Writing SOPs for AI agents: what a person never needed

Most procedures in a finance team were written for someone who already sits in the office. They say "check the bank rec looks right", "flag anything unusual" and "use the usual template". A colleague reads those lines and fills the gaps from context: last month's file, the person at the next desk, a feel for the client. An AI agent has none of that. Writing SOPs for AI agents is mostly the work of finding the sentences that quietly rely on a person's judgement, and writing down what that judgement actually is.

What a person brings that the manual never says

Take a hypothetical step from a monthly supplier statement reconciliation: "Compare the statement to the ledger and investigate any differences." A person who has done it twice knows a lot that is not on the page.

• Differences under a few pounds on this supplier are usually rounding and get written off.
• Invoices dated in the last two days of the month often sit in next month's batch, so they are timing, not errors.
• The supplier sometimes sends a credit note by email instead of on the statement, so check the AP inbox before chasing.
• If the difference is over a certain amount, the finance manager wants to know before anyone contacts the supplier.

None of that is written down, because nobody needed it written down. The person picked it up by sitting next to someone, or by getting it wrong once. Unlike an outdated SOP, this manual may be perfectly current. It just assumes a reader who knows the rest.

Writing SOPs for AI agents: five things to make explicit

An agent will follow what is written with patience a person cannot match, and it will not reliably guess what is missing. So the useful question for each step is: what would a capable new starter on their first day have to ask? Write down the answers. In practice they fall into five groups.

1. Inputs, named exactly

"The bank statement" is not an input. "The CSV export from the bank portal for the main current account, covering the first to the last calendar day of the month" is. Say where each file lives, what it is called, and what to do if it is missing or covers the wrong dates.

2. Thresholds instead of adjectives

Words like "material", "unusual", "significant" and "looks right" carry the most judgement and the least information. Replace each with the rule the team already applies, even if nobody has said it aloud. "Unusual" might become "any payment over twice the supplier's average for the last six months, or to a supplier not paid in the last year". If the team cannot agree on the number, that is worth knowing before anyone, human or agent, runs the task alone.

3. A definition of done

People know when a task is finished because they have seen it finished before. Write it down: the reconciliation balances to zero or every difference has a note, the file is saved with this name in this folder, the review task has been raised.

4. What to do with exceptions

The normal case is rarely the hard part. Most of the time in recurring work goes on items that do not match, which is why we suggested keeping an accounting exceptions log. For an agent, the common kinds of exception need a written response: handle it this way, or stop and hand it back. The second option matters as much as the first.

5. When to stop and ask

A good junior knows when they are out of their depth. An agent needs that boundary written down. List the conditions that send the work back to a named person, with a short note on what was found: a difference over the threshold, a missing input, a new supplier, anything touching a client's tax position. "If in doubt, ask the finance manager" is a perfectly good instruction, as long as that person is named on the task and the rule is written, not just a habit.

A before and after

Here is the same hypothetical step, rewritten.

Before: Compare the supplier statement to the ledger and investigate any differences.

After: Compare each line of the supplier statement (PDF in the supplier's folder for the month) to the open items in the supplier's ledger account. Differences under £5 in total: post to the rounding account and note it. Invoices dated in the last two working days of the month and missing from the ledger: check next month's AP batch and mark as timing if found. Before chasing any other difference, search the AP inbox for credit notes from the supplier dated in the month. Any remaining difference over £250: stop and send the list to the finance manager, with the line items and what was checked. Done when every line is matched or explained, and the working file is saved as SUPPLIER-YYYY-MM.

The figures are illustrative and every firm will set its own. The point is the shape. The second version is longer, and someone covering the task for the first time could follow it without asking a single question. That is the test, for a person and an agent alike.

Who writes it, and how

The person who runs the task today is the only one who knows what belongs in the second version. That makes this work worth crediting, not something squeezed into a Friday afternoon. It is also a chance for them to hand over the dull part of a task and keep the judgement calls, which are usually the parts they enjoy.

Three habits help:

• Write while doing. Capture the steps during a real run of the task, not from memory. The gaps show up when you watch the work, not when you describe it.
• Ask the new starter questions. Have someone who has never done the task read the draft and list every point where they would have to ask. Each question is a missing line.
• Let corrections feed the manual. When an agent starts by drafting and a person approves, as in our post on letting AI in drafts first, every correction the reviewer makes is a sentence the manual did not have yet. Add it.

This is also part of what we are designing Beaks to do: keep the manual attached to the recurring task, written as the work happens, so it is the same document a person and an agent both work from.

One thing to do this week

Pick one recurring task with a written procedure and read it with a pen. Circle every adjective that hides a judgement ("unusual", "material", "correct", "as normal") and every step that does not say where its input comes from. Sit with the person who runs it for twenty minutes and replace each circle with the rule they actually use. Even if no agent ever touches that task, the next person to cover it will thank you.