Kennisbank

AI and administration: what changes per task

Administration is not one block, but 84 separate tasks

Asking "can AI take over administration" is actually asking the wrong question. The administrative department consists of dozens of tasks that have little in common with each other. Entering a purchase invoice is a different kind of work than calling a debtor who hasn't paid for three months. In our taxonomy for this domain, we distinguish 84 tasks, and based on their nature they fall into three categories: tasks AI can already take over now, tasks where AI does the work but a human approves or rejects it, and tasks that remain human work. The pattern behind that division is more useful than a list of 84 rules.

Where AI takes over: structured, repeatable input

The tasks that can be fully taken over share one characteristic: they are structured, repeatable, and testable against fixed rules. Entering purchase invoices is the clearest example here. Amount, VAT, and general ledger account follow logical patterns that a system can recognize and apply consistently, especially with suppliers that have been entered before. Reading in and matching bank statements also falls into this category: linking a transaction to an open item is essentially a comparison between amounts, dates, and references, something software does better and faster than a person scrolling through the same rows of figures.

Drafting and sending sales invoices fits logically alongside this, provided the underlying data on delivery and pricing agreements is already fixed. What these tasks have in common is that there is no judgment question involved. There is a correct outcome, and the system only needs to find and execute it. Errors can be checked afterward against hard criteria, which keeps the risk of full automation limited.

Where oversight remains necessary: judgment about exceptions and relationships

The largest group of tasks in administration falls into the middle category: AI does the preparatory work, but a human approves or rejects it, for a reason. Sending payment reminders to debtors is a good example of this. Flagging an overdue invoice and drafting a reminder is something a system can do perfectly well, but whether that reminder should actually go out - for a customer currently negotiating a payment arrangement, or for an invoice still in dispute - requires a view that knows the context.

The same applies to preparing and processing outgoing payments. A system can compile a payment batch based on outstanding accounts payable, but the authorization of it is a control moment with a purpose: is this amount correct, has this supplier been approved, is the due date right. That oversight is not a temporary interim step until AI can do it better; it is a deliberate structuring of responsibility and control over cash flows. Processing expense claims and travel costs falls into the same category: recognizing a receipt and entering the amount can be automated, but the question of whether an expense is business-related and reasonable remains a judgment made by someone with knowledge of the policy.

Where human work remains: conversation, negotiation, and responsibility

A third group of tasks remains human work, not because the technology isn't ready for it, but because the task itself revolves around something other than data processing. Following up on outstanding debtor items is the clearest example of this: contacting a customer about an overdue payment and agreeing on how and when it will be settled is a negotiation in which tone, relationship, and flexibility play a role. That is not a data-entry task with one correct outcome, but a conversation with an outcome that varies per customer.

Monthly payroll administration likewise has a core that remains human work, even though individual steps within it can be automated. Entering salary changes and generating pay slips can largely be supported, but final responsibility for correct and timely payment, including the judgment call in deviating situations such as illness, leave, or dismissal, rests with a person. When a task comes close to personnel decisions, its own legal requirements apply, on which this page makes no statement. Preparing the VAT return also remains, despite a great deal of automated preparation, a task for which someone takes final control and submission responsibility, if only because liability for an incorrect return does not rest with a system.

Same pattern, different department

This division - structured input to AI, judgment work to oversight, relationships and responsibility to people - is not a coincidence specific to administration. Looking at what work AI can take over in customer service shows a similar pattern: automating routine questions works well, complex complaints with emotion in them do not. And on what AI can take over in the installation sector, the same question arises around planning and administration surrounding the work itself. The shape of the task, not the sector, determines where AI can and cannot be deployed.

What this means for a specific administration differs per organization: the mix of tasks, the systems already connected, and the way responsibilities are divided all play a role. An indication of that picture for your own situation is available through ftetoai's free quickscan: twelve questions, no account required, with an indication of what portion of the hours in your profile could be taken over by AI today. The full work scan that calculates at task level is still under construction; we deliberately do not offer that here yet.