Datmos finishes Top 5 in the Bloomreach Hackathon with an AI winback agent

2 octobre 2026

Every retailer loses customers. The hard part is catching them before they're gone. A loyal customer goes quiet, and by the time someone builds a winback campaign by hand, that customer is already buying somewhere else. Churn rarely makes noise. It just shows up later as lower repeat revenue.

That's the problem Datmos took on at the Bloomreach Hackathon, held September 21 to 28. We finished in the Top 5 with EverLoyal, an AI agent built for customer retention. Every morning, the agent finds the customers slipping away and sends each one a personal offer. By 9 AM, the emails are in their inboxes. Nobody builds a campaign by hand.

Here's what the agent does, how it works, and what it taught us about trusting AI agents with real customers.

What was the Bloomreach Hackathon challenge?

Bloomreach, Google Cloud, Shopify, and Databricks challenged teams from around the world to build one AI agent that gets all four platforms working together. Teams had eight days. Only working demos counted.

More than 250 active participants joined, forming 83 teams. Only 26 teams made it to a final submission, and EverLoyal placed in the Top 5 of those.

Why is winning back customers so hard to do by hand?

Winback is simple on paper. You find customers who stopped buying and give them a reason to come back. In practice, it's a chain of manual steps that rarely gets done well or on time.

A marketing team has to export customer lists, write the copy, and create discount codes. Then it builds each campaign one by one. After the send, it often has to guess whether the coupon drove the sale or the customer would have come back anyway.

So we asked a direct question: how much of that work could run on its own, with a human only checking the result?

What the EverLoyal agent does

EverLoyal handles the full winback cycle in four moves:

It spots customers who are drifting away.
It writes a personal email for each one.
It sets a discount your margin can afford.
It lets a marketer review everything and stop the send in one click.

The agent works fast, but it never goes off script. It only chooses from messages and offers your team has approved. And it explains every choice in plain English.

How does the AI agent run each morning?

Think of EverLoyal as a marketer who starts at 5:30 AM and has never needed a coffee. Here's a day in the life of the agent:

Orders arrive. Shopify orders land in Databricks.
Drifting customers are found. Databricks flags three groups: at-risk customers, VIPs going quiet, and one-time buyers.
Copy is written. Gemini writes the emails, picks the best version for each person, and explains why.
The offer is set. An offer agent checks margin, sets today's discount, and creates single-use Shopify codes.
The journey is built. Bloomreach's Loomi Connect builds the journey through the Model Context Protocol (MCP). It holds back a control group: 90% of customers get a coupon and 10% don't.
A human checks. QA receives all six email versions. One click can hold the send.
Emails go out. Personal emails reach inboxes by 9 AM.
The day is summarized. Loomi and Gemini write a daily summary of what happened.

Watch the EverLoyal agent demo

How does the agent learn from each send?

The loop doesn't stop at the inbox. When a customer clicks "Use my code," a Databricks App opens with the discount already applied.

Orders, opens, and clicks then flow back into the system. That data shapes tomorrow's emails. Each morning, the agent starts with more context than it had the day before.

The moment it clicked: the agent protected margin on its own

During testing, margin dropped to 27.1%. The target was 35%. No one stepped in. The agent saw the gap and cut the discount from 15% to 10% on its own.

The next day, margin came back at 43.8%. The agent raised the discount back to 15%.

This is the behavior that matters. A discount that ignores margin can win back a customer and still lose money. EverLoyal treats margin as a rule, not an afterthought.

What did we learn about building AI agents marketers can trust?

Eight days of building taught us four lessons. Each one applies well beyond winback.

Explainability. Every pick comes with a plain-English reason. If a marketer can't see why the agent chose a message or a discount, they won't trust it with their customers.

Guardrails. The AI only chooses between options people have approved. It decides fast, but inside limits your team sets.

Control groups. Without a control group, you can't tell if the coupon worked. The 90/10 split shows whether the offer drove the sale or just gave away margin.

Human in the loop. One check at 7 AM, where it counts. The marketer reviews the output, not every step that produced it.

Why does this matter for retail and commerce teams?

Your customer data, your margin rules, and your brand voice are things a competitor can't copy or buy off the shelf. EverLoyal shows what happens when you put them to work.

This is what we mean by Talent at the Core, AI at the Edge. Your team sets the rules and the standards. The agent does the repetitive work at the edge, every morning, without dropping a step.

Want to see EverLoyal in action?

Guillaume Thélot Parthon de Von

Director of AI

Guillaume is Director of AI at Datmos, bringing 11 years at Schlumberger, from field engineer to Data & Operations Excellence Manager, plus three co-invented patents in automation and machine learning. He leads Datmos's AI Task Force, turning operational discipline and deep technical expertise into AI capabilities that drive measurable impact for clients.