Platform Convert & Engage Decisions

Beyond segments. One decision per person.

A real-time decision engine that ranks the next-best experience for every visitor, on every signal and every channel: which product, which offer, which message, or nothing at all.

  • Decides in milliseconds
  • Respects margin, stock and frequency
  • Every decision explainable
A road that splits into four paths, one highlighted and ending at a flag
Decided in 38 ms214 signals · 61 eligible actions

Next-best experience

  1. Lounge upgrade0.92
  2. Extra baggage0.71
  3. No contact0.12
Revenue per visitor Measured against a randomised control
Attach rate Relevant add-ons accepted
Discount cost Incentives only where they change behaviour
Catalogue coverage More of your range gets seen

Why change

A segment is an average. Nobody is average.

Rules and segments were the right tool when decisions were made weekly. They cannot keep up with a customer who changes their mind mid-session.

Big catalogues overwhelm

The products that fit the customer's intent stay buried on page six.

Every slot shows the same thing

Bestsellers appear whatever the session is telling you.

Commercial reality is ignored

Out-of-stock, low-margin and ineligible items waste the impression.

Scores are frozen overnight

Batch models miss what the customer did thirty seconds ago.

What it does

Every signal in. One best action out.

Predictive models estimate what each customer wants. Business rules decide what you are willing to offer. An arbiter weighs the two and picks, in real time.

01 Real-time arbitration

Offers, content and channel, ranked together.

Candidates from every team and every model compete in one ranking for each customer moment. The score blends likelihood, value to the business and cost of the action, and "do nothing" is always on the list.

  • Session-aware: the last click counts as much as the last year.
  • Anonymous visitors included, with cold-start fallbacks for new people and new products.
  • Sequence-aware, optimising the next few steps, not just the next click.
Returning shopperViewing running shoes · cart has socks Live
1Trail shoe in her size, in stock nearby Product page · "similar, better grip"0.88
2Insoles bundle Cart · completes the basket0.64
315% off voucher Email · likely to buy without it0.21
4Bestselling road shoe Filtered · her size out of stock—
38 ms · 214 signals · policy v31Illustrative

02 Rules you can read

The business sets the boundaries. The model works inside them.

Eligibility, margin floors, inventory, contact policy and fairness constraints are written as plain rules that merchandisers and marketers own. Change one and simulate the effect on last week's traffic before it goes live.

  • Merchandising overrides and pins, with an expiry so they do not linger.
  • Net-value optimisation, accounting for likely returns and cancellations.
  • Diversity and exploration, so the engine does not collapse onto ten bestsellers.
Policy · Homepage "For you"v31 live
In stock at customer's nearest hubHard filterOn
Contribution margin at least 18%Hard filter · owner: FinanceOn
Max 3 promotional contacts per weekShared with every channelOn
At least 4 categories per row, 10% explorationSoft constraintOn
Pin: monsoon collection in slot 2Expires SundayTemporary
−2.1%simulated clicks
+6.4%simulated margin
92%catalogue coverage
Simulated on last 7 days before publishingIllustrative

03 Explainable by default

Ask why, and get an answer.

Every decision is stored with the signals used, the candidates considered, the rules applied and the probability with which the action was chosen. That makes it debuggable for engineers, auditable for risk and usable for honest measurement.

  • Reason labels you can show the customer: "because you viewed…".
  • Decision replay, to see what a new policy would have done.
  • Graceful degradation, with cached fallbacks if a model is slow or unavailable.
Why this decision?Decision 7f3a…e21
Viewed 3 trail shoes in this sessionStrongest signal+0.31
Size 6 in stock 4 km awayEligibility passed+0.19
Bought running socks last monthComplementary category+0.14
Rarely buys on discountVoucher ranked down−0.11
Returned a road shoe in MarchSimilar style ranked down−0.07
Chosen with probability 0.81 · 0.19 reserved for explorationIllustrative

Agents at work

Models decide. Agents keep them honest.

The decision itself is made by statistical models in milliseconds. Agents work around it: proposing new candidates, watching for drift, simulating policy changes and explaining results to the people accountable.

  • Drift and performance monitoring, with automatic fallback to the last good policy.
  • Policy registry: version, deploy, compare and roll back.
  • Fair-exposure analytics, across sellers, categories and customer groups.
Agent activity  Live
Policy v31 promoted from 10% to 50% trafficGuardrails healthy for 48 hoursAutomated
Stock feed stale for footwearFell back to cached availability, flagged ownerAutomated
Proposed: lower margin floor to 15% for clearanceSimulated +3.8% revenue, −1.2 pts marginNeeds approval
Exposure skew detected: 61% to two brandsDiversity constraint tightened within limitsAutomated
Candidate offer using health-condition signalSensitive attribute not permittedBlocked
Every step logged with inputs, reason and outcomeIllustrative

You decide what the agents may do here.

AutomatedInside the policy

Ranking, exploration, retraining on schedule, staged rollouts and fallback when something degrades.

Human-approvedChanging the policy

New objectives, margin and frequency rules, new data used as a signal and full rollout of a new model.

RestrictedHard stops

Sensitive attributes as signals, offers a customer is not eligible for and pricing outside approved bands.

What's inside

A decision engine, not another recommendations widget.

One service for web, app, messaging, contact centre and store, so the customer gets a consistent answer everywhere.

Event stream

Impressions, clicks, carts, purchases and returns in real time.

Catalogue and stock sync

Product metadata, availability and price kept current.

Recommendation strategies

For you, similar, related, trending and frequently bought together.

Next-best-action arbitration

Offers, content and channels ranked in one list.

Anonymous and cold-start

Sensible decisions for new visitors and new products.

Low-latency APIs and SDKs

Ranked lists served in tens of milliseconds, with fallbacks.

Business rules

Price, stock, category, margin and contact policy as readable constraints.

Placements and overrides

Merchandising control with expiry dates.

Exploration and diversity

Avoid popularity loops and keep learning.

Explanations

Reason labels and full decision logs.

Simulation and replay

Test a policy on past traffic before it sees a customer.

Model registry

Retraining, monitoring, versioning and rollback.

How it runs

From a click to a decision that learns.

Each step is a named, logged stage. Nothing autonomous is switched on until its data, permissions, monitoring and rollback have been checked.

  1. Capture first-party events
  2. Sync catalogue and stock
  3. Build session and customer features
  4. Generate candidates
  5. Rank and apply rules
  6. Serve through a low-latency API
  7. Test outcomes and retrain

How you'll measure it

A decision is only good if it beats the control.

We agree the numerator, denominator, cohort and success gate with you before anything goes live, and compare against a baseline or a holdout, never against a hunch.

  • Incremental revenue per visitorRevenue per eligible visitor in treatment minus control.
  • Recommendation conversionOrders after an eligible exposure, compared with a randomised control.
  • Cross-sell attach rateOrders containing the targeted add-on, divided by eligible orders.
  • Net value after returnsContribution after returns and cancellations, not gross sales.
  • Catalogue coverageDistinct eligible products recommended, divided by eligible products.
  • Serving latency (p95)The 95th-percentile response time at agreed load.

Questions

Things people ask before they start.

How is 1:1 decisioning different from segmentation?

Segments group people and treat the group the same. Decisioning scores every eligible action for each individual at the moment of interaction, so two people in the same segment can and often should see different things.

Will it just push whatever sells most?

No. Popularity loops are a known failure. Exploration and diversity constraints keep the range visible, and coverage is reported so you can see it.

Can merchandisers still control what is shown?

Yes. Pins, exclusions, margin floors and campaign priorities are rules they own, and each change can be simulated on recent traffic before it is published.

What happens if a model is slow or fails?

The API answers from a cached fallback inside the latency budget, the incident is logged, and the last known good policy stays in place until the issue is fixed.

Composable by design

Useful alone. Sharper when connected.

It can rank from a product feed and a click stream alone. Connected, it knows the whole customer and proves its own value.

Keeps your existing stack

Serves your website, app, messaging and contact centre through one API. Reads catalogue and stock from your commerce platform.

Runs where your data lives

Small language models trained on marketing data are deployed inside your cloud or data centre, so customer data and the models reasoning over it stay under your control.

Getting started

Start with one placement and a fair test.

Choose one high-traffic placement, connect events and catalogue, and run the engine against your current logic with a randomised split.

  1. Connect events, catalogue and stock
  2. Agree rules, objective and guardrails
  3. Run a randomised test on one placement
  4. Roll out, then add the next surface