Monday at a small kitchen fitting company. Before 9am the inbox has a question about lead times, two requests for a rough price, someone asking if they cover a nearby town, a customer whose drawer runner has failed, and a supplier chasing an invoice. By 10am there is a complaint about a delayed worktop and a request for a discount because a neighbour got one.
Some of that list could be handled by software without anyone noticing the difference. Some of it absolutely could not, and getting that wrong would cost the business a customer or a chunk of margin.
That is the useful way to think about an ai employee for small business: not as a whole person, but as a set of tasks that either transfer cleanly, do not transfer at all, or need rules before they do.
The honest inventory
Here is the split, area by area, for a website assistant working a small business front desk. The left column is work that transfers well. The right column is work that should stay with a person, even if the software could technically produce a reply.
| Area | Can do well | Stays human |
|---|---|---|
| General questions | Answer hours, areas, services, process and policies from your content, at any hour | Explain why a policy exists to an unhappy customer |
| Prices | Quote published prices or ranges and explain what drives cost | Final quotes, discounts, price matching, "can you do better" |
| New enquiries | Ask qualifying questions, capture name, contact and job details | Decide whether a job is worth taking |
| Appointments | Capture a booking request with preferred times | Confirm the slot, juggle the diary, handle double bookings |
| Existing customer problems | Acknowledge, open a numbered ticket, gather photos and details | Diagnose the fault, decide on a revisit or refund |
| Complaints | Apologise, log it, set a clear follow-up time | Resolve it, offer goodwill, own the relationship |
| Handover | Transfer a live conversation to whoever is on duty | Everything that happens after |
| Other systems | Send lead or ticket details to a CRM or Slack via webhook | Decide what the data means |
| Unknown questions | Say it does not know, capture the question, record the gap | Write the missing answer |
| Other languages | Reply in the language the visitor writes in | Review anything contractual in that language |
Read down the left column and a pattern appears: every item is either retrieving something you already wrote or recording something for later. Read down the right and every item involves a decision, authority or accountability.
The grey zone, where setups go wrong
Most disappointments with assistants do not come from obvious cases. Nobody expects software to negotiate a discount. They come from tasks that sit between the columns and were never given a rule.
- Availability. "Can you come Thursday?" sounds factual. It is not, unless the assistant can see the diary. Without that, the rule should be: capture the request, say it is not confirmed.
- Refunds and returns. Explaining the policy is fine. Telling a customer they qualify is a decision. Write the line explicitly in your content.
- Urgent requests. "My oven is sparking" is not a ticket for the morning. Give the assistant an emergency route to show before anything else.
- Opinions. "Which worktop is best for a family kitchen?" can be answered from your guides, but "which should I choose?" invites a recommendation you may not want given without seeing the room.
- Existing orders. "Where is my worktop?" needs order data. If the assistant does not have it, it should log the request, not guess.
The fix is the same each time: decide the rule, write it where the assistant will find it, and test it with the awkward phrasing. When should a chatbot hand over to a human lays out clean triggers for the handover end of this.
A worked week, with numbers to swap
Take that kitchen company again, with invented figures. In one week it receives 70 website and email enquiries.
Sorting them by the table above:
- 31 are general questions or published price questions. These transfer fully.
- 18 are new job enquiries. The detail-gathering transfers; the decision and the quote stay human.
- 9 are booking or visit requests. Capture transfers; confirmation stays human.
- 7 are problems or complaints from existing customers. Logging transfers; resolution stays human.
- 5 are suppliers, spam or other. Mostly ignored.
If a person spends an average of 5 minutes on each of the 31 fully transferable enquiries, that is about 2.5 hours a week. The partial ones save less each, perhaps 2 minutes of back-and-forth per enquiry on the remaining 34 relevant ones, which is roughly another hour. Call it three and a half hours a week handed to software, and none of the decisions.
That is a useful amount of time for a small team. It is not a salary. Anyone promising the second should be able to show their working. Run your own version through the chatbot ROI calculator.
What it cannot do, whatever the settings
A few limits are structural, not configuration choices.
A website assistant does not answer the phone. SpideyChat, for example, works on your website only and has no phone number. It also does not write appointments into a calendar; it captures the request with the details attached for a person to confirm. It cannot know anything you have not given it, so an out-of-date price list produces out-of-date answers. It cannot notice that a regular customer sounds unlike themselves today. And it cannot be accountable: when something goes wrong, a person has to own it.
None of that makes it less useful for what it does. It just defines the job. What an AI employee can actually do gives more examples on the practical side.
Run your own inventory this week
- Export a week of enquiries. Emails, form submissions, chat logs and messages. Anything a customer typed.
- Tag each one. Can, cannot or grey, using the table above as your guide.
- Write rules for every grey item. One or two sentences each: what the assistant says and what it passes on.
- Estimate the time. Minutes per transferable enquiry, multiplied out.
- Decide from the number. If the can column is small, a better FAQ page may be all you need. If it is large, write the job down properly using an AI employee job description and test it in the demo with your trickiest grey-zone questions first.