Which processes to automate first: a scoring method
Automate first the process that is frequent, rule-based, fed by digital inputs and cheap to get wrong. Score each candidate from 1 to 5 on five criteria, add the points, and build the top one or two. The method takes an afternoon and prevents the usual mistake of starting with the most painful process instead of the most suitable one.
The five criteria
Each criterion counts the same in the basic version. If your company has one dominant constraint, such as an old ERP, weight that row double.
| Criterion | Score 1 | Score 5 |
|---|---|---|
| Frequency | A few times a month | Many times a day |
| Rule clarity | Depends on individual judgement | Staff follow the same steps every time |
| Input quality | Paper, phone calls, free-form chat | Structured forms, system records, consistent documents |
| Error tolerance | A mistake costs money or breaks a legal duty | A mistake is visible and easy to correct |
| System access | Legacy tool with no API or export | Modern apps with documented APIs |
How to run the scoring
- Ask each team lead for the three recurring tasks that consume the most hours. Ten to fifteen candidates is plenty.
- Write one line per candidate: trigger, steps, systems touched, monthly count.
- Score every candidate on the five criteria with the people who do the work, not only their managers.
- Add the scores. Anything at 20 or above is a strong first candidate. Between 14 and 19, look at what holds it back. Below 14, leave it for later.
- Estimate hours saved per month for the top three and choose the one with the best ratio of saving to effort.
A worked example
Take a wholesale company with three candidates. The scores below are illustrative and show how the method separates them.
| Process | Frequency | Rules | Inputs | Tolerance | Access | Total |
|---|---|---|---|---|---|---|
| Routing inbound support email to the right queue | 5 | 4 | 4 | 5 | 4 | 22 |
| Entering supplier invoices into the ERP | 4 | 4 | 3 | 2 | 3 | 16 |
| Preparing custom quotes for key accounts | 2 | 2 | 2 | 2 | 3 | 11 |
Reading the result
Email routing wins. It happens constantly, the categories are known, and a wrongly routed message is noticed and moved within minutes. An AI classification step fits well, and staff see a benefit in the first week.
Invoice entry is valuable but scores lower on tolerance and inputs. Build it second, with extraction checked against purchase orders and a person approving anything that does not match. Custom quotes stay manual for now, because the work relies on judgement and relationships. A drafting aid may help there later.
Where AI steps change the score
- Input quality rises. A model can turn free-text email or mixed PDF layouts into structured fields, which lifts a 2 to a 3 or 4.
- Rule clarity can rise. Sorting by intent or urgency no longer needs keyword lists that someone must maintain.
- Error tolerance does not change. If a mistake is expensive, it stays expensive, so keep human approval on those steps.
- System access does not change. No model fixes a tool without an interface. Plan an export, a database connection or a replacement.
Processes to leave alone for now
Some tasks score low for good reasons and should wait. Automating them early burns budget and goodwill.
- Processes that are about to change because of a reorganisation or a new system.
- Tasks done so rarely that nobody agrees on the correct steps.
- Decisions with legal weight where the reasoning must be a person's own.
- Anything whose only data source is a spreadsheet that one employee maintains by hand.
Common questions
How many processes should we automate in the first round?
One or two. A small first round lets you learn how your systems behave, set up monitoring and build trust with staff. After that, later workflows reuse the same connections and go faster.
Should we measure time saved?
Yes, but keep it simple. Record how long the task takes by hand for a week before the build, then count automated runs afterwards. Add error and rework counts if the process had quality problems.