The usual selection method is to automate whatever the loudest department complains about. That is a reasonable proxy for pain, but a poor proxy for payback — the loudest process is often loud precisely because it is full of judgement calls, which is what makes it expensive to automate.

Here is the profile that actually predicts a fast return, followed by the six process shapes that fit it and the diagnostic questions we use during discovery.

The profile

Processes that pay back fastest share five characteristics:

• High volume — several hundred items a month at minimum. Below that, the build cost rarely amortises regardless of how annoying the work is.
• Text-shaped — the inputs are documents, emails or forms rather than physical objects or conversations.
• Rule-dominant — most decisions follow rules that could be written down, even if nobody has written them down.
• Expensive labour — currently done by someone whose time has a high opportunity cost.
• Tolerant of escalation — a low-confidence item can be routed to a human without breaking anything.

That last one is the filter people skip. A process where every item must be resolved immediately, with no escalation path, is a poor first candidate no matter how well it scores on the others.

The six shapes, ranked by payback

1. Invoice and purchase order matching — 3 to 5 months

Consistently the fastest payback we see. High volume, highly structured, rule-heavy, and the exception path already exists because someone already handles mismatches manually. Extraction quality on invoices is now good enough that the constraint is your approval workflow rather than the model.

Diagnostic: does anyone in finance spend more than a day a week keying or matching supplier documents?

2. Support and service ticket triage — 3 to 6 months

Classification, prioritisation and routing on inbound tickets. Payback is fast because volume is high and misrouting is expensive in a way that is easy to measure — every reassignment is a delay you can already see in your reporting.

Diagnostic: what percentage of tickets get reassigned at least once? Above ten percent, there is real value here.

3. Document review and extraction — 4 to 6 months

Contracts, claims, clinical records, due diligence packs. Payback depends heavily on who does the work today; when it is fee-earners or clinicians, the numbers become compelling quickly. Insist on citation back to source — a summary a reviewer cannot verify against the original creates a new review task rather than removing one.

Diagnostic: is anyone billing at a senior rate for first-pass reading?

4. Customer and patient intake — 4 to 7 months

Reading inbound requests, extracting fields, checking completeness, creating the record. The gain is often less about labour than about turnaround time and error rate — and in healthcare, completeness at intake is what determines whether something gets returned by a payer three days later.

Diagnostic: what proportion of intake items come back incomplete? Above ten percent, the automation pays for itself on rework alone.

5. Compliance evidence collection — 6 to 9 months

Slower to pay back in year one, but the calculation changes across audit cycles. Access reviews, change records, vulnerability remediation evidence — collected continuously rather than assembled in a panic each cycle. Year two typically costs a fraction of year one, and the benefit recurs indefinitely.

Diagnostic: how many person-days went into your last audit evidence pack?

6. Vendor and supplier onboarding — 6 to 10 months

Lower volume, so payback is slower, but the risk reduction is meaningful. Document collection, insurance verification, sanctions screening and record creation, with an audit trail that a regulator will accept.

Diagnostic: how long from first contact to a supplier being able to invoice you? If it is measured in weeks, there is a case.

What does not work yet

Being clear about this is more useful than another list of opportunities.

• Anything requiring genuine negotiation. Judgement under conflicting incentives is not a solved problem, and pretending otherwise damages trust in the whole programme.
• Low-volume, high-consequence decisions. If it happens twice a month and being wrong is catastrophic, a human should do it. The build cost will never amortise anyway.
• Processes nobody can describe. If three experienced people describe the process three different ways, you have a standardisation problem first. Automating it will encode one person’s version as truth.
• Work that is mostly chasing people. Partially automatable, but the bottleneck is usually the other party’s responsiveness, not your effort. Expect modest gains.

How to choose

Score your candidates on the five profile characteristics, then apply one further test: can you measure it today? If you cannot state the current cycle time, touch count and error rate, you will not be able to prove the automation worked — and an unproven first automation makes the second one much harder to fund.

Pick the highest-scoring process that you can already measure. Build that one properly, with the identity, logging and ownership architecture that everything afterwards will inherit. Then use the measured result to fund the next three.