FlowsBook a callSign inCreate free accountSign up
AI employees and automation

How much autonomy to give an AI agent, decided one action at a time

How much autonomy to give an AI agent: split its work into looking, deciding and acting, set limits on each, and widen them only as its track record grows.

Part of the guide AI employees

By the xMap team · · 5 min read

The practical answer to how much autonomy to give an ai agent is: plenty for looking things up, some for deciding, and very little for acting, at first. Split what the agent does into those three kinds of step, give it free rein on reading data, put rules and a second check around its decisions, and keep a person in front of anything it sends or changes outside your company until its record earns more. Autonomy is not one dial for the whole agent; it is a setting for each action.

This article is for teams putting an AI agent to work with real tools: searching places, reading their own tables, drafting messages, calling other systems. It explains the three kinds of step, the questions that decide each limit, a made-up example from an estate agency, and the specific controls an agent has in xMap Flows.

Looking, deciding, acting

Kind of stepExamplesRisk if it goes wrongSensible starting autonomy
LookingSearch places near an address, read your client table, check the forecastA wasted searchHigh: let it look as it sees fit, within a budget
DecidingWhich listing fits a buyer, whether an enquiry is a complaint, how to score a leadA wrong choice someone acts onMedium: written rules plus a check
ActingEmail a client, post in a channel, add a row to your CRM, book a visitSomething said or done in your nameLow: a person approves, at least at first

Sorting an agent's steps. Most of the risk sits in the last row.

Looking is cheap and reversible: a bad search can be thrown away. Acting is the opposite: an email to a client cannot be unsent. That is why it rarely makes sense to give an agent one level of trust for everything.

Four questions for each action

  • Can it be undone? A draft can be edited; a sent message cannot. The less reversible, the more checking.
  • Who sees the result? Your own team forgives a rough line. A customer or the public does not.
  • How often does it happen? Ten times a day you will stop reading approvals carefully. Once a week you will read every word.
  • How clear is the rule? "Only listings under the buyer's budget" can be checked by a machine. "Sound warm but professional" needs a person's eye for a while.

A worked example

This example is made up. An estate agency in Melbourne gives an AI agent the job of answering buyer enquiries that arrive through a form on its website. The agent can look up the agency's listings table, places near each property (schools, cafes, stations), and draft a reply.

  1. Looking: high autonomy. It may search the listings table and nearby places as often as it needs for each enquiry, up to a set number of tool calls, inside the AI employee's monthly budget.
  2. Deciding: rules plus a check. It picks up to three listings that match the buyer's budget and suburb. A second step checks every draft: it names only listings in the table, quotes no numbers not in the data, and stays under 150 words. Drafts that fail go back to be rewritten.
  3. Acting: a person first. Every reply waits for someone on the sales team to approve it, and they can rewrite it on the approval page. After six weeks they add a Branch: replies to simple viewing requests go out on their own, while anything mentioning an offer still waits for a person.

The agency widened autonomy one step at a time, on the evidence of what the agents changed (or did not change) in the drafts.

Mistakes to avoid

  • Full autonomy because the demo looked good. A demo has easy cases. Give it a few weeks of real ones first.
  • Approval fatigue. If every item needs approval and people approve in bulk without reading, the check is gone. Narrow what needs approval instead.
  • Letting it choose every setting of a tool. Decide which settings it fills on each call (the place to search around, say) and fix the rest yourself (the radius, the number of results).
  • No limit on tool calls. An agent that searches again and again on a hard question costs more and answers later. Cap the calls per message.
  • Never loosening. Strict forever wastes people's time. Review the approvals every few weeks.

The controls an agent has in xMap Flows

In xMap Flows an agent is the AI agent block inside an AI employee, and each of the three kinds of step has its own controls.

Looking

The agent can only use the blocks you link to its Tools dot, such as Nearby places or My data. In each tool you write {{agent}} in the settings it may fill on each call; every other setting stays as you set it. Most tool calls caps how many it makes for one message, and it stops calling paid tools when the run's credits would not cover more.

Deciding

Its instructions set the rules in plain words. Check the work after it checks each answer against your rules and sends failures back to be redone. On a chat page, Protection settings keep it on topic and take personal data out of answers. Test cases in the Quality panel let you check answers against expected ones before and after a change.

Acting

A sending block linked as a tool is never run while the agent thinks: in a test run it only says what it would have sent, and in a real run the sends are queued and made with the run's other sends. On top of that, each AI employee has one of three levels: Ask me first puts an Approval before what it sends, each item on its own with drafts you can rewrite; Act and tell me sends and leaves a note; Fully on its own sends and stays quiet unless something breaks. Conversations answered on the spot (a chat page, a WhatsApp reply) are not held, so put the checks before the reply instead.

Templates to try these controls on: Chat about places near anyone and An intro and two follow-ups to new dental clinics. The AI employees guide covers how agents fit into the wider team.

Questions

Should an AI agent ever act without approval?

Yes, once its record shows the drafts rarely need changing and the action is low risk, such as an internal list or a reply to a simple, repeated question.

Can I see why an agent did something?

Yes. Each run keeps a trace of what it was asked, which tools it called with what, what came back and what it answered.

What stops an agent from spending too much?

A cap on tool calls per message, the run's credits, and the AI employee's monthly budget, which pauses it at its limit.

Can the agent ask a person when it is unsure?

A flow can include Ask a person, which asks someone by email, WhatsApp or Slack and carries on with the answer.

Templates that do this

IndustriesReal estate

Read next

More on ai employees and automation · Every article · Templates

Put it to work

An AI employee in xMap Flows does this on live data, on its own timetable. Start from a template or describe it in plain words.