Agent factories will not stay in IT

Agent factories are leaving the IT department for procurement, finance and customer service, and which function benefits first will depend on the state of its process, not on the tool it buys.

By an agent factory we mean a loop in which work enters in one place, agents carry it through a series of stages, and people stand at the decision points. Software teams already work this way. A ticket lands in a queue. An agent sorts it, drafts a specification, prepares the change, gives it a first review and attaches evidence that it works. A person approves the specification, confirms the result and decides on the release. Most of the rest of the company is about two steps behind.

Why developers got there first

Not because they are cleverer than buyers or accountants. Their work has met three conditions for years that most business functions still do not.

The input is written down. Tickets, code, change descriptions, version history. An agent has something to read.

The output can be checked. Tests pass or they do not. The application starts or it does not.

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

Now look at procurement. Requests arrive by email, by phone and on a note left on someone’s desk. The choice of supplier is justified in conversation. Whether the purchase was a good one becomes clear a quarter later, if anyone checks at all. An agent in that process has nothing to read, nothing to compare its output against and no trail anyone could learn from.

Four steps from chat to factory

Step	Developers	Procurement	Customer service
1. Chat beside the work	asks a chatbot, pastes the answer into the editor	asks a chatbot to draft a request for quotation	asks a chatbot for a polite reply to a complaint
2. Suggestions in the tool	the editor completes the next lines	the purchasing system suggests a category	the ticketing system proposes a reply template
3. One person’s agent	a developer steers an agent step by step on a laptop	a buyer steers an agent that compares quotes	an agent pulls together a customer’s history for one adviser
4. Factory	a ticket enters the loop; a person approves the spec, the verification and the release	a request enters the loop; agents check budget, contracts and quotes; a person picks the supplier	a complaint enters the loop; agents reconstruct the case and draft a decision; a person makes it

Most business functions sit at step one or two, with individuals at step three, often without their managers knowing, as we wrote in the piece on people who already use AI. The gap between steps three and four is wider than the table suggests. At step three an agent extends one person. At step four the work stops being private and becomes a company process.

This is a change project, not a purchase

Moving to a factory looks like a buying decision: pick a platform, buy licences, connect the systems. The platform is the smallest part of the job. The bigger part is a set of questions no vendor can answer for you. Who writes the request, and in what form. Who approves the supplier once an agent has prepared the comparison. What happens when the agent is unsure. Who owns the outcome when a machine did half the work.

If the process is not defined, the factory will not define it. It will freeze whichever version happened to make it into the agent’s instructions, as we argued about agents in undefined processes. The second condition is less visible: the team needs room to change how it works. If a department is already carrying three other programmes, a fourth will not land, however well it is designed. The constraint is change capacity, not technology.

Two jobs instead of one

In teams that work as a factory, everyone has two jobs. The first is making the product. The second is improving the factory that makes it: refining agent instructions, checking how much work went through the loop without corrections, and what that cost.

Buyers and service advisers face the same split. The buyer is still accountable for whether the supplier was a good choice, because they know the market better than any agent. But a growing share of their week goes into making the next request need fewer of their corrections. When an agent compares quotes badly, the buyer fixes the instruction, not just the output.

That second job does not happen by itself. If there is no time for it in the plan, the buyer will fix the output, close the case and move on, and the factory stays where it was in week one.

Where automation does harm

A factory with nobody in it. Part of the market promises loops in which nobody signs off. In procurement that means an agent placing an order with a new supplier on its own. In customer service it means an agent promising a discount to a customer in a dispute. That does not save work. It moves it to the legal team. The agent analyses, prepares and proposes. A person decides, signs and answers for it.

Work that suddenly becomes visible. At step three, how someone works with an agent is their own business. In a factory every correction is visible, and the cost of each person’s work is easy to count. If the board uses those numbers to rank people, the best will leave first. The pushback you then see will not be irrational, as we described in the piece on resistance to change. Factory numbers are for improving the factory.

Automation where mistakes are expensive and surface late. Classifying purchase requests can go through the loop from day one. Changing payment terms with a key supplier should not.

How to tell it is working

Vanity metric	Outcome metric
how many quote comparisons and replies the agent produced	how many times a person had to touch one case before it closed
how many people hold a licence for the tool	what share of cases went through the loop with no corrections after approval
how many tokens the department used	what one closed case costs, and how that changes month to month
“we rolled out agents in three departments”	whether time from request to order fell against the baseline

A factory shortens the road to the truth. It does not change the truth. If a purchase was poor because the market moved, the loop will show it sooner. It will not make it a good purchase.

A test for this week

Pick one task outside IT that comes back every week. Write it down on one page as if handing it to a new hire: what comes in, what should come out, how to recognise a good result, and who signs off. If it does not fit on a page, or three people write it down three different ways, the first project is the process, not the agent. That kind of diagnosis is where VUCALE’s work begins.

What remains is a choice. You can wait for agent factories to mature in IT and only then think about the rest of the company. Or you can start writing down the work in procurement and customer service now, so that on the day the tools are ready, you are not starting with the question of how the process actually works.