How to use ai in customer support?
Using AI in customer support works when you treat it as a new junior team member: give it clear material, check its work at first, and widen its responsibility as it proves reliable. Businesses that skip the checking stage tend to produce one memorable wrong answer. This rollout is designed for small teams with limited time.
Step 1: find the repeat questions
Export the last 200 to 500 conversations or emails. Tag them by topic. Most small businesses find that a handful of topics, such as opening hours, delivery times, returns, pricing and booking, make up well over half the volume. Those are the AI's first job.
Step 2: write the source of truth
The AI should answer from your own pages, not from general knowledge. Gather the pages, policies and FAQs that state the correct answer. Fix contradictions between them before loading anything; the AI will reproduce them.
Step 3: choose the setup
| Decision | Safe starting point |
|---|---|
| Channels | Web chat and a shared email inbox first; add SMS, voice or Slack later |
| Approval | A person approves every reply for the first two weeks |
| Scope | Only the repeat topics from Step 1; everything else goes to a person |
| Tone | Two example replies in your voice |
| Handoff | A visible "talk to a person" option and a note of what the customer already said |
Step 4: run in review mode
In Elite Support AI, a human can approve or reject any reply before it goes out, and answers come from your own pages. Use that mode at first. Record each reply as correct, needs edit or wrong, and fix the source page when the same error appears twice, rather than only fixing the reply.
Step 5: measure
- Share of replies approved without edits.
- Number of wrong answers per 100.
- Time to first response.
- Share of conversations handed to a person and why.
- Customer follow-ups that repeat the same question, which signal an unclear answer.
Step 6: widen gradually
When a topic reaches a very high approval rate over a couple of weeks, let the AI send those replies directly while still reviewing a sample. Keep high-stakes topics human: refunds, complaints, medical or legal questions, and exceptions.
Pitfalls
- Letting the AI quote prices or policies that are out of date.
- No clear route to a person.
- Ignoring after-hours messages that the AI is now answering.
- Choosing on per-seat or per-resolution pricing without modelling your volume; the cost calculator helps.
What it costs to start
| Plan | Price | AI replies per month | Sites |
|---|---|---|---|
| Free | $0 | 50 | 1 |
| Starter | $29 per month | 1,000 | 1 |
| Growth | $79 per month | 4,000 | 3 |
| Scale | $199 per month | 15,000 | 10 |
A sample two-week schedule
- Days 1 to 2: tag past conversations, pick five topics.
- Days 3 to 4: clean up the pages that answer them, load them, write two sample replies in your voice.
- Days 5 to 10: review mode, every reply approved by a person; log results.
- Days 11 to 14: fix source pages for repeat errors, decide which topics can go out with a sampling review.
At the end, you will have a measured accuracy figure for each topic rather than a general feeling, which is the best basis for deciding how much to let the AI send on its own.
Frequently asked questions
How long does setup take?
For a small site, an afternoon to load pages and a couple of weeks of review before widening scope.
Will AI replace my support person?
It reduces repetitive work. People still handle exceptions, complaints and judgement calls.