When not to automate

1. The volume is too low

A painful task done four times a year is cheaper by hand than any automation, however elegant. Automation pays back through repetition. If the process happens less than weekly, or takes under an hour a month in total, leave it.

Instead: write a checklist so it is at least done consistently, and revisit if the volume grows.

2. Nobody can write the steps down

If the process lives in one person's head and every case "depends," it is not a process yet. Automating it means encoding guesses, and a model will produce confident-sounding guesses that a person still has to check.

Instead: have the person who does it write down the steps and the decision rules for a month. Most of the "it depends" turns into rules once it is on paper. Then reassess.

3. The process changes every few weeks

Automation encodes how work is done today. If how it is done changes constantly, because the business is still figuring it out or because nobody owns it, the automation will be rebuilt constantly.

Instead: assign an owner, stabilise it for a quarter, then automate the version that survived.

4. The step is a decision with money, people, or legal weight

Credit approvals, hiring, tenant screening decisions, tax positions, disputes, terminations. These should be made by an accountable person, and in several cases the law expects it. Automation can assemble the file and apply the checks. It should not make the call.

Instead: automate everything up to the decision (gathering, checking, routing) and put the decision in front of the right person with the context already assembled.

5. Every case needs judgement

Some work is genuinely bespoke: complex negotiations, custom design, sensitive customer situations. If there is no routine share, there is nothing for the automation to take. Adding an AI step just adds a step.

Instead: look at the admin around the judgement work: scheduling, document assembly, status updates. That part is often automatable even when the core work is not.

6. The real problem is a decision someone is avoiding

A backlog of unmatched invoices is sometimes a process problem and sometimes a sign that nobody has decided which vendor terms apply. Automation makes a process faster. It does not make a decision that is being avoided.

Instead: make the decision. Then see whether the process still needs automating. Often it shrinks.

7. The inputs are too inconsistent to check

AI can read messy documents, but if the same field means different things from different sources and there is no reference data to check it against, the exception queue will hold nearly everything. You have added a review step and removed nothing.

Instead: fix the reference data first: a clean vendor list, a SKU cross-reference, agreed definitions. Then the checking works, and the automation becomes worth building.

8. An off-the-shelf tool already does it

Your payroll provider may already offer certified payroll reports. Your CRM may already sync with QuickBooks well enough. Your accounting software may already read invoices. If a tool you own does the job, turn it on.

Instead: ask whoever you are talking to whether the built-in option would work. An honest answer to that question is a good test of a provider. Our certified payroll page and CRM and accounting integration page both start with that question.

9. You want the customer conversation itself automated

Drafting a reply for a person to review is useful. Sending it unread, or putting a bot in front of an upset customer, is a reputation risk that saves very little. The routine parts of customer communication (status updates, confirmations) can be automated safely. The conversation cannot.

Instead: automate the triage and the drafting. Keep a person on send. See AI email automation for how that line is drawn.

10. Nobody will own it after launch

Automation is not a one-time purchase. Exception queues need working. Rules need tuning as edge cases appear. Systems change. If no one on your side will own the exception queue, and no one will monitor the system, it will quietly degrade and eventually be switched off.

Instead: name the owner before the project starts, and make sure the support arrangement covers monitoring and tuning. Our how we work page sets out what that looks like.

The four tests, in one place

A process is a good candidate when it passes all four. Two out of four means wait.

1. It repeats daily or weekly.
2. Most of its steps follow rules you could write down.
3. It costs real hours or real errors. Put a number on it with the process cost calculator.
4. You can measure the before and after.

If yours passes, the examples library shows what the automation would look like, and a free assessment gives you a written answer for your specific process.