Every sales page for website assistants describes the good days. This one is about the bad ones, because a business that knows where the tool breaks can design around it, and a business that does not finds out from a customer.
The failures are not random. After watching enough assistants in real use, the same handful of patterns keep appearing. Almost none of them are fixed by buying a cleverer product. They are fixed by content, instructions and knowing when a person should take over.
Here are the main ai employee limitations, what each looks like in a transcript, and what contains it.
Failure one: the content is thin, stale or contradictory
This is the big one. An assistant that answers from your content can only be as good as that content. If your website says "prices from £450" and a PDF from two years ago says £395, the assistant has two sources that disagree, and whichever it picks, somebody is going to be disappointed.
Thin content produces a quieter failure. The assistant does what it should, says it does not have that information and offers to take details, but it does so for a third of all questions. That is honest, and it is also not much use.
A well-grounded assistant will not fill the gap by inventing something; SpideyChat, for example, says it does not know and records the question as a knowledge gap. The containment is on your side: remove old documents rather than piling new ones on top, and treat the gap list as a to-do list. The AI FAQ generator is a quick way to draft answers to the gaps you find, which you then check and correct. Training an AI agent on a messy website covers cleaning up at the source.
Failure two: questions that could mean two things
People type in shorthand. "How much for a bathroom?" could mean a full refit, a tiled shower, a single tap replacement or a clean. A person on the phone would ask. An assistant that answers the most likely meaning will be confidently wrong some of the time.
The design fix is to tell the assistant, in its instructions, to ask a short clarifying question whenever a request could reasonably mean more than one thing. It sounds like this:
Visitor: how much for a bathroom
Assistant: Happy to help with that. Are you thinking of a full bathroom refit, or a smaller job like replacing a shower or suite? The pricing works quite differently for each.
Visitor: full refit, small bathroom
Assistant: Full refits are quoted after a survey because of the range involved. The website gives a typical starting point for a small room, and a survey is free. Would you like me to take your postcode and arrange for someone to get in touch?
One extra message, and the answer lands.
Failure three: emotional conversations
Someone whose pet has just died writes to the vet. A customer furious about a botched job writes at midnight. A person worried about a relative in care wants to know what to do.
An assistant can respond politely to all of these, and it will. The failure is subtler: a calm, well-structured, information-rich reply to someone who is distressed can read as cold, and a policy explanation to someone who is angry reads as arguing. Software also cannot judge when a situation is genuinely urgent in the way a trained person can.
Contain it with instructions and routing:
- Acknowledge, briefly. One sentence that recognises the situation, without gushing.
- Do not defend. No policy quotes to an angry customer in the first reply.
- Get to a person. Offer a handover straight away, or take details for a callback if nobody is on duty.
- Point to emergency services for anything involving safety or health. The assistant gives information, not advice, and urgent or clinical matters belong with a qualified person or the emergency number.
Failure four: questions about one customer's account
"Has my payment gone through?" "When is my engineer arriving?" "What did you quote me last March?"
A website assistant answers from the content you give it. It cannot see your job system, your accounts package or last year's quotes, so it cannot answer these, and nor should it pretend to. This is not a bug to report. It is a boundary to design for, usually by opening a numbered support ticket or handing to a person who can look it up.
The mistake to avoid is letting the assistant answer the general version of a personal question. "When is my engineer arriving?" answered with "our engineers usually arrive within a two-hour window" sounds helpful and tells the customer nothing about their own visit. Instruct it to recognise questions about a specific booking, order or invoice and route them straight away.
The failures, and what contains each
| Failure | What you see in transcripts | What contains it |
|---|---|---|
| Thin or stale content | Many unanswered questions, or answers that match an old document | Weekly gap review, removing outdated sources |
| Ambiguous questions | Confident answers to the wrong version of the question | Instruction to clarify before answering |
| Emotional conversations | Polite but cold replies, visitors asking for a person | Short acknowledgement, fast handover |
| Account-specific questions | Visitors frustrated that it cannot check their order | Tickets or handover for anything personal |
| Negotiation and exceptions | Visitors pushing for discounts or special terms | Boundaries in the instructions, pass to a person |
The last row deserves a line. Assistants should not negotiate on your behalf. If someone wants a discount, an exception to the cancellation policy or a rush job, the assistant's job is to capture it clearly and pass it on.
Design for the bad days before launch
Take the five rows in that table and test each one yourself before your assistant goes live on your main pages. Ask a question your content does not cover. Ask something vague. Write a message as an angry customer. Ask about a specific order. Push for a discount. Read each reply as if you were the visitor.
Anything that fails, fix in the instructions or the content, then test again. After launch, the same five checks slot into a regular review; a weekly review routine for supervising your AI employee builds them in. And for the moments where the right answer is a person, how a good AI employee hands over to a human covers doing that cleanly.