4.2 · Governance & privacy· 2 min read
Limits, logs and when a person steps in
Every agent works inside written limits, hands over to a person with context when it should, and logs what it saw, decided and did. A named person on your team answers for the result.
Agents get room to act because there are rules around them. Without the rules, you'd have to check everything, and then there's no point.
Every agent has limits
Agreed with you and written down for each agent, before it does any work:
- What it may do on its own. For example: draft a note, update the CRM, create a task.
- What it must never do. For example: change pricing, make commitments, contact a customer someone has flagged as sensitive.
- How often it may contact a customer. So no one hears from your company four times in a day.
- How sure it must be before acting. Below that, it asks.
Limits start tight and widen only when the log shows the agent has earned it. They never go away completely.
When a person steps in
An agent hands over to a person when:
- it isn't sure enough,
- it's one of your key customers,
- the deal is above an agreed size,
- a competitor is named,
- the customer sounds unhappy,
- it's something it hasn't seen before,
- the topic is legal, security or compliance.
It hands over with context: what happened, what it knows, and what it suggests. Not just an alert.
Everything is logged
For every action an agent takes, the log records:
- what it saw at the time,
- what it decided, and why,
- how sure it was,
- what it did,
- and whether a person changed it.
The log is how you answer "why did that happen?" It's also how the loop improves: the cases people handle become the patterns agents learn from. In the Sprint, we read it with you every week.
When two tools disagree
It will happen. The rule for who wins is set in advance:
- A person's judgement.
- What the people who talk to the customer know.
- Scores and predictions.
- Fixed rules.
Then we fix the cause, usually a definition two people were using differently, so it doesn't happen again.
People stay accountable
An agent can do the work. It can't own the result. For every agent, a named person on your team answers for what it does.
People watch for exceptions rather than approving every action. Nobody approving everything, and nobody absent.
Questions this chapter answers
- How do you stop an agent from sending the wrong thing to a customer?
- Every agent has written limits agreed with you: what it may do alone, what it must never do, how often it may contact a customer, and how sure it must be. It hands over to a person when it isn't sure, when it's a key customer, when a competitor is named or the customer sounds unhappy, and every action is logged with the reason.
- Who is accountable when an agent gets something wrong?
- A named person on your team, for every agent. An agent can do the work but can't own the result. The log shows what it saw, what it decided and why, so the cause can be found and fixed.