Your AI Is Only as Good as Its Answer File

Most AI assistants that disappoint their owners aren't badly built. They're badly fed. The model can write a clean sentence and hold a polite conversation, but it doesn't know your price floor, your service radius, which brands of equipment you refuse to install, or that you stopped doing weekend appointments in March. So it guesses, and a guess delivered confidently to a customer is worse than no answer at all. The fix isn't a better model. It's an answer file: a single, maintained source of truth about how your business actually operates.

What actually goes in the answer file

Think of it as the document you'd hand a new front-desk hire on day one, except written so a machine can use it. It isn't your website copy. Website copy is written to persuade; an answer file is written to be correct. Most service businesses can get a usable version down in a few pages.

Start with the boring operational facts, because those are what customers ask about and what AI most often invents. Your service area, stated as actual towns or postcodes rather than "the greater metro area." Hours, including the holidays you close. Which services you offer and, just as importantly, which ones you don't — the roofing company that doesn't touch flat commercial roofs needs that written down, or the bot will happily book one.

Then pricing, in whatever form you're honest about. If you quote after a site visit, say so and say what the visit costs. If you have a diagnostic fee, a minimum charge, or a consultation that's credited toward treatment, that belongs here in plain language. Most awkward AI moments come from pricing questions the assistant wasn't given permission to answer, so it either stonewalls or improvises.

Add your policies: deposits, cancellations, rescheduling windows, warranty terms, financing options, insurance or membership handling. Add the qualifying questions a good human asks before booking — for a med spa, that might be pregnancy, medications, or recent sun exposure; for an HVAC company, the age of the system and whether there's currently no heat. And add the disqualifiers, the situations where the right answer is "we're not the right fit for this."

Write it for a machine, not for a brochure

The format matters more than people expect. Short declarative statements beat flowing paragraphs. One fact per line. Explicit rather than implied — "we do not service properties more than 45 minutes from the shop" is usable; "we proudly serve the tri-county area" is not.

Include the phrasing customers actually use, not just your internal terms. Your team says "mini-split"; the customer types "that ductless AC thing on the wall." Your injector says "neuromodulator"; the caller says "Botox." An answer file that only contains trade vocabulary will fail on exactly the questions beginners ask, which is most of them.

Write the negative space too. Tell the assistant what it must never do: quote a firm price on complex work, promise a specific appointment time before checking the calendar, give medical or clinical advice, speculate about whether insurance will cover something, or negotiate. These prohibitions do more for output quality than any amount of tone instruction, because they cut off the failure modes that actually cost you trust.

Finally, put a date on it and name an owner. An undated answer file is a slow-acting problem — six months on, nobody remembers whether the pricing section reflects the current rate card, and the assistant is confidently quoting last year.

What to automate, and what stays a human call

Automate the retrieval and the delivery. Once the file exists, the assistant should pull from it on every conversation, and you should never be copy-pasting facts into a prompt by hand. Automate the reminders too: a recurring task that pings whoever owns the file to review it monthly, and a rule that any price change, new service, or policy update triggers an edit before it goes live anywhere else. Automate the drift check by sampling real transcripts each week and flagging answers that don't trace back to a line in the file.

What stays human is the content itself. Deciding what your cancellation policy should be, where your price floor sits, which jobs you'd rather decline, how much detail to give a nervous first-time patient — those are business decisions, and no model should be making them by inference from your marketing pages. Handing that judgement to an AI is how a business ends up with an assistant that's technically fluent and quietly off-brand.

The escalation rules are also yours to set. Which questions trigger a hand-off to a person, how fast that hand-off happens, and what the customer sees in the meantime. The file can encode the rule; only you can decide what it is.

How to know it's working

Read transcripts. Not all of them, and not forever — a sample each week for the first month, then monthly. You're looking for three things: questions the assistant couldn't answer, answers it gave that weren't in the file, and conversations where a customer asked the same thing twice because the first reply didn't land.

Each of those has an obvious fix. Missing answers become new lines in the file. Invented answers mean the prohibitions need tightening. Repeated questions usually mean the wording is technically right but unclear, which is a rewriting job, not a model problem.

This is unglamorous work, and it's the difference between an AI assistant that deflects customers and one that books them. The model is a commodity — everyone has access to roughly the same ones. What your competitor can't copy is a well-maintained, specific, honest description of how your business actually works. That file is the asset. The assistant is just what reads it out loud.