In any decently run office, a new starter does not get the company card, the alarm code and admin rights to the accounts package on the first morning. Access arrives as trust is earned. Nobody finds this insulting. It is simply how you avoid an expensive week.
Software gets a strangely different treatment. Owners switch on every available feature because it came in the plan, then struggle to work out which one caused the odd behaviour they are seeing.
The better approach to ai employee permissions is the one security people call deny by default. Everything starts off. You switch things on one at a time, each for a stated reason, and you watch what happens before the next change. SpideyChat ships this way already: every action is off until you enable it for a particular assistant.
Why one at a time is the whole point
The benefit is not caution for its own sake. It is diagnosis.
If you enable four actions on Monday and on Wednesday a customer complains that they were asked for their email three times, you have four suspects. If you enabled one action on Monday and another on Thursday, you know exactly where to look. Small firms rarely have time for detective work, so avoiding the need for it is worth a lot.
It also forces you to write down why each action exists. "Because it was there" is not a reason, and actions without a reason tend to be configured badly.
The recommended order
| Order | Action | Risk if misconfigured | Main thing to watch |
|---|---|---|---|
| 1 | Search the knowledge base | Low | Answers drawn from outdated or conflicting content |
| 2 | Capture a lead | Low | Asking for details too early, or missing key fields |
| 3 | Transfer to a human | Medium | Visitors waiting with nobody there to pick up |
| 4 | Open a support ticket | Medium | Tickets with too little detail to act on |
| 5 | Send a webhook | Higher | Wrong or partial data arriving in other systems |
Leave at least a few days between steps, or long enough to see a meaningful number of conversations. On a quiet site that might be a week.
1. Knowledge search
This is what lets the assistant answer from your own content rather than general knowledge, so it comes first. Without it you have a friendly box that cannot tell anyone your prices.
Watch for answers that are correct according to an old document. When you find one, delete or update the source rather than adding a correction on top, or you will have two versions fighting.
2. Lead capture
Most of the commercial value sits here, and the risk is mild: at worst you collect a poor lead or annoy someone by asking too soon.
Read the first twenty captured leads carefully. Could someone on your team act on each one without going back to the visitor? If not, change what the assistant asks before it collects contact details. Chatbot lead capture without killing the conversation covers the timing in detail.
3. Transfer to a human
This one looks harmless and is the easiest to get wrong. The action itself is fine. The failure is organisational: the assistant offers a person, and there is no person.
Only switch it on once you have agreed who is on duty and when. Outside those hours the assistant should say plainly that the team is away and take details instead. A good exchange sounds like this:
Visitor: Can I talk to someone about my order?
Assistant: The team is offline until 8am tomorrow. If you give me your order number and email, I'll make sure the first person in sees it, and they'll reply by email.
Watch the time between a transfer request and a person actually joining. If it is regularly more than a few minutes, narrow the hours the option is offered rather than letting people wait. How a good AI employee hands over to a human goes further into the mechanics.
4. Support tickets
Tickets suit businesses with existing customers who report problems. Each one gets a number, which the customer can quote later, and it lands in an inbox your team works through. The ticket inbox sits on the Pro plan, which matters if you are currently on Starter.
The thing to watch is thin tickets. "It's not working" with an email address is not a ticket anyone can resolve. Add instructions so the assistant asks which product, what happened and when, before it opens one.
5. Webhooks
A webhook sends a JSON POST to another system whenever something happens, which can connect to Zapier, Make, a CRM or a Slack channel. It is powerful and it goes last, because errors no longer stay inside the assistant. A missing field becomes a broken CRM record, or a noisy Slack channel that the team learns to ignore.
Test it with your own conversations first. Fill in a lead as a fake visitor, then check that every field arrived where it should. The integrations page shows the common destinations, and how chatbots feed your CRM with clean lead data covers field mapping.
When to switch something back off
Turning an action off is not a failure. It is the same as taking someone off the phones while they learn the new price list.
- Switch lead capture off if it is collecting large volumes of junk and your team has stopped reading the list. Fix the questions first, then return it.
- Switch transfer off during holidays or staff shortages, rather than letting visitors click for a person who never arrives.
- Pause a webhook if the receiving system changes, until you have retested the fields.
- Leave knowledge search on almost always. If answers are bad, the problem is content, and turning search off just makes it worse.
Your next step
Open the settings for your assistant and write a one-line reason next to every action that is currently on. Any action you cannot justify in a sentence, switch off today. Then pick the next action you actually need, enable only that one, and put a note in your calendar to read the resulting conversations in three or four days. If you are still deciding which actions a role needs at all, start with writing a job description for your AI employee.