Most automation programmes do not fail on technology. They fail on sequencing. A team picks the process that annoys them most, discovers it has forty exceptions and three unwritten rules, spends four months on it, and never gets to the second one.
Start where the rules are stable
The best first automation has four properties: high volume, stable rules, a measurable before-and-after, and a low cost of being wrong. Invoice and document processing usually scores well on all four. The documents arrive constantly, the fields you need are consistent, the time saved is countable, and a misread invoice is caught downstream rather than causing harm immediately.
Compare that with something like pricing approvals — lower volume, judgement-heavy, exception-driven, and expensive to get wrong. It is a worse first project even though it may be a bigger prize eventually.
Count the manual touches, not the hours
When we scope a process, the number we care about is manual touches: how many times a human has to pick something up, read it, and put it somewhere else. Hours are a downstream consequence and much harder to attribute honestly.
Counting touches also tells you where to stop. A pipeline that removes nine of ten touches is a success. Chasing the tenth — usually a genuine judgement call — often costs more than it saves and makes the system brittle.
Design the exception path first
The most common failure we are called in to fix is automation with no exception handling. Everything works until something unexpected arrives, and then the pipeline either stops silently or, worse, processes it wrongly and keeps going.
Decide up front what happens when the system is unsure: who is notified, what they see, and how the item re-enters the flow once resolved. Silent failure is the single worst outcome in automation, because it destroys trust in every number the system produces afterwards.
Instrument before you launch
If a process cannot be measured, it is not finished. Volume processed, exception rate, time from arrival to resolution, and where the remaining manual touches sit — all of it should exist from day one. Otherwise the inevitable question six months later, of whether this was worth doing, has no answer.
Common questions
What is the best first process to automate?
One with high volume, stable rules, a measurable before-and-after and a low cost of being wrong. Document and invoice processing usually scores well on all four. Avoid starting with the process that annoys people most — it is usually annoying because it is full of exceptions.
How long before AI workflow automation pays back?
It depends far more on the process you pick than on the technology. A high-volume process with a countable baseline can show a return within a quarter. One with forty exceptions may never pay back, which is why the selection step matters more than the build.
Do we need a data warehouse first?
Usually not for a first automation. You need reliable access to the documents or records involved, not a full analytics platform. Building a warehouse before automating anything is a common way to spend a year without shipping something people use.
