A new person joining your team gets two weeks of patience. Someone shows them where things are, they sit in on calls, they are allowed to ask daft questions and nobody expects them to handle the awkward customer on day one.
A new website assistant usually gets two hours. The owner points it at the website, asks it three questions, is pleasantly surprised, and puts it live on every page. Then the odd answers start arriving and confidence drops fast.
Onboarding an AI employee goes much better if you borrow the new-hire routine. The fortnight below assumes a small business with one person, often the owner, giving it fifteen or twenty minutes on most days.
Days 1 and 2: the induction pack
A new starter gets a folder. Your assistant needs the same thing, and the website alone is rarely enough, because websites are written to sell rather than to answer.
Gather these before you teach it anything:
- The pages that matter: services, prices, areas covered, FAQs, terms.
- The documents your team actually uses to answer customers: price lists, brochures, the policy PDF, the email you always paste when someone asks about cancellations.
- A plain-language description of the business, written as if briefing a temp: who you serve, what you do not do, what a good enquiry looks like.
- Twenty to thirty real questions copied from your inbox, contact form and messages over the last few months.
Keep item 4 separate. It is your exam paper, not teaching material. If you feed those questions in as Q&A pairs, the assistant will pass the test and you will learn nothing about how it copes with the questions you did not think of.
Then teach it. SpideyChat accepts all four approaches: a written description, a crawl of your site of up to 150 pages, uploaded documents (PDF, DOCX, TXT, CSV or Markdown) and individual Q&A pairs. If your site is untidy, training an AI agent on a messy website is worth reading before you crawl it.
Days 3 to 5: shadowing, the other way round
With a person, shadowing means they watch you work. With an assistant it runs in reverse: you watch it work, privately, before customers do.
Do not add the script tag to your live site yet. Instead, ask it every question from your exam paper, word for word as customers wrote them, including the spelling mistakes. Mark each answer as right, wrong, or "right but I would not say it like that".
The wrong answers nearly always fall into one of a few causes. The information was never supplied. It was supplied in two places that disagree (an old price in a PDF and a new one on the site is the classic). Or the question was vague and the assistant guessed at the meaning. Fix the source, not the symptom, then ask again.
A well-built assistant should respond to a missing answer by saying it does not have that information and offering to take details. If yours invents something instead, stop and find out why before going further.
Days 6 to 9: a first shift with a narrow brief
Now it goes live, but on a short lead. Pick one or two pages where enquiries start, often the contact page and your main service page, and keep every action switched off except knowledge search and lead capture. Our post on starting with no permissions explains the reasoning for each one.
This is the stage where daily transcript reading earns its keep. Here is a rough worked example you can adjust.
Say the assistant holds 12 conversations a day on those pages. You read every one on days 6 and 7, which is about 24 conversations and perhaps 25 minutes each day. From day 8 you read only the ones with a poor rating, a handover request or an unanswered question, plus three picked at random. That drops you to around 10 minutes. If your volume is three times higher, sample more heavily rather than reading everything.
What you are looking for:
- Questions nobody predicted. Add the answer to your content the same day.
- Answers that are accurate but clumsy. Adjust the persona or add a Q&A pair in your own wording.
- Enquiries captured without the detail you need. Tighten which questions it asks before taking contact details.
- Visitors who wanted a person. Note the time of day; this decides your on-duty hours later.
Days 10 to 14: widening the job
By now the unanswered questions list should be shrinking and the transcripts should feel boring. Boring is the signal to give it more responsibility, one change at a time.
| Day | Change | What to check the next day |
|---|---|---|
| 10 | Add it to the rest of the service pages | Are new page types producing new question types? |
| 11 | Switch on transfer to a human, with someone on duty | Did takeovers happen quickly, and did the person have context? |
| 12 | Turn on a proactive greeting on one page | Are more conversations starting, or just more "hi" and nothing? |
| 13 | Switch on support tickets, if you use them | Are tickets numbered, complete and landing with the right person? |
| 14 | Connect a webhook to your CRM or Slack | Did every lead arrive with all its fields? |
Only move to the next row when the previous one looks clean. If day 11 turns up a problem, day 12 waits. Stacking two changes on one day means that when something breaks, you will not know which one did it.
What day fifteen looks like
At the end of the fortnight you should have an assistant running on your enquiry pages with a known set of actions, an unanswered questions list that is shorter than it was, a clear idea of when visitors want a human, and a content set you trust.
From here, daily reading becomes a weekly half hour. Book it now as a recurring slot, because it is the first thing to slip once the assistant seems fine. A weekly review routine for supervising your AI employee lays out how to spend those thirty minutes.
If you have not started yet, copy your thirty real questions into a document today. That exam paper is the most useful thing you will make, and you can put it to any assistant, including the one on /try, before you commit to anything.