Why we started Tricorne Labs
Three engineers who lead robotics and AI work by day, and why we decided the software sold to small businesses deserves the same standard.
Between the three of us we have spent about eighteen years writing software that other people depend on. Robotics and machine learning programmes. Observability platforms carrying enterprise traffic. Data platforms inside the automotive industry. A health insurer's first mobile app.
None of those are forgiving environments. When a machine learning system gets something wrong, you do not hear about it in a support ticket — you find out downstream, once the wrong answer has already been acted on. You learn a particular kind of engineering discipline in places like that: you assume you are wrong, you make the system say how confident it is, and you never let it take an irreversible action on its own judgement.
Then we started looking at the software sold to small businesses.
The gap
An owner-operated insurance brokerage runs on email, a CRM, a spreadsheet, an intake form, and a carrier portal that has not been redesigned since 2011. The owner is the salesperson, the account manager, and the person who re-types the client's date of birth into the fourth system that afternoon.
There is no shortage of products aimed at them. Most are a thin wrapper over somebody else's API with a chat box bolted on the side. They demo well. They fall apart on contact with a real business, because a real business has fifteen years of context, a specific way of talking to clients, and a very low tolerance for a tool that emails a customer something wrong.
That last part matters more than anything else on the feature list. The reason owner-operators are cautious about AI is not that they think it is useless. It is that their business is their client relationships, and they are being asked to hand those to something that will confidently invent a policy detail.
What we decided
We started Tricorne Labs to build for that gap, with three commitments we took directly from our day jobs.
Confidence is part of the output. A classifier that returns an answer without saying how sure it is has hidden the most important number. Everything our models produce carries a confidence score and an escalation level, and those are visible to the user, not buried in a log.
The human gate is the design, not a setting. In our first product, no AI-written reply reaches a client until a person has read it and pressed approve. There is no configuration flag that turns that off, because the moment there is one, someone turns it on and the product becomes something we would not sell.
Build the whole stack. Data model, integrations, and the AI layer on top of both. You cannot draft a good reply from a thread you only half-parsed, and you cannot fix a bad integration with a better prompt.
Where we are
Our first product is Tricorne — an AI operations desk for owner-operated service businesses. It connects to Gmail or Outlook, sorts the inbox before the owner opens it, drafts replies from that business's own tone and policies, turns intake forms into client records, and holds every outbound message for approval.
We use it on our own inbox. Every rough edge we would ask a customer to live with, we live with first.
We are a small company and we intend to stay a careful one. If you run a business that lives in its inbox and you want to talk about it, we would genuinely like to hear from you.