control group, on a rule written down before the start.
Every result feeds the next round, so the gains build over time.
How the team finds, sizes, decides, activates and measures, inside Hilbert.
What changed.
Why it likely happened.
What to do, and for whom.
Anyone on the team asks in plain language, and the answer comes with its sources.
Category promos went to a wide pool: anyone active in the past year who hadn't bought the category in the last month, with no read on who was likely to buy it next.
Hilbert's predicted-intent audiences, 53–72% smaller, tested head-to-head against the old pool across four categories, October to November 2025.
More than 2× purchaser rates, 35% vs. 12% post-promo retention, and ~$6M a year from purchaser lift, retention and discount savings.
Invoiced once, at the decision gate; 45-day terms.
The floor only applies below $150,000 of extra net contribution margin. Above it, Hilbert is paid only from what it brought.
After the gate, a licence sized to scope (products, surfaces, model families), never to usage.
If Hilbert brings $250,000 of extra net contribution margin
3.3× what Orchard Market pays
Illustrative. The real result is measured in your own A/B tool.
Scores, audiences and test results stay with Orchard Market, whatever the outcome.
Every opportunity Hilbert surfaces and every one activated is logged in one shared list.
Metric, groups, window and read-out date are written down before the test and not revisited.
Both sides agree the test design in writing; neither shapes it to inflate or understate the result.
If the models do not hold on Orchard Market's own history at the backtest, either side can stop before the live test at no cost.
Service levels and order forms are set in the Pilot SOW.
The test window has to avoid it.
Finance wants a number before planning closes.
Scores ready to use in the new programme.
Usually 3 to 6 weeks. Business sessions and technical discovery run in parallel.
Procurement and legal close the agreement.
One or two sessionsAgreement ready to sign.
Backtest · 27 Nov 2026
Decision · 4 Jan 2027
Integration plan: what data, how, and how long.
Shared data model, a score for every customer, checks on last year's history, platform access.
How well the models would have predicted last year's results. Either side can stop here at no cost.
Hypothesis, groups and period fixed in writing.
Daily reads and a weekly written read-out.
The result against the agreed lift, the history checks and the proposal to scale.
Sets the final scope, approves the agreement and decides at the gate.
Two sessions and the decision
Chooses what is tested, sets the target lift, reviews weekly and approves each list before it runs.
One scoping session, then 30 minutes a week during the test
Data in scope, access, and customer identifiers as they stand today.
About two hours of discovery, then two or three technical sessions
The activation tools, the control group in their campaigns, and the scores inside their tools.
One working session, then the weekly review
Agrees how the incremental result is counted, in their own margin terms.
One session before the test
TBD: account lead as the single point of contact; three forward-deployed engineers for the data model, the models and the connections; a growth team that designs the activation with your team; founders at kickoff and at the decision.
Full time from kickoff to the decision
Everything needed to sign, in one place. The status is as of 6 Oct 2026.
Limited to the data in scope, for the length of the PoC, and revoked at the end.
Scope, metric, groups, dates and service levels in one document.
45-day terms
Hilbert answers your questionnaire or holds one session with your security team.
Hilbert processes Orchard Market's data only on Orchard Market's instructions and does not use it for any other client.
Signed before any data is shared.
HilbertThe proposal in four lines, and the one decision ahead.
Next step · Next meeting · 6 Oct 2026
The tools Hilbert reads from and sends results back to, as listed on hilberts.ai.
Results from comparable Hilbert deployments.
Integration effort · ~98% Hilbert · ~2% Orchard Market
The metric is contribution margin after discounts and media. Not sales, not return on ad spend.
It has to beat what your team already does, on the same shoppers.
Your experimentation tool reads the result, with definitions fixed before the test.
Hilbert carries the integration; your team provides access and decisions.
Hilbert sits between the client's data and the tools its teams use: ingestion, a shared data layer, company-specific features, the models, the agents and activation, as an extension of the stack rather than a replacement.
The schema's meaning is captured at onboarding, business rules live in a workspace context, conflicts are surfaced instead of resolved silently, and every answer states its filters, windows and assumptions.
Deep-learning scores for every customer from the first transaction (value, drift, win-back, next purchase), refreshed daily and trained on the client's own history.
Models are checked with the standard ranking and calibration measures and watched for drift; the agent layer runs through unit tests, expert and model judges, retrieval checks and reference outputs.
Order history, product catalogue, channel attribution and behaviour in; scores and audiences out to paid and CRM. No personal data: identifiers are hashed at source.
AWS-native, encrypted at rest and in transit, each organisation's data and agent memory kept apart, single sign-on with forced MFA, role-based access, SOC 2.
Every result is read against a group we don't contact, never before and after; paid channels get causal lift.
The features, the model logic and the outcomes are open to the client's data team and discussed, not hidden inside a vendor product.