Kennisbank

AI and work with high error costs

Why error costs change the assessment

For most tasks, an assessment looks at how much time a task takes and how predictable it is. For work with high error costs, a third factor weighs more heavily: what an error costs if it slips through. An incorrectly addressed newsletter is annoying. An incorrectly calculated dosage, an error in an annual financial statement, or an incorrect release in a production process can cause damage many times greater than the time the task itself takes. That asymmetry changes what AI takeover means there, even if the task itself is just as easy to automate as elsewhere.

It is not about whether AI can handle the task in terms of content. Often it can. It is about what happens in the tail end of the cases where it goes wrong, and how costly that tail end is. That same task, carried out in a context without significant error costs, can be fully taken over. Carried out in a context where an error results in a fine, a claim, or irreparable damage, that task more often remains at the level of "partial, with oversight," or even shifts back to human work, even though the underlying action is identical.

Which tasks shift and which do not

Tasks that shift towards AI takeover with high error costs are often tasks where the error itself is cheap to detect before it has an effect: a second check, a validation rule, a threshold value that a human must confirm. Think of the preliminary drafting of a document that a person ultimately signs off on, or the calculation of scenarios that are always still assessed before they become a decision. That preparatory layer can often be removed by AI, while the confirmation remains with a human. Exactly how that division looks depends on what is described on the page about work that falls under oversight: there, the separation between execution and approval is worked out with reasoning.

Tasks that do not shift are tasks where the error only becomes visible after the damage has already been done: an incorrect diagnosis that has already been made, an incorrect release that has already been sent, an incorrectly signed contract that has already gone out the door. For that kind of task, the assessment does not shift towards more AI takeover but towards more emphasis on the oversight moment itself. That directly affects how a signing process is set up: which steps in it AI can prepare and which step a human must continue to confirm is described on the page about the document signing process.

Why two companies with the same tasks end up different

Two companies with the same task on the payroll regularly arrive at a different outcome in a work scan, and that is not an inconsistency. The outcome depends on how the rest of the company is organized around that task.

A company whose systems connect well to each other can often let an error-sensitive check run automatically with a recorded outcome that can be traced back. A company where data still moves manually from one system to another must first make that transfer error-resistant itself before AI can take something over there without increasing the risk. What that means in concrete terms is worked out on the page about working with systems that are not connected to each other.

The degree of customization also plays a role. A task with high error costs that also barely ever runs the same way twice is harder to hand over to AI than a task that, while risky, does follow a fixed pattern. Where that boundary lies for customized work is described on the page about work with a lot of customization. And for companies working with shift work or varying staffing, there is the additional factor that the oversight moment itself can involve a different person with a different level of experience per shift, which colors the error-cost trade-off differently again, as can be seen on the page about shift work and varying staffing.

A final factor that explains the difference between companies is where the error-sensitive tasks are concentrated within the company. For many companies, a substantial part turns out to lie with HR administration: contracts, leave arrangements, appraisals, where an error quickly becomes a legal issue. How that department turns out in practice is on the page about work in the HR department. If it concerns personnel decisions themselves, separate statutory requirements apply to those; that assessment is not part of this scan.

What you can do now

The classification described here applies in broad terms. Where your own work stands exactly depends on the tasks you actually have carried out and on how your systems, your staffing, and your error sensitivity play out together. The free quickscan provides an initial direction: twelve questions, no account needed, resulting in an indication of what portion of the hours in your profile can be taken over by AI today. The full work scan, which calculates down to the FTE level per task, is still under construction. What is already possible today is getting an initial picture of where in your company error costs shift the assessment framework the most.