Understand the work before you automate it
By Sascha Cutura

Process understanding has become the quiet dividing line between AI projects that deliver value and those that stall.
Many organisations understand parts of their work, or how it was designed to run. Fewer understand how it behaves day to day. Once AI is introduced, especially when a system is expected to take action, that gap becomes hard to ignore. Automation accelerates whatever is already there.
The work is held together by judgement
Most operations rely heavily on people to keep processes moving. People know which steps can be skipped, which exception is safe, when to escalate and when to wait. Much of this knowledge is informal and rarely written down, and it is often what makes the system work.
When automation is added without that reality in view, those invisible decisions disappear. What looked tidy on a process map turns out to be fragile in practice. The technology behaves as expected. The process does not.
This is why process visibility belongs at the start of an implementation, not as a cleanup exercise afterwards. If a system is going to act inside a process, it needs to see where work starts, how it moves, where it pauses, and what happens when things do not go to plan, which is most of the time in real operations.
A simple test helps. If you cannot explain the process in plain terms, including the exceptions, you are not ready to automate it with AI.
Look at the inputs as well as the steps
Even a well-described process will struggle if the inputs are messier than the plan assumed.
Most automation projects are designed around a tidy picture. A document arrives, the system reads it, the data goes where it should. Then the real items turn up. Invoices as PDFs, scans and photos taken on a phone. Orders in the body of an email, in an attachment, and occasionally in both, with different quantities. Forms that changed layout last spring. Spreadsheets where column F means something different depending on who filled it in.
In our experience, this is where implementations slow down, overrun or quietly fail. The process logic is often sound. Nobody looked properly at what was actually coming in.
Before automating anything that handles documents or data, look at a real sample from a normal month, including the awkward ones. What types of item are there, really? A team will say "we get invoices" and the sample will show several distinct kinds, each needing different handling. How consistent is each type? What is missing, wrong or ambiguous? Those cases generate exceptions, and exceptions are where the cost of an automated process actually sits.
Once you understand the mix, classifying inputs becomes a step in its own right. Standard items can go straight through. Known variations go to a rule built for them. Genuine unknowns go to a person, with the context they need. Modern capture and AI tools are much better at this than the template systems of a few years ago. They still work best when someone has decided what the categories are, what good looks like, and what should happen when the tool is unsure.
Fix what you can before the system goes live
Sorting the inputs also shows you where to stop the mess at source. You cannot make every supplier use the same layout. A surprising amount can still be prevented: a web form that replaces free-text emails, a required purchase order field, one shared definition of a customer, a short request to your largest senders to use one address and one format. None of these are technology projects, and each one removes a category of exception before the automation ever sees it.
Across more than 700 process optimisations in 90+ companies over the last 17 years, the most common finding has been work that no longer needs to exist. A weekly report compiled by hand and read by nobody. A second approval kept long after the risk it covered disappeared. Details typed from one system into another because an integration was never finished. Remove those steps, and the case for AI usually narrows to the parts of the process where judgement, volume or speed genuinely justify it.
Leaders sometimes treat this work as slow. In practice it is what makes implementation possible. When teams can explain how a process works, who owns each step, where decisions sit, and how exceptions are handled, automation is easier to design and safer to deploy. Without that clarity, AI simply magnifies the confusion that was already there.



