Count first, automate second
Every automation project that fails starts with a tool. Every one that works starts with a list of steps and the hours each one eats per month.
Spend a week counting. Ask the team to note what they did by hand and how long it took. The result is almost never what management expected.
Do not automate a broken process
Encoding a bad process makes it faster and much harder to change. If three people approve something for historical reasons, delete the approvals first.
The cheapest automation is deletion. The second cheapest is doing the step once instead of twice.
Start where data is retyped
The highest-value integration is nearly always the place where someone reads a value from one screen and types it into another.
That work is slow, it is dull, and it is where mistakes enter the business. Connecting the two systems removes the hours and the error class at once.
- Orders into accounting.
- Enquiries into the CRM.
- Anything that currently lives in a spreadsheet somebody rebuilds every morning.
Where AI belongs, and where it does not
AI earns its place where the input is messy and the output gets checked: sorting incoming mail, pulling fields out of an invoice, drafting a reply someone approves before it sends.
It does not belong anywhere that a confident wrong answer costs real money and nobody looks. That is not a limitation waiting for a better model. It is how the technology behaves.
- Ask any vendor for the cost per user per month at your real volume.
- Ask what happens when the model is wrong. A vendor who cannot answer both has not shipped this before.
Keep a person in the loop where it counts
Money, contracts and anything public keep a human checkpoint. The automation still saves the time, because reviewing a draft is far faster than writing one.
Stop at the right time
Work down the ranked list and stop when the next item costs more to automate and maintain than it saves. There is always a tail of rare, fiddly cases that should stay manual forever.
Trying to automate that tail is how automation projects lose their budget and their sponsor.
Measure it afterwards
Record the hours before and after. Without that, nobody can tell whether the project worked, and the next one is harder to fund.