AI agents in daily work: the real issue is control

Since May 2026, every support ticket for a software product we run lands with an agent. It picks up the ticket, logs in, opens a coding session in Cursor, fixes the bug and ships the fix through GitHub. That is about 50 tickets a month.

At first we reviewed every fix. After a few weeks it was clear that about 95% were right the first time. Today we wave most of them through and look more closely at the rest.

I am putting this first because most writing about AI agents stays hypothetical. It lists what agents “can” do. I care about what they do next week at a service firm with 20 or 80 people, and what goes wrong along the way.

What our agents do today

Support tickets through to the fix. Described above. The agent has access to the helpdesk, the code and deployment, and to nothing else. We mostly run it on Grok in Cursor, because that model is fast and cheap enough for this kind of bug.

Project work in continuous dialogue. Most of our work today runs from coding agents such as Claude. Leadtime is connected through an MCP server. MCP is an open standard that lets an AI assistant access another system. I talk a task through with the agent, and whatever comes out of it is recorded properly in the ERP as a task, a time entry or a note. Some days almost everything runs this way.

The agent built into Leadtime. I ask it for numbers from the system or have it prepare things, such as creating and checking invoices. They only go out once a person has looked at them. Where money is involved, that stays the rule.

All three have one thing in common: the agent works inside the system, with its own access and its own role. It does not sit in a chat window next to the system while someone copies the result across.

When an agent gets stuck, it is usually the model

We have not really had an agent fail. When one gets stuck on a task, switching to a stronger model usually does it.

That has a consequence for planning: choosing the model is a cost decision. A large model for every small task gets expensive. A small model for everything fails on the hard cases. If you run agents in production, you have to decide per task which model it gets, and you need to see what that costs.

Where I do not use agents

In sales, and above all in outreach. You can almost always tell when cold outreach was written by an agent, and the recipient draws a conclusion about the whole company.

Ticket replies are different. When a client asks about status or how a feature works, an agent can answer, provided it takes the answer from the system and does not make it up.

“We already have ChatGPT, that’s enough”

I hear this a lot from managing directors. My estimate is that anyone working this way has seen maybe 3% of the potential. I have not measured that; it is my impression from many conversations.

A chat window answers questions. It does not pick up tickets, log time, draft invoices, or know which project is running over budget. The value only appears once the agent is connected to the systems where the work happens.

How far most companies are from that shows in the latest Bitkom survey in Germany: 57% of companies with 20 or more employees use AI. But of those using or planning AI, only 11% use agents, 29% are planning to and 31% are still discussing it.¹ I expect the next two years to decide who ends up on which side of that gap.

The real problem is control

Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls.² Of those three, I think the last is the biggest.

In most companies I know, every employee is at a different level of AI maturity and uses their own tools. One writes proposals with a private subscription, another has code reviewed by an assistant nobody approved. It is a sprawl, a shadow economy next to the official systems.

• In a survey of almost 5,000 employees by the German Economic Institute, 29.1% of people using AI at work said they use tools not provided by their employer.³
• According to Bitkom, 8% of companies say private AI tools are widespread among their staff, and another 17% know of individual cases.⁴

Companies tend to underestimate this, because people rarely mention using something that is not allowed. For a service firm it is more than a data protection issue. When work happens in private chat windows, it is missing from the system: no time on the project, no note in the ticket, no trail for the colleague who takes over tomorrow.

AI makes the individual faster, but the company learns nothing from it.

The cost gap

On top of that comes the huge gap between pricing models. Individual subscriptions are extremely generous: a fixed monthly fee buys a lot of compute. But you cannot simply use them in a business setting, and they are usually not meant for agents that work on their own.

If you run agents properly through the providers’ APIs, you pay per token, that is, per piece of text processed. Costs then depend on the task and the model, not on the number of users. If you cannot see that per task, you either get a surprise at the end of the month or throttle the agents out of caution until they stop being useful.

What a controllable system needs

You really have to get deep into the subject. From our own work, it comes down to four things:

Leadtime, the platform for IT service firms, agencies and consultancies, is built this way: agents have their own accounts and roles and work in the same process as the team. Work out in the calculator how much time that frees up in your routine. It is a model with all assumptions visible.

Frequently asked questions

Sources