Why our AI agents never move money on their own

Most conversations about AI agents start with autonomy: how much can the agent do by itself? We start from the other end. Before an agent touches a real workflow, we decide what it may never do alone.

The rule is short. An agent can read, sort, draft, prepare and recommend. A person approves anything that moves money, carries a signature or reaches a customer. And no agent can ever approve its own work.

It sounds conservative. In practice it is what lets organizations put agents into production in weeks instead of debating them for quarters.

Where the rule comes from

We did not design it on a whiteboard. It comes from running agents on our own group of companies every day.

The administration and finance of the group used to run through one overloaded inbox: invoices, contracts, payment runs and investor reporting across several companies and countries. Today agents sort the mail, prepare the reports, draft the paperwork and line up the payments. A person approves every payment and every signature.

That split was not a compromise. The expensive part of back office work is rarely the click that releases a payment. It is everything before it: finding the invoice, matching it to the contract, checking the amount, preparing the batch. Agents take that work. The approval itself takes seconds, and it is exactly the moment where a human judgment is worth having.

What the rule looks like in four different workflows

The same pattern holds across very different jobs.

Workflow	What the agent does	What a person approves
Group back office	Sorts mail, prepares reports, drafts documents, lines up payments	Every payment and every signature
Synthetic developer	Takes tickets, asks questions, writes and tests code	Every change before it ships, and the agent cannot approve its own work
Fundraising pipeline	Maps investors and introductions, drafts every message	Every message before it goes out
Company memory	Reads meetings, chat and tickets, writes the shared record	Changes before they land in the record

In each case the agent does the volume and the person keeps the decision.

Why approval makes adoption faster

Three things change when the rule is explicit from day one.

Trust arrives earlier. Teams accept an agent much faster when they know it cannot surprise them. The question stops being "do we trust AI?" and becomes "do we trust this draft?", which people answer every day for each other.

Mistakes become cheap. An agent that prepares a wrong payment costs a person a few seconds to reject it. An agent that sends a wrong payment costs a phone call to the bank, or worse. The rule turns the second kind of mistake into the first.

Learning becomes visible. Every rejection and every correction is a signal. Because the agent works through approvals, you can see exactly where it needs better instructions, better data or a narrower scope. Without approvals, the same errors would be invisible until something breaks.

How we set it up

A few practical details matter more than the model you choose.

1. Give the agent its own identity. Our synthetic developer has its own email and accounts, like any new hire. That way every action is attributable, and permissions can be granted and revoked like a person's.
2. Make the approval step part of the tool people already use. Approvals happen in the inbox, the team chat or the task board, not in a separate dashboard nobody opens.
3. Separate the maker from the checker. An agent never approves its own output, and where several agents work together, one cannot sign off another's work in place of a person.
4. Set a budget. Agents run inside a fixed spending limit, so cost is known in advance and a runaway loop cannot become an expensive one.
5. Log everything. Every action and every approval is recorded. A year later you can still see what happened and why.

When to loosen the rule

Approval is not forever. Once a workflow has run for a while and the rejection rate on a given task is near zero, it can make sense to let the agent complete that narrow task alone, with sampling and logs instead of a sign off on each item. The point is that this is a decision you make with evidence, task by task, rather than a leap of faith at the start.

What should almost never be automated away is the approval of money, signatures and anything a customer will read in your name.

Where to start

Pick one workflow with real volume and a clear owner. Write down what the agent may prepare and what a person must approve. Then build the agent into the tools the team already uses and measure what gets corrected.

If you have a workflow like that in mind, we are happy to look at it with you.