Count the questions your business answered last week. Now count how many you’d already answered a hundred times before.

“Do you service my area?” “What’s your return policy?” “How long does an install take?” “Are you open Saturday?” Every one of those took a person away from something harder — and every one that arrived at 8pm waited until morning, where some percentage of the askers quietly went to a competitor who answered first.

You already wrote the answers. They’re in your FAQ page, your policy docs, your onboarding emails, the SOP binder nobody opens. The move is to turn that existing content into an assistant that answers instantly, around the clock — without letting it improvise.

The one rule that makes this safe

The assistant only answers from your documents. This approach (the industry calls it retrieval-based, or RAG) means the software first finds the relevant passage in your content, then phrases an answer from it. If your documents don’t contain the answer, it says so and hands off to a human.

That single constraint is the difference between an assistant that says “We service ZIP codes 30030–30060, and Saturday visits book out about a week ahead” and one that cheerfully invents a discount you’ve never offered. A general-purpose chatbot pointed at the open internet will embarrass you. One fenced to your own content mostly can’t.

What to feed it

Start with what exists:

  • Your website FAQ and service pages
  • Policy documents (returns, warranties, scheduling, payment terms)
  • The five emails you find yourself rewriting every week
  • Onboarding or “what to expect” documents
  • Price sheets — if you’re comfortable quoting them publicly

Then spend one honest afternoon on the gap: write down the twenty questions your team actually gets, and check each one has a clear written answer. In most businesses this exercise finds five to ten questions everyone answers slightly differently. Fixing that inconsistency is valuable even before any software is involved.

The build, step by step

A veterinary clinic, an HVAC company, and a software consultancy would all follow the same path:

  1. Collect the documents above into one folder. Delete anything stale — the assistant will faithfully repeat your 2023 prices if you leave them in.
  2. Pick a platform that offers “chat over your documents” with your own branding. This is a crowded, commoditized category now; the capability matters more than the logo.
  3. Set the boundaries. Configure the assistant to: answer only from the provided content; say “I don’t know — let me get you to a person” when unsure; and always offer a path to a human (email, callback form, booking link).
  4. Test it like a skeptic. Before launch, throw fifty real questions at it — including ones it shouldn’t answer (“can you knock 20% off?”, “what do you think of [competitor]?”). Tune until the failures are graceful.
  5. Launch quietly. Put it on your website first. Watch the transcripts for two weeks. Then consider text/SMS if the volume justifies it.

Keep a human in the loop

“Deflect 30–40% of repetitive questions” is a realistic settled-state target — but the goal is not zero human contact. Design the handoff:

  • Every conversation offers a talk-to-a-person path, prominently.
  • Questions the assistant couldn’t answer get logged and reviewed weekly. Each one is either a missing document (write it) or a signal that a human should own that topic (route it).
  • Anything involving a complaint, a refund, or an upset customer routes straight to a person. Instantly. No assistant is worth an angry customer arguing with software.

The transcripts, by the way, become the best market research you’ve never had: an unfiltered log of what customers actually want to know, in their own words.

The cost, plainly

  • Software: typically $30–$150/month at small-business scale, usage-dependent.
  • Setup: the content collection and testing is a day or two of honest work. The software part is hours.
  • Maintenance: thirty minutes a week reviewing transcripts and patching content gaps. Skip this and quality decays; do it and the assistant gets steadily better.

Payback math: if it handles thirty questions a week that used to take four minutes each, that’s two hours a week back — before counting the after-hours leads that no longer go cold.

Pitfalls we see

  • Feeding it everything. Internal docs with margins, staff notes, or venting have no business in a customer-facing tool. Curate.
  • No owner. Someone specific reviews transcripts weekly, or nobody does.
  • Pretending it’s human. Label it as an assistant. Customers don’t mind bots; they mind being tricked.
  • Letting it negotiate. Prices, discounts, exceptions — human-only. Configure it to hand off, not haggle.

We build assistants like this — it’s the “intelligent chatbot” piece of our AI Digital Assistant service, and the same careful, fenced approach runs the chat on this very site. If you want a straight answer on whether your content is ready and what it would take, book the free assessment. Twenty minutes, no pressure.