How Alma scaled AI from personal productivity to shared company infrastructure

Alma is a leading Buy Now Pay Later (BNPL) solution built for merchants across Europe. Headquartered in France and active across 10 markets, Alma processes millions of transactions a year and serves thousands of merchants - each with different languages, business models, and payment expectations.

At that scale, operational precision and cross-team coordination quickly become critical.

When Alma deployed Dust company-wide in September 2025, adoption didn't happen overnight. It grew progressively: team by team, use case by use case.

To support that rollout, Alma created a cross-functional Dust Squad spanning Revenue, Ops, Product, and Engineering. Most departments had a dedicated point of contact with protected time every week to test workflows, surface blockers, and share what was actually working. Squad members embedded themselves inside operational teams, ran open training sessions, and gradually built a network of champions. Dust also became a recurring topic at Almatime, Alma's company-wide monthly gathering.

Some teams found strong use cases quickly; others took longer to get traction. Squad members acted as deployment strategists to drive adoption: they mapped who was using Dust and who wasn't, dug into the reasons behind slow uptake, explained features, and identified which internal tools needed to be connected to the platform to unlock more powerful agents.

Within six months, 87% of the company was active monthly, with 93% of those users still active the following month.

Early on, the impact looked like what most companies experience with AI adoption: faster drafting, easier access to information, more structured problem-solving. Over time however, Alma started using Dust less as an individual assistant and more as operational infrastructure shared across teams.

The shift showed up in the numbers. Raw blank-chat LLM usage dropped from over a third of all conversations to just 14% today. Meanwhile, 37% of conversations now connect to external tools, and 78% run through a custom or platform agent.

Teams weren't just using AI more. They were building systems that other people could use and improve. Three examples show what that looked like in practice.

Revenue teams need rich context on every merchant they approach: business model, payment setup, technical stack, market specifics, and account history.

To support that work, the Revenue team built @prospection, a shared agent connected to Salesforce. It helps prepare outreach by qualifying opportunities and generating personalized emails, LinkedIn messages, and call scripts adapted to the merchant's market, language, and role. By the time a sales rep starts engaging with a prospect, much of the groundwork is already prepared.

Today, around 90% of the Sales team uses the agent across segments, generating more than 2,000 conversations per month.

The deeper impact was on how the team works together. Feedback from one market improved outputs for the others; refinements made by one team became improvements for everyone. @prospection became a shared layer the whole Revenue organization iterates on, not a personal productivity tool.

Product documentation is useful across the company, but keeping it up-to-date is usually difficult to prioritize.

At Alma, Product Managers already write product briefs at the end of projects: what changed, why it matters, rollout details, operational impacts, and dependencies across teams. The issue wasn't creating information - it was turning that information into documentation that stayed maintained over time.

The Product built @productopia to close that gap. When a project closes, the agent reads the brief and posts a suggested documentation update in a dedicated Slack channel. The PM reviews it, adjusts the wording if needed, and publishes the final version to Notion. The agent took an existing step in the process and made it faster and less likely to be skipped.

Within one quarter, up-to-date product documentation jumped from 55% to 87%, while Product Managers reported spending less than half the usual time maintaining it. @productopia was later expanded into a company-wide product agent, letting any employee ask product questions directly, with answers grounded in documentation that updates as projects close.

Alma's Operations and Risk teams manage thousands of complex interactions every month. Reviewing payment fraud cases requires gathering information across several internal systems, applying the correct framework, and assembling structured reports, all before the actual investigation work can begin.

The analysts doing this work built an assistant agent themselves. It retrieves relevant payment information, organizes supporting context, and generates structured reports for review. Analysts still make every decision - the agent handles the assembly work that previously consumed time before they could get there.

The team didn't wait for a product or engineering roadmap to solve this. The people who understood the problem best built the solution directly, then embedded it into their browser environment so it fit into existing tools rather than adding another step to the workflow.

That's what decision augmentation looks like in practice: not replacing judgment, but removing the operational overhead that gets in the way of it.

Alma is still early in figuring out what AI adoption at scale should look like in practice, and many of these workflows continue to evolve.

One pattern has become increasingly clear: when the people closest to operational problems are equipped to build and improve solutions themselves, useful ideas spread faster and compound across teams.

The next phase is about being more deliberate. The Dust Squad now has full department coverage, backed by explicit leadership buy-in. The question has shifted from "are teams using AI?" to "what behaviors are actually changing?”

To answer that, Alma is mapping agents across two dimensions: what they do (improving an existing process, enabling a new one, supporting a decision, or making knowledge accessible to a team) and how critical they've become (does the process stop if this agent disappears, or does someone just work a little slower?). Together, those two questions shape how each workflow gets maintained, invested in, and evolved over time.

The foundation is in place. The work now is making sure it scales in a way that's deliberate, not just fast.