How Do I Map a Business Process for an AI Agent?
Map the process from 20 real cases with the person who runs it, not from the SOP, and write it down as four lists: what arrives, every decision as a rule with a number in it, every system the agent reads or writes, and who clears each exception by when. Then run the map against 50 past cases before anyone writes a prompt, and don't build until it explains at least 9 in 10 of them.
Why companies get the map wrong
Most companies hand the agent builder the SOP. I've read a lot of SOPs, and the typical one describes about 60% of what the team does. The other 40% lives in one person's head: the supplier whose invoices always come in two parts, the customer who pays on the 3rd whatever the terms say, the order you hold because the sales rep forgot the PO.
An agent built on the SOP handles the 60% well and sends the rest to a queue. Someone clears that queue by hand, and 3 months later the project gets written off as "AI isn't ready for us". The model was fine. The map was missing 4 in 10 cases.
The second mistake is mapping too much at once. A whole order-to-cash flow has 30 to 50 decisions across 4 teams. Nobody can agree all of them in one workshop, so the map never gets finished. Pick one process that one person owns and starts and ends in the same team.
The five steps I use to map a process
1. Pick one process, one owner, enough volume
I look for a process that runs at least 200 times a month, has one named owner, and touches no more than 3 systems. Supplier onboarding, invoice coding, new customer setup and expense review usually pass. Below 200 cases a month, the build costs more than the time it saves. If two managers each think they own it, settle that first or pick something else.
2. Watch 20 real cases, not the SOP
I sit with the owner for 2 sessions of 2 hours and they work through 20 cases from the last month while I write down every step. I ask "why did you do that?" at every pause. I pick the 20 so that at least 5 are the awkward ones: the late, the partial, the duplicate, the one with a missing field. Those 5 teach me more than the other 15.
3. Write the map as four lists
I don't draw a flowchart. Flowcharts hide the rules inside boxes, and nobody updates them after month 2. I write four plain lists in one document.
| List | What goes in it | Supplier onboarding example |
|---|---|---|
| Inputs | Everything that arrives, in every format it arrives in | Web form, emailed PDF, bank letter, VAT certificate |
| Decisions | Every judgement, written as a rule with a number | Annual spend over £50,000 needs a second approver |
| Actions | Each system the agent reads or writes, marked read or write | Reads Companies House, writes a draft supplier record in the ERP |
| Exceptions | What the agent can't settle, who gets it, by when | Bank details don't match the company name: AP lead, same day |
The decisions list is where the work is. "Check the supplier is legitimate" is not a rule. "Company is active on Companies House, incorporated more than 12 months ago, and the bank account name matches the registered name" is a rule. If the owner can't give me the number, the decision stays with a person for now.
4. Mark who does each step
Next to every line on the map I write one of four labels: agent reads, agent decides, agent acts, person approves. Anything that moves money, changes bank details or sends something to a customer gets "person approves" on day one. In a typical first process, about 70% of the steps end up with the agent and 30% stay with a person. That 30% shrinks later, one rule at a time, once the log shows the agent got it right.
5. Test the map on 50 past cases
Before anyone writes a prompt, I take 50 closed cases from the last quarter and walk each one through the map on paper. For every case I ask one question: does the map produce what the person did? I don't build until the answer is yes for at least 45 of the 50. Every miss becomes a new rule or a new exception. This test takes 2 or 3 days and it's the cheapest week of the whole project.
From Hayat's operating seat
I'm Hayat Amin, an AI-native CFO, AI operator and AI coach for CFOs who teaches finance leaders how to build AI into finance operations. I spent twenty years in the C-suite, with three exits and three FT100 listings. Now I build and run AI agent systems inside companies, across finance, sales pipeline, onboarding and compliance evidence, and I coach CFOs one to one.
Every agent I run started as one of these four list documents, and the document is still there after go live. When the new joiner onboarding agent explains a policy in the joiner's own language, it's reading from a decisions list a person signed. When a rule changes, the owner edits one line in that list and the agent follows it the next morning. I learned to keep maps like this before agents existed. When I sold companies, buyers asked who decides what, on which number, and where it's written down. A good process map answers all of that.
Do I need SOPs before I build an AI agent?
No. Most SOPs I read describe the process as it was designed, and leave out the workarounds people built since. You need the 20 cases and the person who handled them. If an SOP exists, I use it as the first draft of the map and mark every step the cases contradict. Those contradictions are where an agent would have gone wrong.
How long does it take to map a process for an AI agent?
For one process with a single owner, about 2 weeks. That's 2 sessions of 2 hours with the person who runs it, 3 or 4 days to write the four lists, and 2 or 3 days to test the map against 50 past cases. A process that crosses 3 teams takes 4 to 6 weeks, mostly waiting for people to agree on who owns which exception.
Who should map the process, IT or the team that runs it?
The team that runs it, with one person who has built agents holding the pen. IT knows the systems and the access rules, so they review the actions list. They don't know why the credit controller holds an order when a customer is 45 days overdue. The person who does the work owns the decisions list and signs it off.
Where I come in
This is work I do inside companies. I pick the first process with you, sit with the person who runs it, write the four lists, test them on 50 cases, and then build and run the agent on top. If you're still choosing the process, read which processes to automate first and how to make agents safe enough for real operations. Finance leaders can start at AI for CFOs. Which of your processes would pass the 45 out of 50 test today? See how I work as your AI agent operator, or start at meethayat.com and tell me.