When “AI agent” is mentioned, the first question in most managers’ minds is the same: Will this thing get into my systems and break something? The question is fair. Unlike an ordinary chat tool, an agent takes steps on its own: it reads data, compares, produces a result. If you let it write that result somewhere, it writes its mistakes too.
In this article we describe a way of working that keeps this risk to a minimum. We call it an “approval-gate agent”. It has three rules: read-only access, human approval, records. We set up this arrangement for our customers; in our own work we also operate with the principle of approval and record.
First rule: the agent reads, it does not write
The agent we set up for our customers only reads your systems. It does not create, change or delete records. It connects to your accounting program, your customer relationship software and your shared folders with a read-only account. This single rule closes the biggest concern from the start: however wrong the agent may be, it cannot touch your data.
So how does the work get finished? The agent produces an output: a list of differences, a summary, a reminder. The person who writes the change into the system is still your employee. The agent says “these three orders look different in the two lists”; the correction is made by a human.
Second rule: everything that goes outside passes through human approval
An e-mail, a report or a suggested record update produced by the agent comes before one of your employees before it goes outside. The employee looks, corrects if needed, approves. At the start, nothing is sent without approval.
This gate does not rely on the assumption that the agent will make no mistakes; on the contrary, it accepts that it can. A wrongly flagged difference is caught on the approval screen, and the rule is corrected accordingly. Over time, routine tasks can be tied to pre-approved rules; any message that contains a price, a commitment or an attachment always goes with human approval.
Third rule: every step is recorded
What the agent read, what it suggested and who approved what are recorded in an auditable way. When a question comes up, the answer to “why did the agent suggest this?” is in the record. This record is also used for measurement: the work time before and after, the rate of correctly flagged differences, the false alarm rate and the approval time are reported regularly. We talk about whether the agent is useful not by guessing but with these numbers.
An example: comparing two lists
Think of a distribution company. Open order reports come from dealers every week; some in Excel, some in PDF. The same orders also sit in the company’s own customer relationship software. An employee compares the two sides by hand every week; every difference that gets missed means a missing delivery or excess stock.
The approval-gate agent works on this job as follows: it reads the two sources, flags the differences and presents the list to the employee. It does not write to the software. The employee looks at the list, approves the correct ones and makes the correction themselves. The comparison that took hours each week comes down to reviewing an approval list.
The same arrangement suits other recurring tasks: matching supplier invoices with order records, a weekly summary of alerts from the monitoring system, collecting contract renewal dates from several lists and sending reminders.
Which tasks fit and which do not
A task that fits is one whose rule can be put into writing: “these fields in these two lists must match”, “notify this person before this date”. If it repeats weekly or daily and the data sits in more than one place, it is a good candidate for an agent.
There are tasks that do not fit as well. In tasks where the rule changes every time and cannot be put into writing, the agent goes in the wrong direction. A setup that asks the agent to write to your system or change records is also outside the arrangement we recommend; we keep the agent read-only and leave the change to a human.
How we set it up
We proceed in four stages. In discovery we write the scope document and the success criteria together: which data, which comparison, which output. In the pilot, the agent works with exported data and with human approval; we give a weekly report. In production we go live with a read-only service account, monitoring and an audit record. In operation, monthly maintenance and reporting continue.
What we need from you: in discovery, read-only access to the data sources and an employee who can explain the rule; in the pilot, a person to approve the results.
We work this way too
Before recommending this arrangement to our customers, we set it up in our own work. An agent performs the monitoring and backup checks at regular intervals, summarizes the findings and presents them to the responsible engineer; the engineer makes the decision to intervene. An agent prepares the daily work summaries and the preliminary check of incoming invoices, and the employee concerned checks the results. For quote drafts, an agent gathers the relevant information from our product catalog; an authorized employee approves the price and the terms.
Our agents do not make payments; invoicing, deletion and every write operation happen only after a person’s approval, and every step is logged. Every outside message that contains a price, a commitment or an attachment passes through a person’s approval; routine reminders go out under pre-approved rules. Every write and every message is recorded. Even where we give write permission, the approval gate and the record rule operate exactly the same.
Also on this topic: A custom AI agent for you · We work this way too
Is there a recurring check in your work? Let us look together, in a short call, at whether it suits an agent.
