Automation is a staffing decision

The first AI project in a services firm is rarely a science problem. It is a staffing problem. Someone is already copying notes into a CRM, drafting the same proposal paragraph, or reconciling a spreadsheet that should have been a system. A model can take the first pass. It cannot take the blame.

Start with a single queue of work that is frequent, text-heavy and already reviewed by a person you trust. Give the model the same instructions you would give a new analyst, including the examples of a bad answer. Measure two numbers: hours returned to the reviewer, and errors that would have reached a client if nobody had looked.

Do not begin with a customer-facing agent. The cost of a confident wrong answer in front of a buyer is higher than the cost of a slow draft inside the office. Internal first. Narrow second. Public last, and only with a named owner and a way to switch it off the same day.

Treat the licence like a hire. Who trains it, who reviews it, what it is forbidden to touch, and when you will fire it if the error rate does not fall. Firms that skip that conversation collect pilots. Firms that keep it collect time.

If you cannot name the person who checks the output, you have not automated the work. You have hidden it.

HI