Automation
Automate the process you actually have
Automation projects fail when they encode the idealised process rather than the real one. A practical sequence for getting it right.
The most common way an automation project fails is not technical. It is that the team automated the process as described in the kickoff meeting, and the business runs on a different process entirely — one full of exceptions, judgement calls and workarounds that nobody thought to mention because they are invisible from the inside.
The documented process is a summary, not a spec
Ask how quoting works and you will hear a clean four-step sequence. Watch someone actually produce a quote and you will see them check a spreadsheet that is not in the four steps, apply a discount rule that exists only in their head, and email a colleague about an edge case. All three of those are load-bearing. Automate the clean version and the system will be wrong for a meaningful share of real cases on day one.
Start where the exceptions are rare
Not every process repays automation. The ones that pay back fastest are high-volume, low-judgement and well-bounded: routing enquiries, sending reminders, syncing records between systems, generating standard documents. The ones that punish you are low-volume and judgement-heavy, where the exception rate is high enough that people end up checking every output anyway — at which point you have added a step rather than removed one.
This applies with particular force to AI-assisted steps. A model that is right 90% of the time is genuinely useful when a human reviews the output as part of an existing workflow, and actively harmful when it is wired directly into a process nobody inspects.
A sequence that works
- Observe the real process, including the exceptions and the workarounds
- Quantify it: how often does each path actually occur?
- Automate the dominant path first, and route exceptions to a person
- Instrument everything, so you can see how often the exception path fires
- Expand coverage only where the data says it is safe
The goal is not to remove people from the process. It is to stop spending human attention on the parts that never needed it, so it is available for the parts that do.
More insights
- AI or ordinary automation? Match the tool to the problemMost tasks sold as AI problems are better solved by deterministic automation. How to tell which is which before you commit budget.
- When to build custom software, and when to just buy itBuilding costs more than the quote and buying costs more than the subscription. A practical test for telling which one your problem actually needs.
Find out which of these applies to you
30 minutes, no pitch. If we’re not the right fit, we’ll tell you that too.