Make or Buy in the age of agents. A model.

Essay #002 · 2026 · Reading time 6 min · Field: Make/Buy · Market: Mittelstand/Trades

Ten years ago, make-or-buy was usually easy to calculate. There was a table. On the left, the cost of building: salaries, infrastructure, maintenance. On the right, the cost of buying: license, customization, dependency. One side ended up with the better number. You signed.

With AI agents, this table no longer works. Not because the numbers have become harder: the columns themselves no longer fit. A “license” today is a subscription to a service whose technology changes monthly. “Building” today mostly means connecting existing components. And “maintenance” means constant readjustment.

This essay proposes a simple grid to sort the decision. It is not a finished framework, just three questions. They have proven useful in my own decisions at arocom and in client projects.

Three questions instead of a cost calculation

The classic make-or-buy question knew two factors: cost and control. Building was expensive and controlled, buying cheap and dependent. For agent systems, that is not enough. I ask three questions.

First question: how close is the system to your core business? An agent processing internal invoices is far from the customer. An agent handling customer conversations is right in the middle. The closer to the customer, the more you should build yourself. The reason is not that building is better, but that you need to adapt quickly there. That outweighs the price.

Second question: how important is your own data? Does the system work with publicly available data, say for research, summaries or translations? Then buying is cheaper and better. Does it work with internal, often sensitive data such as customer history, costing or process knowledge? Then the question becomes strategic. If you put such data into someone else’s system, you hand over more than the license invoice shows.

Third question: how fast does the task change? A compliance check for credit contracts changes slowly. A marketing assistant changes monthly. Tasks that change quickly need a system you can rebuild quickly. That does not have to be in-house. But it is rarely the big enterprise solution.

The three questions yield eight combinations. Most are clear-cut. Far from the core business, no sensitive data, stable task: buy. That is the calendar assistant. Close to the core business, sensitive data, fast changes: build. That is the sales agent for a niche business. The interesting cases are the mixed ones. There, the grid helps you find the center of gravity.

A case from manufacturing

A mid-sized industrial company, two hundred employees, specialized manufacturing with many custom designs. The managing director asks himself: should we deploy an agent that drafts technical quotes?

Quote preparation is a bottleneck in the company. Technical clarification with the customer, alignment with engineering, preliminary costing, writing the quote: half a working day of a senior design engineer per quote. With orders increasing, this slows growth.

A vendor promises a ready-made solution, trained on the internal data. Price: 60,000 euros for rollout, 4,000 euros per month. Attractive, because fast.

The in-house team proposes building it: 140,000 euros in the first year, then 50,000 per year. More expensive, slower, but tailored.

The three questions give a clear answer:

• Closeness to the core business: high. Quotes are direct customer communication. Every inaccuracy shows.
• Own data: very important. Quality depends on how well the system knows customer history, past projects and the internal margin logic. This data is highly sensitive.
• Rate of change: medium. The structure of quotes rarely changes. What changes are prices, materials, constraints. That is master data, not agent logic.

The consequence: the bought solution is risky because the system sits too close to the customer and needs data that is too sensitive. Building is necessary. But because the task changes slowly, the build does not have to be elaborate. A lean, hybrid approach is enough: a base framework from the market, integration and customization in-house. That might cost 80,000 euros in the first year and far less in the years after.

Here the grid recommended neither the most expensive nor the cheapest option, but the one that fits best.

What the grid does and does not do

The grid is not a calculator. It does not tell you which vendor to buy or which architecture to build. It tells you where the fundamental decision lies and which questions to ask your people before you decide.

That is more than it sounds. In the projects I have seen, the mistake was rarely the chosen option. The mistake was that the decision turned technical too early. Vendors were compared before it was clear what actually had to be decided. Interfaces were discussed while the real question stayed open: do we want to let this data leave the company at all?

The grid forces you to ask these questions before the tool questions.

Three bad decisions I keep seeing

“We buy now and build later.” Looks cheap short-term and gets expensive long-term. If you buy, you build on technology you do not control. In the cases I know, the later switch cost more than building in-house from the start would have.

“We build everything ourselves.” The decision driven by fear of losing control. It is also expensive, and it ties up people who are needed elsewhere. Not every task is core or needs custom work.

“We will decide when the technology is more mature.” Sounds reasonable. In my experience, it is usually the worst choice. While you wait, your competitors gain experience. And decisions still get made in your company: by employees who simply use the tools. Just not by you.

The decision stays with you

The grid structures the decision. You still have to make it yourself. Essay #001 describes how employees pull ahead of their bosses on AI when leadership does not decide. The same applies here: if you hand the make-or-buy question to external consultants or your own IT, you hand over a core question of your business. Where agents work and where they do not determines how your company operates in five years. That is not an IT decision — it is yours.

The first step is small: take one process where a tool purchase is currently under discussion in your company. Answer the three questions in writing: closeness to the customer, data sensitivity, rate of change. Only then look at vendors.

— Axel Roth