An employee doesn't understand something in a software package and asks a question. No outage, no broken system, just unclarity about use. This work today ends up at the service desk or with functional management, usually through a ticketing system, sometimes by phone or simply at the desk of the colleague who "knows everything" about that one package.
The question at hand is not whether this work is annoying or valuable. The question is: does the pattern of this task align with what AI can handle with text today.
Three axes are decisive here: volume, customer contact, and structuredness.
The volume is high. The same questions recur, time after time, from different people: how do I export this report, where do I find that setting, why am I getting this error message. That repetitive character is exactly what automation runs on: a task that occurs a hundred times identically is a different task from one that occurs once and is different every time.
Customer contact is limited. The person asking is a colleague, not an external customer with whom a relationship needs to be maintained. Less restraint is needed to give an answer directly and without detour, and there is less risk that a less fluent answer will be experienced as unprofessional.
The structuredness is high. The question often has a fixed form ("how do I do X in system Y") and the answer, in most organizations, already exists somewhere: in a manual, a knowledge base, old tickets. AI doesn't need to invent something new here, but to retrieve something existing and reformulate it to the situation of the person asking.
On top of that, error costs and compliance score favorably. A wrong answer about an export button rarely leads to damage, and there is generally no legal requirement prescribing that this answer may only be given by a human. Physical actions are not needed: this is purely a language and knowledge task, which explains the score on that axis.
Creativity scores low, and that is precisely where the limit lies. A question that falls outside the standard patterns — an unusual combination of settings, a problem that actually touches three systems, a user who cannot quite articulate what is going wrong — requires someone who can probe further and can deviate from the script. Judgment therefore scores middling: part of the questions can be answered unambiguously, another part requires an assessment of what the person asking actually means.
This is therefore not a complete takeover. It is a division: the questions that occur often and have a fixed answer are promising for AI, with — where necessary — an employee who reviews the answer before it goes out the door. The questions that deviate remain human work, not because AI wouldn't want to do that work, but because the pattern is not suited for it.
This only works if the underlying knowledge is in order. An up-to-date knowledge base and well-structured FAQs are not a side issue, they are the precondition. An AI system that gives answers based on a manual from three versions ago gives wrong answers with just as much confidence as correct ones. Companies where documentation is fragmented, resides in the heads of individual employees, or has never been maintained, will not be able to benefit from this pattern right away — not because the task is different, but because the precondition is missing.
At an organization with one simple, stable software package and a small, experienced user group, the share of questions that can be answered via AI is likely large: little variation, much repetition. At an organization with dozens of systems, much customization, and regularly changing employees, that is different: more variation, more questions that fall just outside the standard pattern, and thus a larger share that remains with a human.
This task is also not separate from the rest of the service desk. Anyone looking at how AI can register and prioritize incidents or at how an outage is communicated to users sees a comparable pattern of repetition and structure, with slightly different outcomes per axis. A broader overview of what that leads to for customer service as a whole can be found at what work can AI take over in customer service.
In companies where this has already been set up, the service desk sees fewer repeat questions coming in to people and more time left over for the questions that do require probing further. That is not a vision of the future: it is happening today, in part, at organizations with a knowledge base that is in order. At other organizations it is not happening yet, simply because the foundation for it is still missing.
What an organization does with the freed-up hours — a different role setup, a different division of tasks — is a choice that is up to the organization itself and for which its own legal requirements apply; that is not a question that can be answered with a task analysis.
To see how this pattern plays out in your own situation, an indication can be obtained via the free quickscan: twelve questions, no account needed, with an estimate of what part of the hours in your profile can be taken over by AI today. The full work scan, which breaks down the work of an entire company into tasks, is still under construction.
Vraag maar. Ik ken de kennisbank van deze site; wat ik niet weet, zeg ik erbij.
Answers come from this site’s knowledge base. Not tailored advice, and not a scan of your company.