When to automate a process — and when automation makes it worse

The four gates

A process should clear all four. Three out of four is the profile of an automation that gets built, half-trusted, and quietly worked around within a quarter.

Worth it

Frequency times time saved exceeds build plus ongoing maintenance. Maintenance is the term people forget; assume a recurring cost, not a one-off.

Ready

The process is stable and agreed. Two people run it the same way, and every decision inside it can be written as a rule.

Feasible

The systems expose the data reliably, and neither the tools nor the process is mid-change.

Safe

Failure is reversible, detectable within a day, and owned by a named person after go-live.

The automation readiness score turns these into ten questions and a recommendation band.

The arithmetic, done honestly

Annual saving is frequency per year times minutes saved per run, converted to hours, times a loaded hourly cost. Against it: build hours plus an annual maintenance allowance — assume something between a tenth and a fifth of the build, more if the process touches systems you do not control.

When manual is the right answer

• The rules are still being argued about. Automating a disputed rule turns a disagreement into infrastructure.
• The volume is genuinely low. A weekly ten-minute task is not a project; it is ten minutes.
• The decision needs context. Anything a customer would want a human to weigh should stay with a human, with the workflow preparing the context.
• The process is about to change. Automating during a redesign guarantees rework.
• Nobody will own it. Unowned automation degrades silently and is discovered during an incident.

The middle path most teams miss

Between manual and fully automated there is a large, underused middle: automate the trigger, the context assembly and the record-keeping, and leave the decision with a person. You capture most of the time saving with a fraction of the risk, and the workflow becomes a helper rather than a hazard.

1. Detect

   The system notices the condition — a form arrives, an SLA is about to breach, a milestone slips.

2. Assemble

   It gathers the context a person needs and attaches it to a task, in one place.

3. Decide

   A human makes the judgement call, quickly, because nothing had to be looked up.

4. Record

   The outcome is written back in structured form so reporting stays clean.

Sequencing several candidates

When five processes qualify at once, build in this order rather than by enthusiasm:

1. Highest frequency, lowest judgement. Fast to build, felt immediately, and it earns the credibility for the harder work.
2. Anything with a customer-visible clock. Response times and acknowledgements protect the relationship while the rest is being fixed.
3. Anything that currently depends on one person. Automation here removes a single point of failure, which is worth more than the hours saved.
4. Reporting assembly. Usually a large, invisible cost, and the output improves decision speed for everyone.
5. Everything else. By this point the register, the alerting habit and the ownership model exist, so the marginal build is cheap.

Related reading

The full discipline is business automation; the tooling trade-off is in n8n versus Zapier.