AI Integration

Tasks that can tolerate an occasional error and are reviewed by a person: summarising, classifying, drafting, extracting data from text. What isn't worth it: a decision with money or legal commitment attached, unreviewed. The model is wrong confidently — more dangerous than being wrong obviously — so correct use assumes the error and builds review on top of it.

A question asked before choosing, not after. It depends on the provider and the plan: some use inputs for training by default and some don't. We define what is sent at all, and keep out what must not leave. And if the data is genuinely sensitive, a self-hosted option exists and costs more — a trade-off presented plainly rather than hidden.

It is billed by consumption rather than a flat subscription, and that surprises people. Cost follows the number of requests and the length of the text, so a feature that looks cheap in a trial can become expensive at real volume. We estimate cost against expected volume before building, and put an operational cap in place so the bill doesn't surprise you.

Rarely, and promising that usually ends in disappointment. What happens in practice is that it removes the boring part of a task and leaves the judgement to a person — the draft is written automatically and reviewed, the sorting is automatic and the decision manual. Anyone building on a promise of full replacement discovers the gap at the first unusual case.