Agent factories will reach accounts and purchasing, not just developers

Agent factories will not stay with developers: in a mid-sized company they will reach accounts, purchasing and sales, and whether they work will be decided by the company’s IT, not by the choice of agent. A factory is a loop in which agents carry a task from stage to stage and people stand at the decision points. Development teams already work this way. The rest of the company will get there for the same reasons: cost, speed and control.

So the question for a forty-person company is not whether it arrives. It is whether there will be anything to build it on when it does.

Four stages developers have already been through

The easiest way to see what is coming is to look at the team that is two steps ahead.

Stage	For a developer	In accounts	In purchasing
1. Chat beside the work	asks a chatbot, pastes the answer into the editor	asks a chatbot for a spreadsheet formula	asks a chatbot to draft a supplier email
2. Suggestions in the tool	the editor completes the next lines	the accounting package suggests a ledger code	the ordering system suggests a supplier
3. An agent driven from the desk	steers an agent step by step on their laptop	steers an agent that reconciles a statement	steers an agent that compares quotes
4. The factory	a ticket enters the loop; a person approves the plan and the release	an invoice enters the loop; agents match it; a person posts the exceptions	a request enters the loop; agents gather quotes; a person chooses

Most departments are at stage one or two today. The gap between three and four is bigger than the table makes it look. At stage three the agent is an extension of one person and their laptop. At stage four the work stops being private and becomes a company process.

Why developers go first

Not because they are cleverer than accountants. Their work has long met three conditions the rest of the company usually does not.

The work is written down. A ticket, a description of the change, a version history. The agent has something to read.

The result can be checked. Tests pass or they do not. The program runs or it does not.

Every change leaves a trail. Who changed what, when and why can be reconstructed a year later.

Now look at your own office. Invoices arrive in three mailboxes, one of them personal. Approving a purchase means replying “ok” in a thread nobody will find again. Month-end lives in a spreadsheet one person understands, and that person is on holiday this week. It is the same problem we described in your continuity plan is one person. An agent will not fix it. It will inherit it.

What the factory needs from company IT

Those three conditions sound like process questions. In a mid-sized company they are largely questions about the IT environment. That is the part SovereOS works on, and it is where the factory plugs in.

One way in. Invoices, requests and queries need to land in a shared mailbox or queue, not in personal folders. A factory with six front doors does not know what is on its table.

Files where both a person and an agent can reach them. On the company’s shared storage, in open formats, with permissions granted to groups rather than individuals. An agent that reads supplier contracts has to know where they are, and must not see the payroll on the way.

An identity for the agent in the directory. An agent acts with somebody’s permissions. If that is simply an employee’s account with their saved passwords, nothing separates what the agent did from what the person did. That is why identity is the real lock-in: the directory has to know about agents too, and be able to revoke their access in the same step as the person’s.

A record. What the agent was asked, what material it used, what it produced and who approved it. Without that you cannot measure the loop, or answer a customer who asks where a number came from.

None of this is new. It is the same groundwork we described for shadow AI and for a company laptop for AI agents. A factory simply stops forgiving its absence.

Two jobs instead of one

In teams that already run a factory, everyone has two jobs. One is the product. The other is improving the factory: fixing the agents’ instructions, checking how much work went through without corrections, and what it cost.

The controller and the buyer have the same future. The controller still signs off the month, because they hold context no agent has. But more of the job becomes the close loop itself, which lets through fewer exceptions each month. When an agent mismatches an invoice, the fix goes into the instruction, not just the ledger.

The agent prepares. A person decides, signs and answers for it. That order does not change.

Where automation does damage

A factory with nobody in it. Some of the market promises loops in which nobody approves anything. In an office that is simply dangerous. An agent that sends a payment reminder to a customer with an open complaint, or posts a quarter-end correction on its own, does not save work. It creates it. Anything that reaches a customer or a payment stays a draft until a person approves it.

Work that suddenly becomes visible. At stage three, how somebody talks to their agent is their business. In a factory every instruction and every correction is visible, and it is easy to work out what each person’s work costs. If the board uses those numbers to rank people, the best ones leave first. Factory numbers are for improving the factory, not for league tables.

Automating a process nobody wrote down. An agent will speed up a mess as happily as it speeds up order. Faster, and with better formatting.

How to tell it is working

Vanity metric	Outcome metric
how many emails, reports and descriptions the agent produced	how many times a person had to touch one task before it closed
how many people hold an AI licence	what share of tasks went through the loop with no corrections after approval
how many requests the department sent	what one closed task costs, and how that cost moves month to month

To be fair about it: month-end can still slip because a supplier never sent the invoice. A factory shortens the road to the truth. It does not change the truth.

Where to start on Monday

Pick one task in one department that comes round every week and has a clear input and output: reconciling a bank statement, the weekly sales report, comparing quotes for materials. Write it down as if you were handing it to someone new: what comes in, from where, what should come out, how to tell it is right, and who approves it.

Then ask three questions about IT. Does the input land in one place? Are the files the task needs somewhere a person other than their author could read them? Is there a record of who approved what? If any answer is no, you know where to start, and it is not buying an agent.

That leaves a choice. You can wait for factories to mature among developers and think about the office afterwards. Or you can sort out the way in, the files, identity and the record now, so that on the day the tools are ready the company does not first have to find the one person who knows where everything is.