Skip to content

Explainer · August 27, 2026

AI for home services: what actually runs the front office

Chatbots deflect. Agents do work. Here's the difference in plain terms, and the questions that expose which one you're being sold.

Every tool in your inbox now says AI. Underneath the label there are really three different things, and the difference decides whether the 9pm caller gets booked or gets deflected.

The three things sold as AI

  • A chatbot answers questions from a script or a website. Useful for 'what are your hours' — helpless the moment the answer requires doing something.
  • An assistant helps a human mid-task: drafting the reply, summarizing the call. The human still owns every job start to finish.
  • An agent owns a job: it answers the call, checks the real calendar, books the real slot in your field software, sends the confirmation, and logs all of it. Work, not conversation about work.

What separates a real agent from a demo

  • Tools: can it actually write into ServiceTitan, Housecall Pro, or Jobber — or does it 'integrate' by emailing you a note?
  • Memory: does the repeat caller get recognized, or interrogated from scratch?
  • Truth constraints: is 'never invents a price' enforced by the system, or promised in marketing? Ask how it's enforced. Enjoy the pause.
  • An approval model: what waits for your yes — spend, refunds, change orders — and is that gate a mechanism or a policy document?
  • A record: can you read exactly what it did and why, per action, after the fact?

Why the approval model is the whole ballgame

The scary version of AI-runs-my-business and the useless version fail the same way: no governance. Full autonomy with no gate is a liability with your name on the truck. A gate on everything is a chatbot with extra steps. The working answer is graduated: routine and reversible runs on rules you approved; risky and material waits for your recorded yes in one inbox; autonomy gets earned per workflow by a track record, not assumed on day one. That's the model Caliber OS is built around — five named seats doing the work, one owner holding every yes that matters.

And one more honesty test, since you're evaluating vendors who publish articles exactly like this one: ask what they can't do yet. A vendor with no answer to that question has an answer to no question.