The honest answer is: rarely a whole role, often meaningful parts of several. Which sounds like a dodge until you look at how roles are actually composed.
Why whole-role replacement usually fails
A job is a bundle. A bookkeeper does data entry, reconciliation, exception handling, chasing people who have not sent things, and answering questions from you. AI handles the first two well, the third sometimes, and the last two badly.
Replace the person and you have automated sixty per cent of the role and orphaned the rest — which then lands on you, the owner, who is the most expensive person in the business.
This is the most common way an AI hiring decision goes wrong. The saving is real and the residual work is invisible until it arrives on your desk.
The better question
Not "can AI do this job" but "which parts of this job would I be delighted never to think about again, and can a machine do those reliably?"
That reframing changes the answer. Instead of replacing a bookkeeper, you remove data entry and reconciliation and the same person handles more volume, or spends their time on the exception work that actually needs judgement.
A four-part test
For any task you are considering automating:
Is it repetitive? If it happens a handful of times a month, the build will not pay back regardless of how well it works.
Can you define correct? If you cannot state what a right answer looks like before you start, you cannot verify the output, and an unverifiable system is a liability.
What is the cost of being wrong? A mis-scheduled meeting costs an apology. A mis-posted transaction costs a restatement. Match the supervision to the stakes.
Does it need context only a person has? Anything requiring knowledge of a relationship, a history, or an unwritten rule is a poor candidate.
Where it genuinely replaces headcount
In our experience it does not replace people already employed. It removes the next hire.
A business at capacity faces a choice: add a person or increase throughput. That is where the substitution is real, and it is a much easier decision than making someone redundant — no disruption, no severance, no morale cost. You simply do not add the role.
If you are considering AI as a redundancy exercise, we would push back. The returns are better and the execution risk far lower when it is a growth-absorption decision instead.
The part people forget
Someone still owns the system. When it breaks, when the process changes, when a supplier alters their format — a person handles that. You have not removed a role; you have changed what it requires.
Budget for that person's time before you count the saving, not after.
Related: How to Choose Your First AI Automation and, on whether any of this affects what your business is worth, Does AI Adoption Increase Enterprise Value?.