Clarm started as an AI inbound conversion engine: a chat widget that captured website visitors, qualified buyer intent, and routed the good conversations to the right next step. It worked. One healthcare customer saw 6.1x more inbound conversations on the same traffic. We could have spent years just on that.
We moved to building a no-code AI agent builder instead. This is the honest account of why, because the reasoning is the same reasoning a customer uses to decide whether to trust us with more than one workflow.
The chat was the easy part
A chat widget that answers from a generic model is a weekend project and a liability. To make ours genuinely useful, the answers had to come from each customer’s own documents, not the public internet, and a customer in healthcare or banking had to be able to trust what it said in front of their users. That turned out to require a lot of machinery underneath the chat box:
- A memory layer that indexed each customer’s documents and retrieved the right passages.
- Source citations on every answer, so a person could check it and an auditor could trace it.
- A way to say “that is not in your data” instead of guessing.
- Tenant isolation, so no customer’s data could leak into another’s answers.
- An approval path for anything sensitive, and an audit trail of what happened.
- Model portability, so a customer was not married to one AI vendor’s pricing or policy.
Building the chat took weeks. Building the machinery that made the chat trustworthy took the rest of the time. And that machinery was the part customers in regulated industries actually cared about.
The realization: the machinery was general
The pattern showed up on call after call. A customer would see the inbound chat working and ask: “Can it also draft our client review notes?” “Can it run our weekly allocation pack?” “Can it screen new customers against our checklist?” Every one of those was a different workflow. And every one of them needed the exact same machinery: grounding in approved data, source citations, a human-approval step, an audit trail, isolation, a choice of model.
That was the realization. We had not built an inbound chat product with some infrastructure behind it. We had built the infrastructure to make any AI workflow trustworthy, and the inbound chat was just the first workflow we ran on it. Once you can build one governed workflow, you can build any governed workflow. The hard, valuable, general thing was the machinery for building workflows itself.
So we made the machinery the product
The inbound chat did not go away. It became one consumer of the substrate: a surface that answers live questions grounded in your data. Alongside it sits the other consumer: agent workflows that draft scheduled or triggered work for a human to approve. Same memory, same source citations, same approval gate, same audit trail underneath both. We stopped selling a single workflow and started selling the machinery that builds them.
In production that looks like a Swiss private bank drafting client review packs from a voice memo, a fresh-produce importer drafting weekly allocations from its own systems, a healthcare team that went from email-only deflection to chat, voice, and integrated agents — roughly 8x its case volume in twelve months — on one shared knowledge base. Different workflows, one set of machinery, the customer in the approval seat throughout.
Why this is a better deal for the customer
Buying a single AI workflow is a bet that the one workflow you picked is the one you will still need in a year. Buying the machinery is different: you get the workflow you need today, and the next workflow ships in days rather than quarters because the hard part is already built and already approved by your compliance team. The substrate compounds; the model is the part you can swap. That is the case we now make, and it is the case our own history backs up.
Where this leaves us
Clarm is a no-code AI agent builder for non-technical operators in high-trust teams, with the chat product as one of the things it can do. If you want the longer technical version, read What Is Atlas? and From Chatbot to Agent Platform. If you want to see it on your own data, book a pilot discussion.