Ask a room of founders what they want to automate and the answers cluster around the interesting end: an agent that handles customer conversations, something that writes proposals, a system that reads contracts. These are genuinely impressive, and they are almost always the wrong place to start.
The pattern is consistent. Impressive use cases are low-frequency and high-judgment. Low frequency means the time saved is small. High judgment means the build is hard. Together they produce the longest project with the weakest payback, positioned as the first one — which is how automation programmes lose credibility before they produce anything.
The money is in the unglamorous end. The process someone does forty times a week that nobody mentions in meetings because it is too boring to complain about.
Here is a way to find yours.
The Automation Priority Score
Seven factors. Score each 1 to 5, apply the weight, total out of 100. Higher means automate sooner.
| Factor | What you are measuring | Weight |
|---|---|---|
| Frequency | How often the process runs | ×5 |
| Human time | Minutes consumed per run | ×5 |
| Error cost | What a mistake costs when it happens | ×4 |
| Predictability | How well rules describe the cases | ×3 |
| Integration readiness | Whether systems expose the needed data | ×2 |
| AI necessity | Whether reasoning is genuinely required | ×1 |
| Payback clarity | Whether you can name the metric it moves | ×0 or reset |
Scoring each factor
Frequency (×5). 1 = monthly or less. 3 = a few times weekly. 5 = many times daily. Weighted heaviest because automation value is per-run and compounds with volume. A process run twice a month cannot pay back a build no matter how painful each run is.
Human time (×5). 1 = under two minutes. 3 = fifteen to thirty. 5 = over an hour. Frequency and time multiply into total hours, which is the actual saving.
Error cost (×4). 1 = trivially fixed. 3 = costs a customer interaction or rework. 5 = costs money, a customer, or compliance standing. Where this scores high, automation buys consistency even when hours saved are modest.
Predictability (×3). 1 = every case is different. 3 = rules cover most, exceptions are common. 5 = fully rule-describable. Low scores here do not disqualify, but they push the build from deterministic workflow toward an agent, which changes cost — the boundary is covered in AI agent vs workflow automation.
Integration readiness (×2). 1 = data is in PDFs, email bodies, or someone's head. 3 = systems have APIs but need work. 5 = clean APIs on both ends. This is a cost multiplier rather than a value driver, which is why it carries less weight but still matters.
AI necessity (×1). 1 = pure rules. 5 = genuine interpretation required. Deliberately weighted lowest. Needing AI is not a reason to prioritise something; it is a fact about how it will be built.
Payback clarity — a gate, not a score. Can you name the specific number this automation should move? Hours per week, error rate, turnaround time, cost per transaction. If you cannot, the process does not enter the ranking at all, regardless of everything else. Automations without a named metric are the ones quietly switched off within a year, because nobody can defend them at budget time.
Reading the total
Above 70 — automate now. High frequency, real hours, meaningful error cost. These are the boring processes that fund everything else.
50 to 70 — automate in the next wave. Usually genuine, often waiting on integration readiness. Sometimes a smaller intervention resolves most of the pain.
40 to 50 — hold. Revisit if volume grows or systems improve. Building here works but pays back slowly.
Below 40 — do not automate. The process is too rare, too variable, or too cheap to do manually. The right move is usually to simplify or eliminate it rather than encode it.
That last band matters more than it appears. A meaningful share of processes people want to automate should be deleted instead. Automating a report nobody reads produces a faster unread report.
What the ranking usually reveals
Run this across ten to fifteen processes and three things tend to surface.
The top of the list is unglamorous. Order status lookups, data re-entry between two systems that do not talk, invoice matching, delivery exception follow-ups. Nobody demos these. They are where the hours are.
The interesting use case scores mid-table. The agent everyone wanted usually lands at 45-60 — real, worth doing, not first. Reframing it as "next year, after the boring ones fund it" is easier once the number is on the page.
Something scores high and needs no AI. Frequently the highest-scoring process is a deterministic workflow that can be built quickly and cheaply. This is the framework's most useful output, and the reason AI necessity is weighted lowest — separating what to do from how to build it prevents buying reasoning for a job that needs a rule.
Sequencing after the ranking
Three rules that matter more than the scores.
One at a time, properly. Most SMEs succeed with three to five automations in year one and fail attempting ten. Each requires workflow change, team retraining and a monitoring habit — and organisational attention, not build capacity, is the binding constraint.
Fix before you encode. If the process has unclear ownership, inconsistent steps, or exceptions handled by improvisation, automating it locks the confusion in at machine speed. Map it, agree it, then build.
Capture the baseline first. Runs per week, minutes per run, error rate, cost per error — recorded before the build. Without it you cannot prove the outcome, and the next budget conversation becomes opinion against opinion.
A worked pass
Here is how a scoring pass looks across four processes. The numbers are illustrative, not a client case:
- Daily courier exception follow-up — 40+ runs weekly, 6 minutes each, customer-visible errors, largely rule-describable, APIs available. Score 78. Goes first, as a deterministic workflow rather than an agent.
- Marketplace-to-ERP order re-entry — 25 runs weekly, 12 minutes each, expensive errors, fully predictable, clean APIs. Score 74. Second.
- Inbound supplier email processing — 15 runs weekly, 20 minutes each, moderate error cost, genuinely unpredictable formats, data trapped in attachments. Score 58. Third, and the one needing an agent.
- Quarterly vendor performance report — 4 runs yearly, 3 hours each, low error cost. Score 31. Do not automate it. Simplify it instead.
Two of the four are deterministic workflows. One needs reasoning. One should not be automated at all. That distribution is typical.
Running it this week
Pick ten processes people complain about. Score each on the seven factors. Apply the payback gate — anything without a nameable metric leaves the list.
Take the top two. For each, capture the baseline before doing anything else.
Then decide the build shape, not the vendor: deterministic workflow or agent. That determines cost more than any other choice, and the ranges sit in our AI automation cost breakdown.
If you would rather have the scoring run against your actual operation — with the baselines measured rather than estimated, and an honest view of which processes should be simplified instead of automated — book an automation diagnostic. We come back with the ranked list, the build shape for the top candidates, and the ones we would tell you not to touch.




