Agent factories will not stay in the engineering department

The way engineering teams now work with agents will reach finance, procurement and operations, and the departments that get there first will be the ones whose work is already written down. Not the ones with the most licences.

By an agent factory we mean a loop: work comes in at one door, agents carry it through a fixed sequence of stages, and people stand at the points where something is decided. It is not an AI factory in the hardware sense, and it is not a lights-out plant. It is a way of organising office work, and software teams happen to be the first to try it at scale.

Four steps engineering took first

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

Infographic

From a chat window to a factory

  1. 01A chat beside the workAsk, copy the answer, paste it into the real tool. Most business teams are here.
  2. 02Suggestions inside the toolThe editor, the mail client or the spreadsheet proposes the next line, paragraph or category.
  3. 03An agent driven from one deskOne person steers an agent step by step. The work is still that person's, on that person's machine.
  4. 04The factoryWork enters a shared queue, agents take it through the stages, people approve at the decision points.
The jump from the third step to the fourth is larger than it looks: private work becomes a company process.

In finance the third step looks like a controller steering an agent through a supplier reconciliation on a laptop. The fourth looks like every incoming invoice entering one queue, agents classifying, matching and flagging deviations, and the controller posting only the exceptions. In procurement it is the difference between a buyer asking an agent to summarise three quotes and every new supplier request passing through the same checks, in the same order, with the same record.

Why engineering is first, and not the only one

Engineers did not get factories first because they are better with technology than accountants. They got them because their work has met three conditions for a long time.

  1. The input is written down. A ticket, the code, a description of the change, the history of every version. An agent has something to read.
  2. The output can be checked by a machine. The tests pass or they do not. The build starts or it does not.
  3. Every change leaves a trace. Who changed what, when and why can be reconstructed a year later.

Now look at a typical month-end close. Half of it lives in a workbook one person maintains, the adjustments are explained in e-mails, and whether the numbers are right is settled in a meeting. Supplier onboarding is a chain of forwarded attachments. On a production site, the night shift’s judgement about which line to stop is in nobody’s system at all — we wrote about that in the night shift knows things nobody wrote down.

An agent factory in finance will not start with buying an agent. It will start with what engineers have had for twenty years: a recorded input, a checkable output and a trace of every change. Put differently, a model of the business that a person and an agent can both read.

Two jobs instead of one

In teams that already work this way, every engineer has two jobs. The first is the product. The second is the factory that builds the product: correcting the agents’ instructions, choosing which model does which stage, measuring how much work went through without a person and what it cost.

The same split is coming to the controller and the buyer. The controller still signs the month and still knows more about the business than any agent. But a growing part of the job becomes the close loop itself, and making next month’s loop raise fewer exceptions than this month’s. That is a change in who does what, and it has to be decided, not left to happen. When it is left to happen, the system goes live and the organisation carries on as before, as described in the system went live and nobody changed who does what.

Where the factory does damage

Two traps are already visible in engineering, and both will move into the rest of the company.

The factory with nobody in it. Part of the market sells factories without people. We work on unattended operation in plants, and even there the rule is that anything outside the known list of exceptions stops and calls somebody. An office loop deserves the same rule. An agent that sends an upsell offer to a customer with an open complaint, or posts a quarter-end correction on its own, does not save work. It creates work for the legal department. The agent analyses, prepares and proposes. A person decides, signs and answers for it.

Work that suddenly becomes visible. At the third step, how somebody talks to an agent is their own business. In a factory every instruction and every correction is visible to the team, and the cost of each person’s work becomes easy to count. If management uses those numbers to rank people, the best leave first, because they feel like replaceable parts. The factory’s numbers are for improving the factory.

How you know it is working

Vanity metricOutcome metric
Documents, e-mails and reports generated by agentsHow many times a person had to touch one task before it was closed
People with a licence for an AI toolShare of tasks that went through the loop with no correction after approval
Tokens used by the departmentCost of one closed task, and its trend month to month

Be honest about what the factory cannot fix. A campaign can fail on price, and a close can slip because a supplier never sent the invoice. The factory shortens the way to the truth. It does not change the truth. How much really goes through without a person is a number worth measuring before anyone promises it, which we argued in how much really goes through untouched.

Where to start on Monday

In one department, pick one task that comes back every week and has a clear input and output: the sales report for the board, a product description for the online shop, a bank statement reconciliation. Write it down as if you were handing it to a new colleague: what arrives, what must leave, how you tell it is right, and who approves. If it does not fit on one page, an agent will not do it either, and neither will the new colleague. The case for starting with dull, recurring work is in your first agent should do the boring job.

Our part

In AI transformation we start by checking which processes meet the three conditions today and which need writing down first, and we change the roles together with the tools. When a department is not ready for a loop, we say so, and the first piece of work is the written process, not the agent. More about predictes is on the home page.

What remains is a choice. You can wait until agent factories mature in engineering and only then think about the rest of the company. Or you can start writing down the work in finance, procurement and operations now, so that when the tools are ready, the first project is not finding out how your own processes actually run.