Automation fails in predictable ways. Not because the model is not clever enough, but because nobody wrote down what the work actually is before handing it to a machine. These are the five questions we ask on the first call. You can answer them without us, and if the answers are uncomfortable, that is the finding.
1. Who does this today, and what do they do when it goes wrong?
The happy path is the easy part. What matters is the exception: the invoice with no purchase order, the customer who replies in a different language, the file that arrives as a photograph of a screen. Sit with the person who handles those and write down what they do. That list is the specification. If nobody can produce it, the process is not understood well enough to automate yet.
2. How often is it wrong now, and how would you know?
Every manual process has an error rate. Almost nobody measures it, which makes the automated version impossible to judge: any number looks bad next to an imagined perfect human. Spend a week counting. You need a baseline you can be held to, and you will usually find the manual rate is higher than anyone guessed.
3. What does the process cost now, in hours rather than money?
Hours are honest. Money invites arguments about overheads. Count the hours the work takes across a month, including the parts nobody logs: the chasing, the re-keying, the meeting where three people reconcile two spreadsheets. If the honest answer is four hours a month, automate something else first.
4. What happens if the machine is confident and wrong?
This is the question that decides the architecture. Ask what the worst plausible mistake costs, and who finds out. Then set the guardrails from that answer rather than from the model's capability.
- If a mistake is cheap and visible, let it run and log everything.
- If a mistake is expensive but rare, put a person in front of the exceptions only.
- If a mistake is expensive and quiet, do not automate the decision. Automate the gathering, and leave the judgement where it is.
5. Who owns it on the Monday after we leave?
An automation with no owner degrades. Rules go stale, a supplier changes a form, a threshold that made sense in March does not in November. Name the person before you start. If no name is forthcoming, that tells you what the organisation thinks this work is worth.
The best system is the one nobody has to think about.
Two weeks of asking these questions properly is the cheapest part of any automation project, and the part that decides whether the rest is worth doing. It is why our discovery sprint ends with a plan and a decision rather than a commitment.