Ask which job to automate first and the answer is usually whichever one annoyed someone most recently. Irritation is a real signal, but it correlates badly with value — the most annoying task in a business is often small, and the most expensive one is often invisible because it's been absorbed into somebody's normal week.
Here's a ranking method that takes an afternoon, needs no software, and reliably surfaces a different answer than the annoyance list.
Step one: list the recurring work
Not projects — recurring work. Anything that happens weekly, monthly or per-customer and follows roughly the same shape each time. Ten to twenty items is normal. Include the ones nobody talks about: re-keying between systems, chasing information, assembling the same report, checking whether something happened.
Step two: score each one on four things
| Score | What you're measuring | How to get it |
|---|---|---|
| Cost | Annual hours × loaded hourly cost | Minutes per occurrence × frequency × 52 |
| Mechanical | How much needs no judgement | Rough percentage — 90% mechanical, 40% mechanical |
| Reachable | Whether the data is accessible | Does every system involved have an API or export? |
| Blast radius | What a mistake costs | Cheap and visible, or expensive and silent? |
Cost and mechanical you multiply. Reachable is a gate — if the data can't be got at, the score is zero regardless of how attractive the rest looks. Blast radius is a modifier: anything expensive-and-silent gets pushed down the list even if the numbers are good, because the automation will need review steps that eat the saving.
Rank by annual cost × mechanical percentage, then delete anything whose data you can't reach and anything where a silent error would be expensive. What's left at the top is almost always the right first job — and it's rarely the one people nominated.
Step three: sanity-check the top three
For each, answer two questions out loud. If this ran twice as often, would that be better? If not, you have a process problem, not an automation opportunity. And who acts on the output? If there's no name, the automation will produce something nobody owns.
The four archetypes, in the order they usually pay
1. Moving data between systems
Almost purely mechanical, usually high frequency, and errors are visible quickly. This is the highest-value automation in most businesses and the least discussed, because re-keying feels like part of the job rather than a task.
2. Assembling recurring reports
Also highly mechanical, and it carries a hidden delay cost: the business is always looking at last week. The one caveat is that automating a report nobody acts on saves the producer's time and nothing else — so fix the content while you're at it.
3. Noticing things
Watching for a pattern — accounts going quiet, stock cover falling, invoices ageing — is something software does perfectly and humans do only when they have spare attention, which is never. Cheap to build, and it surfaces money.
4. Drafting communication
High value, but it needs a human on the send button, which caps the saving. Still worth doing: writing the draft is most of the work, and having one ready is the difference between a chase happening and not happening.
What to leave alone
- Anything still changing. If the process will be different in two months, do it by hand until it settles.
- Genuine judgement. Pricing decisions, difficult conversations, exceptions. Automate the preparation, not the call.
- Rare tasks. Four times a year by hand is fine. Leave it and stop feeling guilty.
- Anything where a silent error is expensive. Payments, anything customer-facing without review, anything with a statutory format.
- Anything whose data you can't reach. Fix the source system first, or accept a manual step.
Worked example: scoring a real list
Here's the exercise done properly, using a wholesaler with about twenty staff. Six recurring tasks, scored on annual cost and how mechanical they are.
| Task | Annual hours | Annual cost | Mechanical | Score |
|---|---|---|---|---|
| Re-keying orders into the accounts package | 104 | £4,160 | 95% | 3,952 |
| Assembling the Monday sales report | 78 | £3,510 | 85% | 2,984 |
| Chasing overdue invoices | 52 | £2,340 | 60% | 1,404 |
| Checking stock cover before ordering | 39 | £1,560 | 80% | 1,248 |
| Following up quotes | 26 | £1,300 | 40% | 520 |
| Monthly commission calculation | 12 | £600 | 90% | 540 |
The interesting part is what this reorders. Everyone in that business nominated invoice chasing, because it's the one that generates complaints. It came third. The re-keying nobody mentioned — because it doesn't feel like a task, it feels like using a computer — came first by a wide margin.
Note also that the commission calculation scores low on total cost but is 90% mechanical, cheap to build, and produces a silent-error risk if it's wrong. Small jobs with high mechanical scores are often worth doing early simply because they're quick wins that build confidence.
The score is a starting point, not a verdict. Three things override it:
- Unreachable data — a high score is irrelevant if the system can't be read. Check before you plan.
- Expensive silent failure — anything where a wrong answer isn't obvious needs review steps that eat the saving.
- A process about to change — automating something that will be different in two months is waste, however high it scores.
What "mechanical" actually means
The mechanical percentage is the part of the estimate people get wrong, usually by being too generous.
Watch someone do the task and count the moments where they make a decision that isn't in any written rule. Choosing which nominal code an odd invoice belongs to. Noticing that a quantity looks improbable. Recognising that two customer records are the same person. Deciding this particular client shouldn't get the standard chase because of a conversation last week.
Each of those is judgement wearing data entry's clothes, and each one is a reason the task is less mechanical than it looks. That's not an argument against automating — it's an argument for automating the mechanical portion and routing the exceptions to a person, which is a different and much safer design.
A useful reframe: rather than asking "can this be automated", ask "what percentage of occurrences are completely routine?" If the answer is 90%, automate those and queue the rest. You capture nearly all the saving and lose none of the judgement.
Sequencing across a year
Most businesses can absorb three or four automations a year without it becoming disruptive. A sensible sequence:
- First: the highest-scoring mechanical flow, usually data moving between systems. Boring, visible, quick.
- Second: a report or an alert built on the data the first one made available. This is where the first build starts paying compound interest.
- Third: something that drafts communication for approval — chasing, follow-ups, nudges. Higher value, needs the trust the first two earned.
- Fourth: whatever the first three revealed, which is rarely what you'd have guessed at the start.
That last point matters. Almost every business finds that the third automation is something they couldn't have specified in January, because it only became obvious once the data was joined up.
Why the order matters more than the tooling
The first automation sets whether anyone trusts the second. Pick something mechanical, frequent, low-risk and measurable, and you get a visible win that funds the rest. Pick something ambitious and customer-facing, and the first mistake becomes the story — and the next proposal gets refused on the basis of that story rather than its own merits.
So the correct first job is usually a boring one. That's not timidity; it's how you buy permission to do the interesting ones.