AI agents can read an email, look up data and take action in other systems without help. They are everywhere, it seems, and more and more we hear the same question: why not let an agent run the whole process?
For a process that runs a few times a month, an agent can work fine. For a process that runs thousands of times a month, it depends on one thing: does the agent do the same thing every time?
What AI agents are and how they work
An AI agent is software that gets a goal instead of a list of instructions, the core idea behind agentic AI. You tell it what result you want – for example "handle incoming repair requests" – and the agent then decides for itself which steps to take, in which order and in which systems.
That flexibility makes agents impressive, but also hard to rely on in certain situations. Because the agent works out its steps again for every request, two identical requests can follow two different routes. Usually both lead to the right result. Now and then, though, the agent skips a step, and there is no way to tell in advance which request it will be and when.
Benchmarks show how fast this adds up. In Sierra's τ-bench study from 2024, the best model of the time succeeded in only about 25% of retail customer service tasks when it had to get the same task right eight times in a row. Models have improved since, but they still decide their steps anew for every request.
Why reliability matters at high volume
We tested this in a simulation built on repair requests, a process every housing association knows. The housing association and tenant below are fictional.
One fine Friday afternoon the following email arrives in the repairs inbox of a housing association:
Hi, since this afternoon there's been water under the boiler in the kitchen. Photo attached. Can you send a technician to take a look? I'm home all day Monday. Regards, Sanne
The repairs team has to work out what's broken, look up Sanne's home, decide who pays for the repair, book a technician and tell her when they're coming. At a large housing association, that work repeats about 3,000 times a month.
So we tried it. We connected an AI agent to the systems a repairs employee works in and told it to handle incoming repair requests. Then we sent it Sanne's email three times.
The first time, the agent recognised a leaking boiler, treated it as urgent and booked a technician for Monday morning. The second time, it decided the leak could wait and scheduled a regular repair twelve days later. The third time, it sent Sanne a confirmation without ever reserving a technician, so on Monday she would wait at home for nobody.
What makes the third run hard to catch is that the agent reported all three requests as handled. The first sign of trouble would be Sanne calling on Monday afternoon to ask where the technician was.
The agent didn't break down in any of these runs and did exactly what it was built to do: pick its own steps for each email. Two of those choices left Sanne worse off, and the agent flagged neither. We only noticed because we checked what the agent had actually booked. A repairs team handling 3,000 emails a month has no time to check every one, so it has to trust that "handled" really means handled.
Suppose the agent gets one request in a hundred wrong. 99% accuracy sounds good enough to sign off on, but at 3,000 repair requests a month it means 30 tenants staying home for a technician who won't show up.
There is a cost side too. In our simulation, at typical 2026 prices for a mid-range AI model, the agent used about 55 cents of AI per request, which comes to around €1,650 a month for this single process. That is less than an employee would cost. AI itself keeps getting cheaper, but agents use far more of it: Gartner expects the cost per agentic workflow to rise more than fivefold through 2028 (Gartner, August 2026). Even so, the missed appointments weigh far more than the bill.
AI agent vs fixed route
Only one part of the repair process from our experiment needs real understanding: reading an email, sometimes with a photo, and working out what is wrong and how urgent it is. Everything after that follows the association's own rules, so that is how we would build it: AI does the reading, fixed rules decide who pays and which time slot to book, API connections look up the tenant and reserve the appointment. If the maintenance system has no API, an RPA robot will create the work order on screen the way an employee would. Together these steps form a fixed route with the same steps in the same order for every request.
The AI step can still get something wrong, but only in that one place which you can test and back up with simple rules. A rule such as "water on the floor is always urgent" would have stopped the agent from pushing Sanne's leak back twelve days. When the AI isn't sure what a request is about, the route should send it to the repairs team.
Explore what’s possible
See how AI, APIs and RPA work together in practical automations for your organisation.
Explore the automation library →
When to use AI agents
We use AI agents every day and mostly at the moment people don't expect: while building process automations for our customers inside the Yara platform. Working out a fixed route means breaking a process into steps, choosing the right technology for each one and thinking through the exceptions. With YaraBuilder, the AI assistant we use to build automations, you can describe the process the way you would explain it to a new colleague, and YaraBuilder will propose the steps along with suitable technology for each of them. A human – either you or our engineers – will then have to approve or adjust if needed. If a situation comes up later that the route doesn't cover, the process will stop and YaraBuilder will suggest an addition for the process owner to approve.
Letting an agent run a process on its own makes sense in two cases: when the process runs so rarely that a fixed route would cost more to build than it saves, or when every case is different and there are no fixed steps to capture.
Conclusion
Agents are coming into everyday work fast: in Gartner's 2026 survey of CIOs, 17% of organisations had deployed AI agents and more than 60% expect to within two years. Gartner also expects more than 40% of agentic AI projects to be cancelled by the end of 2027 because of rising costs and unclear business value, so the useful question is where to put agents to work.
For most high-volume processes, the answer is a mix. In the repair process from our example earlier in this article, AI reads Sanne's email and judges how urgent it is, and fixed steps book the technician the same way every time. An agent could run a process on its own if that process runs only a few times a month or if every case is different. For a high-volume process like repair requests, an agent is more useful earlier on: it helps design the steps and the automation then runs those steps. When a process runs thousands of times a month and the same request must always get the same result, the steps after the AI reads the request should be set in advance.
For a practical example of this combination, explore supplier invoice intake: AI helps with uncertain cases, while fixed checks handle the next steps.
