Skip to content

AI · Article

You hire an agent, you don't install it

Software is bought per user because a person does the work. Once an AI agent finishes a step of work on its own, that logic stops fitting. Why industrial companies will soon staff agents like a role.

By demi editorial team5 min read

All posts

At some point in the coming months, a purchasing department will ask a question that no longer fits: "How many licences do we need?" What they mean is an AI agent that pre-sorts complaints, matches delivery notes against orders or checks requests for quotation for completeness. The honest answer is: none. Or one. Or three hundred. It is the wrong question.

For decades, software was sold per seat, and that made sense. A person did the work and the program helped. More people meant more seats, so more licences. That logic breaks the moment software finishes a bounded step of work by itself. Then you are no longer buying a tool. You are filling a task.

What has changed

The research group METR measures, on software and research tasks, how long a task can be, in the time a person needs for it, and still be solved by AI models with 50 per cent reliability. In early 2025 that figure was around 50 minutes for the best models, and it has doubled about every seven months since 2019 (METR). The authors add their own caveat: whether their results carry over to real work beyond their test tasks is open. The direction is still clear: in these tasks at least, models no longer handle single inputs only, but whole steps of work.

Vendors understood this before buyers did. Salesforce sells its agents at 2 US dollars per conversation or 0.10 US dollars per action, next to the classic per-user licence, and argues that cost should follow outcomes (Salesforce). That is one vendor's price list, not proof of a whole market. But it shows where billing is moving: from the seat to the finished work.

Why industry feels this first

By 2040, 13.3 million people in Germany's labour force will reach the statutory retirement age. That is 30.0 per cent of everyone who was available to the labour market last year (Destatis). With them go people who often have known for decades why a process is the way it is.

At the same time, the labour market shows something odd. In 2025, for the first time since the pandemic, there were more qualified unemployed people than openings for qualified workers. Even so, one in three of those openings could not be filled on paper, more than 369,500 on average (IW/KOFA). The gap shrank by almost a quarter compared with 2024, in manufacturing occupations mainly because of the weak economy. The researchers explain the rest not by headcount but by fit: the qualifications on offer do not match the profiles companies look for.

That is the core of it. An agent does not close a headcount gap. It closes a gap in a task. And you don't describe a task with a licence count. You describe it with a job description.

What "hiring" means in practice

When you hire a person, you write down what the role does, what it may decide and where its limits are. An agent needs exactly that. Which documents may it read? Which cases may it close, and which may it only prepare? Above which amount must a person approve? If you can't answer these questions, you don't have a task yet, only a hope.

Then comes the trial period. Not a demo with hand-picked examples, but real cases from recent weeks, with clear measures: how many cases were right, how many had to be corrected, how long did it take? It is the same review a new clerk gets after three months.

And there is a supervisor: someone in the line department who checks results, reports errors back and decides in the end whether the agent stays. Without that person nobody is responsible, and without responsibility no result gets better.

Here is a surprising twist. In a field study of 5,179 customer support agents, an AI assistant raised issues resolved per hour by 14 per cent on average and by 34 per cent for novices, with little effect on the most experienced (NBER). The authors read it this way: the tool passes the working methods of the best on to everyone else. For industry, that means the experienced expert is not the one the agent replaces. She is the one who trains it. Her knowledge becomes onboarding, for people and for agents.

What shifts inside the company

Once agents are roles, different people talk about them. Purchasing negotiates performance and price per case, not user counts. IT security asks about rights as it would for any new colleague: what may this role see, and what not? The works council wants to know how results are checked and who is liable when a case was closed wrongly. And the budget moves from IT to the line department that solves the task today with overtime.

That is unfamiliar, but not hard. Every department knows the job description format, many know it better than any requirements document. The reason named most often is missing knowledge anyway: of German companies without AI that had already considered using it, 72 per cent gave this reason (Destatis). Describing a task so that a new colleague could take it over in her first week needs no AI knowledge. It needs knowledge of your own work, and you already have that in-house.

Our assessment

An assessment, not a fact: by the end of 2027, the first Mittelstand tenders will describe an AI agent as a role, with a task list, approval limits and acceptance criteria, and will pay per completed case instead of per user.

So the hard question is not which tool to buy. It is: which task would you trust to a new colleague in her first week, and how would you know after a month that she is doing it well? If you have a clear answer to that, you have already written the job description.

At demi, an AI agent never has more rights than the person it works for, and the rights come from your org chart.

Sources

  1. METR (Kwa et al.), arXiv, 18 Mar 2025: Measuring AI Ability to Complete Long Software Tasks
  2. Salesforce, 15 May 2025: Salesforce Introduces New Flexible Agentforce Pricing to Accelerate the Digital Labor Revolution
  3. Statistisches Bundesamt (Destatis), 23 Jun 2026: 13,3 Millionen Erwerbspersonen erreichen in den nächsten 15 Jahren das gesetzliche Rentenalter (Pressemitteilung Nr. N039)
  4. Institut der deutschen Wirtschaft (IW), KOFA, Feb 2026: Jahresrückblick 2025: Schwacher Arbeitsmarkt – erste Impulse durch Sondervermögen?
  5. National Bureau of Economic Research (Brynjolfsson, Li, Raymond), Apr 2023: Generative AI at Work (NBER Working Paper 31161)
  6. Statistisches Bundesamt (Destatis), 2025: Nutzung von IKT in Unternehmen (Erhebung 2025)

Your configurator

No use cases yet

Your configurator is still empty

Add a use case from any card or case page and see what it brings per year, with the calculation. Free, no sign-up, no commitment: try it out and change any number.

Browse use cases