AI that does
a job.

• Intake and triage. An enquiry arrives as a paragraph of free text. Something has to decide what trade it is, how urgent it is, whether it's even in your service area, and what to ask next. That's a genuinely good use of a model.
• Drafting, not sending. The reply is written for you in seconds. You read it and hit send. The two-second review is what keeps the quality yours.
• Retrieval over your own material. Answers grounded in your pricing, your policies and your past jobs, rather than a model's best guess about how businesses like yours usually work.
• Structured extraction. Turning unstructured mess, emails, transcripts, PDFs, handwriting, into rows that a system can actually use.

A human in the loop wherever the cost of being wrong is real. Drafting a reply: fine to automate. Sending a quote, cancelling an appointment, telling a customer something binding: that gets your eyes on it first. This isn't caution for its own sake: it's that an automated system confidently telling a customer the wrong thing costs more than it ever saved.

Deterministic where it can be. Most of what gets sold as "AI" is a database query and three if-statements. If your problem has a correct answer, it should be computed, not generated. I'll use a model for the genuinely fuzzy part and ordinary code for everything else, which is cheaper, faster, and doesn't invent things.

Failure has to be visible. When a model returns something unusable, the system needs to notice and escalate rather than pass nonsense downstream and hope.