Introducing the Cordial Decision Engine

Retail marketing runs on small decisions. For each customer, on each send: message now or wait? Email, SMS or push? New arrivals, free shipping or 20% off? A retailer with ten million contacts faces tens of millions of these calls a day.

No team can make those calls one customer at a time. So a marketer makes one decision for an audience of a million people, and a million people get treated the same: the same message, the same channel, the same offer, at the same time. Point tools help at the edges, each answering a single question like send time or churn risk, but nothing answers them together, for this customer, right now.

The Cordial Decision Engine does. It answers every decision about a customer in a single pass, in under 100 milliseconds, with a calibrated confidence on each answer.

Try it here, or keep reading to learn more about it:

Why System One

Daniel Kahneman described two modes of thought: System 1, fast and intuitive, and System 2, slow and deliberate. Large language models are System 2 tools. They’re the right choice for writing a brief or planning a quarter, and the wrong one for the millionth “send or wait?” of the day. At that volume, text generation is too slow, too expensive and too unpredictable to sit inside software.

System One models, a class TypeSafe AI named in September, work differently. You give the model a situation and a set of typed questions, each with a fixed list of possible answers. It returns one answer per question, each with a calibrated probability, in one pass. It can’t return anything that isn’t on the list.

That is exactly the shape of a retail marketing decision.

Why retail needs its own model

General decision models handle general questions, like routing a ticket or detecting intent. Retail decisions turn on things they have never learned: how a lapsed full-price buyer differs from a lapsed deal-seeker, when a replenishment customer is ready to buy again, when message fatigue starts costing subscribers, and when a discount earns incremental revenue rather than giving away margin on an order that was coming anyway.

We tested a strong general decision model on our retail benchmark. Every answer was well formed. On the decisions that matter most, including send or wait, send timing and audience forecasts, it did no better than chance.

So we built our own. The Decision Engine starts from an open-weight System One architecture and is trained on five families of retail decisions:

• Communication: send, wait or suppress; the best channel; likelihood to buy.
• Candidate scoring: which message or offer wins, on orders and on margin.
• Lifecycle timing: the right window for replenishment and win-back.
• Audience forecasting: a cohort’s purchase rate, and whether a campaign will hit its goal.
• Next best action: what should happen next in a journey, given what the customer has already received.

Training is ramping. With each run, the shared foundation covers more customer types, categories and scenarios. Each retailer’s own layer then trains on real outcomes as its programs run through Cordial.

Confidence is calibrated, so “70% likely to buy” means about 70%. That makes it usable: teams can act automatically on confident calls, explore within a budget, and send borderline ones for review.

What it changes

• Decide per customer, not per segment. At milliseconds and fractions of a cent per call, every message on the calendar gets its own decision: send, hold, suppress or switch channel.
• Decide in the moment. Abandoned carts, browse sessions and back-in-stock alerts can settle offer and timing when they fire, not in tomorrow’s batch.
• Replace point tools with one engine. Channel, timing, offer, propensity, fatigue and forecasting come from one model and one view of the customer, so the answers agree with each other.
• Add a question without adding a model. “Will she buy at full price if we hold the discount?” is a new input, not a new data science project.
• Know before you send. Forecast a campaign’s odds of hitting its goal, and which change to channel or offer would improve them, before the budget is spent.
• Give agents a System One. Agents plan the campaign. The engine makes every per-customer call inside it and returns a typed answer, not a paragraph to parse.

Tuned to each retailer

A beauty brand’s lapsed customer doesn’t behave like a home goods retailer’s, and a brand protecting margin wants different calls from one chasing revenue. So the engine has two layers: a shared retail foundation trained on simulated behavior, and a lightweight layer for each retailer, trained only on that retailer’s customers and outcomes, inside Cordial.

Your layer optimizes for your goals, whether that’s margin per send, list health or repeat rate. Two retailers can describe the same customer and both get the right answer for their business.

The learning compounds, privately

Cordial already sees far more than what happens to a message. Every send and its response: delivered, opened, clicked, unsubscribed. Every transaction, return and cart. Every search, browse and product view. Every store visit, loyalty event and anything else your systems send to Cordial. Until now, most of that fed dashboards and segment definitions, and a person decided what to do next. With the Decision Engine, all of it becomes training signal for your model, and only yours.

The compounding happens in your private layer, on top of the shared base. None of it flows back into the foundation, never reaches another retailer’s layer, and never improves a competitor’s decisions. Each decision produces an outcome, each outcome sharpens your next decision, and the advantage that builds up stays with you.

Most AI gets better by pooling everyone’s data. This works the other way: you get the benefit of a shared retail foundation without contributing your customers to it.

Just the start

The architecture is built to extend. With a shared foundation and a private layer per retailer, new signals and new predictions can be added without starting over.

Today the engine reads a customer through their behavior: purchases, discount sensitivity, channel engagement, lifecycle stage. Next, we’re layering in the attributes of what they interact with. For products, that’s the attributes of every item a customer views, carts or buys, enriched in each brand’s own language. For messages, it’s the attributes of every send they open, click or ignore: lifestyle or product imagery, headline tone, the number of calls to action, the offer.

That moves the engine from knowing that a customer clicked to knowing what she clicked on, and why it worked. It can then decide not just whether and when to message someone, but which product to lead with and which creative will land.

It’s also general. Any question with a defined set of answers can be put to the engine, without training a new model. Teams have found surprising uses for LLMs by asking them new things, and the same kind of invention applies here, to decisions instead of text. We’re already exploring journeyless journeys in Cordial: instead of a marketer mapping every branch in advance, the engine chooses each customer’s next step as their events arrive. It’s one of many ideas we’re testing.

We’re at the beginning of this, and we’re excited about what it will unlock for our customers.

See it for yourself

The Decision Engine is live in our demo. Describe a customer in plain English (“Maya hasn’t bought in three months, pays full price, loves new arrivals”) and get five decisions back in under 100 milliseconds, each with its confidence.

We’re building brand-tuned engines with a first group of retailers now. If you run retention at scale, talk to us.

Try the live demo

FAQs

What is the Cordial Decision Engine? 

It’s a System One foundation model for retail. In contrast to a typical LLM, a System One model calibrates per-customer decisions in milliseconds — which means you can actually run 1:1 personalization at scale.

What’s an AI decision engine for retail marketing? 

An AI decision engine for retail marketing is a model that makes the individual calls behind every message (whether to send, which channel, which offer, and when) for each customer, rather than for a whole segment. 

How does an AI decision engine work? 

An AI decision engine takes a description of a customer’s situation, answers a set of predefined questions about that customer, and returns one answer per question with a probability attached. The Cordial Decision Engine uses a System One architecture, a class of model named by TypeSafe AI in September 2026, borrowing Daniel Kahneman’s term for fast, intuitive thinking. 

Here’s how it works: 

• You give the model a situation and a set of typed questions, each with a fixed list of possible answers (for example, “send, wait, or suppress?”)
• It returns one answer per question, each with a calibrated probability, in a single pass
• It can’t return anything that isn’t on the list, so the output is always something the software can act on.

That’s what separates it from large language models. LLMs are System 2 tools: good for writing a brief or planning a quarter, but too slow, expensive and unpredictable to make the millionth “send or wait?” call of the day

How is AI decisioning different from traditional marketing automation?

Traditional marketing automation runs the rules and schedules a marketer sets for a segment. AI decisioning makes a separate decision for every customer, every time. With automation, one choice about an audience of a million people means a million people get the same message, channel, offer and timing. A decision engine works out each of those choices per person.

The Cordial Decision Engine changes the workflow in several ways:

• Per customer, not per segment: at milliseconds and fractions of a cent per call, every message on the calendar gets its own send, hold, suppress or switch-channel decision.
• In the moment, not in batch: abandoned carts, browse sessions and back-in-stock alerts settle offer and timing when they fire, not in tomorrow’s batch.
• One engine instead of point tools: channel, timing, offer, propensity, fatigue and forecasting come from one model and one view of the customer, so the answers agree with each other.
• New questions without new models: “Will she buy at full price if we hold the discount?” becomes a new input, not a new data science project.

The takeaway is that automation still runs the program, and AI decisioning makes the individual calls inside it.

How can a decision engine improve win-back and replenishment timing?

A decision engine improves win-back and replenishment by predicting the right window for each customer, rather than relying on a fixed delay set for everyone.

How does AI decisioning improve cart abandonment recovery? 

AI decisioning improves cart abandonment recovery by choosing the offer and timing for each abandoned cart the moment it happens, instead of applying one fixed follow-up to every shopper.

Can an AI decision engine reduce wasted discounts?

Yes. An AI decision engine reduces wasted discounts by predicting, for each customer, whether an offer will earn incremental revenue or just give away margin on an order that was coming anyway.