AI agents in mid-sized companies in 2026: what really works

We think the agent euphoria is about two years ahead of itself – and at the same time we think the practical benefit of AI agents in mid-sized companies is underestimated. The two together sound contradictory, but they aren't: what is shown in product videos doesn't work in 2026. What is unspectacular and narrowly scoped works surprisingly well. This article is an opinion piece from ongoing projects – with the clear intention of drawing the line between the two.

What actually is an “AI agent” in 2026?

Viewed soberly: a language model that doesn't just answer but is allowed to use tools – query data, write documents, create entries, draft emails – and that carries out several steps in a row to achieve a goal. There is nothing more to it. Everything else is a question of architecture: which tools, which permissions, which controls, which stopping point.

The marketing term has detached itself from this. It now describes a promise: that software pursues goals on its own, corrects itself and takes over roles in the company. This promise is the reason why many mid-sized companies are disappointed after a pilot project – not because the technology can't do anything, but because it can do something other than what was announced.

What really works in our projects

Narrow toolboxes instead of all-rounders

An agent with three clearly defined tools and an unambiguous task works reliably. An agent with twenty tools and the brief “support sales” produces busywork. The narrower the brief, the more useful the result – that is the most consistent observation from our work.

Suggesting instead of deciding

The patterns that last in everyday use all end with a person. The agent prepares a draft offer, a person checks and sends it. The agent suggests data corrections, a person confirms them in bulk. The agent drafts a reply, a person decides on the tone. That sounds like half an automation – and it is exactly why these solutions are still being used after six months.

Access to structured company knowledge

The difference between an impressive and a useful agent almost never lies in the model, but in access to reliable information. Those who have their offer logic, pricing rules and project experience in good order get usable results straight away. How we create this basis is described in the article Obsidian as a second brain in your company and on our page AI assistant for companies.

Recurring legwork

Going through search term reports, structuring minutes, checking master data for contradictions, pre-qualifying inquiries, matching invoice items against offers. Unglamorous, easy to delimit, immediately noticeable – and therefore the best starting point.

What doesn't work in 2026

• Long autonomous chains. The more steps follow one another without a checkpoint, the more likely it is that an early mistake goes on having an effect unnoticed. Short chains with interim checks are not a compromise but the sound way to build.
• Write access to core systems without approval. We don't grant unsupervised write access to accounting, ERP or payment processes. The possible time saved bears no relation to the risk.
• “The agent replaces role X.” What gets replaced are subtasks. Roles consist of responsibility, and responsibility can't be delegated to something that isn't liable.
• Agent swarms as an end in themselves. Several agents that commission each other look impressive in a diagram. In practice they multiply sources of error, costs and debugging effort.
• Browser agents on third-party portals. Logins, two-factor procedures and security checks make such automations brittle. Where there is an interface, we use the interface.

Three sentences that make us prick up our ears

• “The agent learns your company by itself.” It doesn't. It reads what you make accessible to it. Without a structured basis, it stays with general world knowledge.
• “You no longer have to define processes.” The opposite is true. Automation forces more clarity, not less. If you can't describe your process, you can't hand it over either.
• “Autonomous, end to end, without supervision.” Possible – if the process is completely reversible and the damage in the event of an error is small. That is rarely the case for offers, invoices and customer communication.

How to get started sensibly

• One process, not five. Choose the one that eats up the most recurring time and can do the least damage.
• Define stopping criteria. When does the agent stop and hand over to a person? Answering this question before the start saves the most rework.
• Log everything. Every step, every tool, every result traceable. Without a log, troubleshooting is guesswork.
• Cap costs. Limits per task and per month, technically enforced, not as a good intention.
• Name an owner. Without a person who evaluates results and refines them, every agent project gets watered down within a quarter.

A concrete example: the draft offer

To make the line between “works” and “marketing” tangible, here is a process from our own work. An inquiry comes in through the contact form. The agent reads it, searches the knowledge base for comparable, already completed projects, draws on the service items and price ranges used there and presents a draft offer with three marked uncertainties: unclear scope, missing information on the number of users, an open question about the interface.

What the agent doesn't do: set the final price, send the offer or make a commitment. A person checks, adds and decides. The time saved is still many times over, because there is no blank screen and no digging through old cases.

The whole debate can be pinned down to exactly this scope: the same setup without human approval would be an “autonomous sales agent” – and would be switched off after the first wrongly calculated offer. With approval, it is boring, inconspicuous and has been in use for months.

How to recognize a serious provider

• They ask about your processes before talking about models. Anyone who lists model names in the first conversation has got the order the wrong way around.
• They name what doesn't work. An offer without a single objection is a sales pitch, not advice.
• They show logs, not just results. Traceability is the central quality question with agents.
• They plan for operation. Models, interfaces and policies change; an agent without maintenance becomes outdated faster than classic software.

The cost question has shifted

Two years ago, model usage was the limiting factor. Today it almost never is: for the vast majority of tasks in mid-sized companies, the running costs of a well-scoped agent are lower than the working time it replaces. It gets expensive in two places – with long autonomous chains that repeatedly process a lot of context, and with integrations into poorly documented legacy systems.

The real cost block remains the same as with any software: clarifying the process, integration, operation. Anyone who mainly discusses model costs in an offer is talking about the smallest item. How we handle operation and further development is described under Managed Digital Office, and you can find the scope of our services under services.

What we expect for the coming years

We don't expect agents to suddenly become autonomous. We expect the tools around them to get better: more reliable interfaces, cleaner permission concepts, better logging, more realistic expectations. The companies that will have the advantage then won't be the ones that shouted “AI first” loudest early on – but the ones whose data and processes are orderly enough for automation to be connected at all.

Conclusion

In 2026, AI agents are a good tool for narrowly scoped, recurring tasks with human approval – and a bad one for anything that carries responsibility. Those who start with them gain time immediately. Those who wait for the fully automated company are waiting for something that won't exist in the foreseeable future. You can find examples from our work under case studies & apps.

Frequently asked questions

What distinguishes an AI agent from a chatbot?

A chatbot answers. An agent is additionally allowed to use tools and carry out several steps to complete a task. What matters is therefore not the language model, but which permissions and which checkpoints you give it.

What should a mid-sized company start with?

With a recurring, clearly defined task where a mistake does little damage and a person approves the result. Draft offers, structuring minutes and data maintenance are proven starting points.

What needs to be clarified about data protection before using AI agents?

Before the start: where the system runs, which data flows into which services and who may access what. We rely on European infrastructure and on models with which your data is not used for training, and we strictly limit an agent's access to the data it needs for its task.