Actions are the difference between an agent that talks and one that finishes the job. They are also the part where "turn everything on and see" is a bad instinct.
Start with one
Pick the action that matches why you installed the thing.
- Losing enquiries out of hours → capture leads
- Drowning in repeat support questions → open tickets
- People asking for a person and not getting one → transfer to a human
Turn on one. Watch it for a week. Then add the next.
Enabling four at once means that when something works — or doesn't — you cannot tell which one did it.
Before you switch on lead capture
Check where the lead goes. An inbox somebody reads, or a webhook into a system somebody looks at.
The most common failure here is not technical. It is that leads were captured beautifully for three weeks into an address nobody had open.
Before you switch on tickets
Make sure someone owns the queue. A ticket with a number is a promise to the customer, and an unanswered one is worse than never having offered.
Before you switch on human handover
Two questions: who is on duty, and what happens when nobody is?
Offering "talk to a person" when there is no person is the fastest way to turn a mild query into a complaint. A good setup either hides the option when nobody is available, or says plainly when someone will reply.
What to leave off
Anything destructive. Cancelling, refunding, deleting.
Not because the technology cannot do it, but because the cost is asymmetric: a booking captured wrongly is an annoyance, and a booking cancelled wrongly is a lost customer and an argument. Those need a person to confirm.
A reasonable progression
Week 1: answers only. Learn what people ask. Week 2: add lead capture, pointed somewhere you check daily. Week 3: add human handover if you have someone to hand over to. Week 4: add tickets if support volume justifies a queue.
Four weeks sounds slow. It is considerably faster than switching everything on, having something go wrong, and turning it all off again.
The question to ask before each one
For any action you are considering, ask: what happens if it fires when it should not have?
- Lead recorded that was not really a lead → you read one extra email. Fine.
- Ticket opened unnecessarily → someone closes it. Fine.
- Human fetched when nobody needed one → mildly annoying. Survivable.
- Booking cancelled in error → a lost customer and an argument. Not fine.
The pattern is obvious once you write it out, and it is the whole basis for which actions belong on an agent's own initiative and which need a person in the loop.
Review it after a month
Look at what each enabled action actually did. Most products log every attempt, including the ones that were refused or failed.
If an action has fired zero times in a month, either the agent has no reason to use it or the description it was given is unclear. Both are worth knowing. Turning it off is a perfectly good answer.
Related: what an AI employee can actually do and when to hand over to a human.