Which business processes should never be automated?

Lists of things not to automate usually run on difficulty, which is the wrong axis and gets less right every year. A task being hard is a statement about the technology at a moment in time. What actually determines whether automating something is a mistake is structural and does not move: who bears the cost of an error, whether the error can be undone, and whether the interaction is itself what the customer is buying.

The first exclusion is irreversible customer-facing action. Anything that spends money, cancels something, makes a binding commitment or sends an irretrievable communication belongs behind a person, not because a system would decide badly but because the recovery path does not exist. A wrongly drafted reply is embarrassing and fixable. A wrongly issued refund, a cancelled appointment the customer had arranged their day around, or a commitment to a delivery date that cannot be met each create a situation the business must then manage, and the management costs more than the automation saved.

The second is anything a regulator or professional body holds a named individual accountable for. Advice in a regulated profession, a clinical judgement, a statement about someone's legal position, a certification of work performed. The constraint here is not risk in the abstract — it is that accountability has been assigned to a person by rule, and a system cannot hold it. What can be automated around such a task is everything either side of it: the intake before, the follow-up after, the document assembly, the reminder. The accountable judgement itself stays where the rule puts it.

The third exclusion is the one businesses most often get wrong in their own favour, and it is where the interaction is the product. If customers choose you because of how they are dealt with — a specialist retailer whose advice is the reason to buy there, a practice whose manner is the differentiator — then automating that interaction removes the thing being sold while leaving the price unchanged. This is not a quality argument about whether the automation is good. It is that the business has automated its own advantage.

There is a fourth that is temporary rather than permanent and deserves separating: anything the business cannot yet describe. A process that lives in one experienced person's head cannot be automated well, because what gets encoded is a guess at what they do. The correct sequence is to write it down, discover in the writing that it is three different processes with exceptions nobody had named, and automate the parts that survive. Businesses that skip this stage typically automate the common case and generate an exception queue larger than the work they removed.

Everything outside these four is a candidate, and the list of candidates is longer than most owners assume. The point of drawing the exclusions carefully is not caution for its own sake — it is that a business with clear boundaries automates more, not less, because it no longer has to treat every proposal as an open question about whether the whole idea is safe.