AI enablement for a Canadian ISP, in three months.

Uniserve is a Canadian internet and telecom provider. A high volume of repeat questions arrives through their residential support channels. The mandate came from the owner, not from a department: a three month AI enablement engagement aimed at Uniserve's own customers, with a working system at the end of it rather than a strategy document.

A technician spends part of the day answering something that has been answered before. A customer who gets the real answer at 11pm does not need a technician at 9am.

Most of the repeat questions already had an answer written down somewhere the customer could not find.

95% of enterprise AI pilots show no measurable P&L return (MIT NANDA, The GenAI Divide, 2025). 93% of AI budgets go to technology and about 7% to the people expected to use it (Deloitte, 2025). This engagement was scoped around the people, and around a hard hosting constraint no generic tool met.

We read the business for three things: where work repeats, where information is re-typed from one place into another, and where a person waits on another person before they can finish. That is the map, and it is the part most AI projects skip.

Then we cut it. Each candidate was written up with the arithmetic beside it, so the ranking was arguable rather than asserted, and the shortlist was chosen on return plus feasibility, in that order. Anything that could not be hosted in Canada was struck for that reason and no other. Nothing was chosen because it demos well.

The chatbot answers residential support questions in writing. The voice agent answers the same questions on the phone, for the customer who would rather talk than type. A customer gets the same answer whichever one they use.

Both were designed to the hosting requirement rather than around it. The data and the models sit in Canada, and that decision was made before the first line of code rather than patched in afterwards.

A request lands in the support queue. It is classified, routed and summarised. A suggested reply is drafted and held. Repeat questions are answered from Uniserve's own material rather than the open internet, with the source attached so the answer can be checked in one click.

A technician reads a short version instead of a thread, and a person still releases anything that goes back to a customer. The gate is the point: the draft sits until a technician reads it.

Roughly one hour of a technician's day given back once customers can get the answer themselves. 1 hr/day × about 250 working days × about $40/hr = about $10,000 per technician per year, scaling with team size. This is a forecast agreed as a target, not a booked result. Uniserve's team size is not published here.

Projection, Uniserve programme model. Nothing on this page is presented as a finished result.

The bar was agreed in writing before the build. When the system is in the team's hands, the measurement window opens, and the arithmetic will be printed here the same way as everything else on this site.