About Companycoach.ai

We are a small team that builds AI systems and then has to live with them. Most of what we know came from operating our own products in production, at our own expense, where nobody was available to blame.

What we believe

Most AI projects fail for unglamorous reasons

Not because the model was not clever enough. Because nobody defined what success meant, because the integration to the CRM was never finished, because there was no way to tell whether the new version was better than the old one, or because the person who championed it left. We spend most of a project on those four things.

A model that produces numbers will invent them

Every figure in a system we build comes from a measurement and is handed to the model as a fact. The model orders, explains and writes; it does not count. This is a hard rule and it has saved us from more embarrassment than any other decision we have made.

Autonomy is earned

New systems launch with a human approving everything that leaves the building. Autonomy widens step by step, measured against a test set. Anyone offering full autonomy on day one has either not run a system in production or is not telling you about the incident.

The cheapest model that does the job

We host open models for the high-volume, repetitive work — classification, extraction, translation, summarisation — which takes the marginal cost of those requests to roughly nothing. Frontier APIs are for the difficult minority. This is not ideology; it is the difference between a system that pays for itself and one that is quietly retired at the next budget review.

How we choose projects

We take on work where the workflow is repetitive, the language barrier is real, and someone on your side will own the result. We turn down work where the volume is trivial, where the rules are already deterministic and better written as code, or where the buyer wants to be able to say they are doing AI.

We are a small team on purpose. It means we are honest about capacity, and that the person you speak to on the first call is the person who builds it.

What we refuse

Systems that give medical or legal advice to the public without a qualified human in the loop.
Fully automated rejection of job applicants.
Scraping platforms whose terms forbid it, or reselling data harvested from them.
Bulk outreach designed to evade consent, opt-outs or sending limits.
Anything pretending to be a human when the person asks whether it is.

These are refusals rather than negotiating positions. It costs us work, and it is the reason we are comfortable being audited by a client’s compliance team.

Our own products

We build and operate our own software as well as client systems: a lead-capture assistant used by medical clinics, an SEO analytics platform with its own crawler and link graph, and multilingual content systems running across a dozen languages. Running them is how we learn what breaks at three in the morning — which is the only part of this work that is genuinely hard to fake.