From legal intake to AI agents for compliance teams

We started with legal intake and matter management. I believed matter management and triage were the value. Our sandbox customers, Allianz, Bosch, Solaris and TD Synnex, ignored both after a few weeks; they only leaned in for the automation. We stripped the product back to intake alone.

Buyers had budget but no urgency. Meanwhile companies were moving money into AI, often without a use case, and the CEO’s strategy team picked Governance, Risk and Compliance, with company policies as the way in.

Every prospect session went into one funnel, in the prospect’s own words. Insights became candidate features, each tagged with the prospects it would serve. That gave the strategy team 3 regulated areas to test: OSHA rules for manufacturing, energy and construction; HR; and anti-bribery compliance, which applies to every industry.

Compliance bought the product. Everyone else had to use it.

The CEO wanted something distinct from the platform, purple instead of blue, before any brand existed. I went colourful and playful first, against the grey platform; the team loved it. Leadership wanted it more serious, so version 2 is calmer. The scope stayed at 4 features: upload, chat with those documents, references, feedback. Plus one fake page, a mocked analytics view, to see what prospects reached for.

Once the prototype existed, the PM and I joined prospect sessions instead of reading the notes afterwards. Notes tell you what was said; being in the room tells you what mattered. It confirmed something I now always push for: when the people building hear prospects first-hand, ownership goes up and the build moves faster.

Answers come only from the policies the company uploaded, never the open web. Every answer points to the passage it came from, highlighted, and every answer can be marked wrong. In a regulated space, that is what made people trust it.

The mocked insights on the first dashboard, contradictions, legibility and best practice, got the strongest reactions in calls. So they became the real Policy Agent overview. Cheap test first, then build.

We kept the trial customers close, watched how they used the product, and reworked the information architecture around roles: admins, compliance teams, experts and employees. Soft launch came in April 2024, before we had a brand, with 5 paying customers: Bertelsmann, Zimmer Biomet, Essity, TD Synnex and Datasite.

Policy AI became AI Agents, each one specialised in a single regulation. Commercial wanted agents in their own space, and we had 1 frontend developer. I explored options inside the existing UI and added one menu layer. It looked like a lot on paper; users got it straight away.

When demand pointed back to legal teams, the patterns carried over: projects for each review, chat grounded in the documents you attach, and extracted fields you can check against the source.

Look for demand signals before the build, not after it. We also lost the argument about feature names: users wanted "Chat", commercial wanted something that sounded sophisticated. I collected customer quotes to make the case.