How to tell whether a process is actually automatable: a 10-point test
A scoring test for deciding whether a business process is worth automating — volume, variance, data access, judgement, and the cost of being wrong — usable in an afternoon without a vendor.
By Justice Annan, Founder"Can we automate this?" is almost always the wrong question, because the answer is almost always yes. Anything can be automated badly. The question worth answering is whether this process, at your volumes, with your systems, will still be worth the maintenance in a year.
What follows is the test we run before quoting any build. It takes an afternoon, needs no vendor, and disqualifies more candidates than it approves — which is the point.
How to score it
Ten criteria, each scored 0 to 3. Score from observation, not from how the process is supposed to work. Where you are guessing, write "guess" next to the number — a score built mostly of guesses is itself the finding.
The ten criteria
1. Volume
How many cases per week? Automation economics are driven almost entirely by repetition. Under roughly 50 cases a week, the build rarely pays for its own maintenance unless each case is expensive or slow. Score 3 for hundreds of cases a week, 0 for a handful.
2. Time per case
Measure it with a stopwatch on real cases rather than asking. Self-reported estimates are usually low, because people forget the chasing and the double-checking. Multiply by volume to get the weekly hours actually at stake.
3. Variance
Out of twenty consecutive real cases, how many followed the same path? Sixteen or more is a strong candidate. Fewer than ten and you are not looking at one process, you are looking at several wearing the same name — and the honest first project is mapping, not building.
4. Judgement content
Split the steps into data movement, rule application, and genuine judgement. Data movement is the cheapest and most reliable thing to automate. Rules are fine when they are written down. Judgement is where language models earn their place — and where you need a confidence threshold and a human path. A process that is 80% data movement is a far better first project than one that is 80% judgement, regardless of how interesting the judgement is.
5. Data access
Can you read and write the data through an API, a database, or a file drop? If the only route is clicking through somebody else's user interface, expect a brittle automation and a maintenance bill that never stops. Deloitte's automation survey has repeatedly found integration difficulty among the top barriers to scaling — 62% of respondents in its 2022 study.
6. Input quality
Structured feeds score 3. Consistent PDFs from a handful of senders score 2. Free-text email and photographed documents score 1. Handwriting and scans of faxes score 0 — still doable, but price it as a document-intelligence project, not a workflow project.
7. Cost of being wrong
What happens if one case is processed incorrectly and nobody notices for a week? A mildly annoyed colleague is a 3. A regulatory filing, a payment, or a customer-facing commitment is a 0 or 1 — not unautomatable, but it needs approval gates, and those change the economics. Be honest here before, not after.
8. A named owner
One person whose weekly numbers improve if this works, who will look at the exception queue, and who can approve a scope decision without convening a committee. No owner, no project. This is the single most reliable predictor of whether an automation is still running six months later.
9. Stability of the surrounding process
Is this workflow being renegotiated, re-platformed, or reorganised in the next two quarters? Automating something a month before its underlying system is replaced is a way to spend money and learn nothing. Score 3 for "unchanged for two years and no plans".
10. Measurability
Can you state today's baseline — cases, minutes, error rate, cycle time — in numbers? If not, you can still automate, but you will not be able to prove it worked, which means you will not get funded for the second one.
Reading the score
| Total (out of 30) | Read | What to do |
|---|---|---|
| 24–30 | Strong candidate | Baseline it and scope a fixed-price build. |
| 18–23 | Viable with conditions | Fix the low-scoring criterion first — usually access, ownership, or variance. |
| 12–17 | Not yet | The process needs standardising or documenting before automation is the bottleneck. |
| Under 12 | Decline | Automating it would industrialise a mess. Spend the money on the process itself. |
Three criteria override the total. A score of 0 on data access, 0 on a named owner, or 0 on measurability disqualifies the candidate regardless of how high the rest scores. Each one guarantees a system that either breaks, drifts into disuse, or can never be proven to have worked.
Four ways this test gets misread
- Scoring intent instead of reality. The documented process is not the process. Score what the twenty cases actually did.
- Treating judgement as a bonus. The presence of interesting judgement work makes a process appealing to automate and harder to ship. It is a cost, not a feature.
- Ignoring the exception rate. A workflow that hands 30% of cases to a human is still a large win at high volume and a nuisance at low volume. Volume decides.
- Forgetting month six. Every automation has an owner, a maintainer, and a monitoring cost. If nobody has budgeted for that, the score is academic.
If your best candidate scores badly
That is a useful outcome, not a dead end. A low score almost always points at one specific, fixable thing: an undocumented process, a system with no API, or an absent owner. Fixing that is cheaper than a failed build, and it is the same work you would have had to do anyway — just before the invoice rather than after it.
Questions we get on this
- What volume is too low to automate?
- As a rough rule, fewer than 50 cases a week rarely justifies a bespoke build unless each case takes a long time or an error is expensive. Below that threshold, better templates, a shared inbox rule, or an off-the-shelf tool usually beats a custom automation on total cost.
- Does the test change for AI agents versus classic automation?
- The criteria are the same; the weights shift. Language models widen what counts as low-variance work, because they tolerate messy inputs. They also raise the importance of the cost-of-being-wrong and measurability criteria, since their failures are probabilistic rather than deterministic.
- Who should score it?
- Whoever actually does the work, with someone who has seen automations fail. Scoring done purely by managers tends to overestimate consistency and underestimate time per case.
- Can a process score well and still be the wrong first project?
- Yes. If it sits in a team that has no appetite for change, or depends on a system your IT function is about to replace, a technically ideal candidate is a political dead end. Pick the second-best candidate with a willing owner.
Sources
- 1Deloitte — Automation with intelligence, 2022 survey results
62% cite difficulty integrating solutions as a barrier to scaling automation.
- 2Deloitte — Global intelligent automation study (press release)
Process fragmentation and IT readiness as the top two barriers; minority report mature process definitions.
- 3Parseur / QuestionPro — Manual data entry survey of 500 US professionals, July 2025
Baseline figures for time spent moving data by hand.
Justice spent 12 years running operations and service desks for two commercial banks, 330,000 sq ft of commercial property and HOA communities before founding SolisAI.
Connect on LinkedInWant this run against one of your own processes? We will score it, show you the arithmetic, and tell you if the answer is no.
Book a free process audit