Who owns your agent factory if it runs at a vendor?

Two young colleagues put a pale-blue binder and a small external drive on a wooden shelf of pastel binders, with an open laptop standing on the shelf

The most valuable part of an agent factory is not the model but what the factory learns about your company, and that part should stay with you whoever’s model does the work.

The vendor’s presentation runs forty minutes. A virtual team member, live in two weeks, ready connections to the purchasing system and to email. The board asks good questions about price, about security, about who can see the data. Nobody asks what stays in the company when the contract ends. In two years that is the question that will matter most.

Where the problem comes from

While agents run on individual laptops, the issue looks harmless. Everyone uses a different tool and gives it broad access to mail and files. The instructions are written down nowhere. The conversations with agents, which are the most accurate record of how the team really works, disappear when the window closes. That is where most companies are today, as we wrote in the piece on people who already use AI.

An agent factory tidies this up. It gathers the work into one loop with one set of permissions, with metrics and with memory. That is exactly why the question of where the loop runs and who owns it becomes a board decision rather than a team’s choice of tool. Why these loops will leave IT is the subject of the piece on agent factories in procurement and customer service.

What you actually own when you have a factory

A factory is not one program. It is a stack of layers, and each can belong to you or to the vendor.

LayerWhat it isWhat you lose if the vendor owns itThe minimum worth keeping in-house
Loop definitionstages, agents, triggers, rules for moving onyou cannot rebuild the loop elsewherea definition in your own repository, with change history
Agent instructionshow an agent should judge a request, a complaint or a quote in your companyknowledge built up over months of correctionsfiles readable without the vendor’s tool
Permissionswhat the agent can see and changecontrol over who sees your dataone permission register that you manage
Memory and run historythe decisions and outcome of every casethe most accurate process documentation the company has ever hadstorage in your cloud or your data centre
Evaluationshow you measure quality, cost and timethe ability to compare vendors on your own casesyour own set of test cases and results
Modelwhich model does the workfreedom to switch when price or quality changesswapping the model without rewriting instructions
Computewhere the work physically runscontrol over where data goesa deliberate choice: cloud with a no-retention contract, your own, or mixed

The model is second from last. Models get cheaper and change every few months. Instructions, memory and evaluations gain value with every week the factory runs. If you could take only one layer with you, take the memory and the instructions.

Three routes and their price

A ready-made factory as a service. The fastest start. The price is that the definition of the work, the memory and the evaluations sit with the vendor, and customisation ends where the settings panel ends.

Your own factory from scratch. The first agent runs within a week. Six months later your best people are maintaining scripts and queues instead of working on the process, and every new stage in the loop is a project of its own.

Open infrastructure on your side. The vendor provides the control layer: a work queue, a dashboard, metrics, a way to run agents. The definition, instructions, memory and data stay in your repository and your cloud. You can use several models, open ones included, and compare them on your own cases. The price is that someone on your side has to understand that infrastructure and answer for it.

There is no answer that suits everyone. VUCALE resells nothing, so it can say this plainly: sometimes the right choice is the commercial platform the company already has, sometimes a sovereign open-source stack, and most often a mix. Advice backed by a partner margin has its answer before it hears the question, which is the argument of the piece on consultancies that resell.

When sovereignty becomes an expensive hobby

Sovereignty can be overdone.

A company of eighty people does not need its own GPU servers to own its factory. Hardware bought so the data never leaves the building, which nobody in the company can maintain, ends with people going back to agents on their laptops, only at higher cost. For most companies it is enough that the definitions, instructions and memory live with them, that the model provider’s contract rules out training on their data and retaining prompts, and that the architecture allows a switch to a different model within a week.

The second trap is the open model as a goal in itself. It can be cheaper and can run on your own hardware, but on some tasks it will be weaker. A factory pays off when you can pick the model that fits each stage: a cheaper one for sorting, a stronger one for preparing decisions.

Write the exit in at the start

The cost of leaving is lowest before you sign. After that it only rises. So the exit questions belong to the purchase, not to a review two years later. In what format will we get the instructions and the run history. How long does the export take. Do the instructions work with another model without rewriting. What happens to the data after termination.

The answers are worth testing, not just recording. The simplest way is a trial on your own cases before you decide, as in the piece on testing the platform on your data: run the loop on a hundred real purchase requests, then export everything and try to rebuild even one stage somewhere else.

How to measure sovereignty

Sovereignty is not measured by a line in the security policy. It is measured by what leaving costs.

Vanity metricOutcome metric
“our vendor meets every requirement”how many working days it would take to move the factory to another model provider
the number of models available in the dashboardwhat share of cases run on models you could host yourself
“everything is in the cloud”whether you can rebuild last quarter’s run history without access to the vendor’s account
the vendor’s number of certificationswhether the agent instructions live in your repository and someone reads them

A test for this week

Ask your team two questions. Where do the instructions you use with agents live today, and where are the conversations with them stored? And: if we had to change model provider tomorrow, what would we have to write again from scratch? If the answer is “everything”, your factory, however young, already belongs to someone else.

The choice is easy to name and hard to make. You can build the factory quickly on someone else’s infrastructure and accept that the knowledge of how your company works will sit with the vendor. Or you can build it a little more slowly and keep that knowledge at home. Both routes have a price. The worse outcome is choosing neither on purpose.