AI Automation Opportunity Scorecard
Compare workflows before choosing an AI pilot. Score each criterion from 0 to 3, apply the hard-stop checks, then test whether the expected value still survives implementation and review costs.
Use one copy per workflow. Score the work as it operates today—not the process you hope to have. Ask the person who performs it and the person accountable for errors to score it together. A high total is a shortlist signal, not permission to automate.
Describe the workflow first
Trigger[what starts the work?]
Output[what must be produced or changed?]
Systems touched[source, destination, and system of record]
Current owner[person accountable today]
| Criterion | 0 | 1 | 2 | 3 | Score |
|---|---|---|---|---|---|
| Weekly volume | Under 10 | 10–49 | 50–199 | 200+ | [ ] |
| Minutes per case | Under 2 | 2–5 | 6–15 | 16+ | [ ] |
| Input consistency | Unstructured | Often varies | Mostly consistent | Fixed schema | [ ] |
| Acceptance rule | Subjective | Partly defined | Clear with exceptions | Fully testable | [ ] |
| Error cost | Critical | High | Moderate | Low / reversible | [ ] |
| System access | None | Manual export | Partial API | Stable API / webhook | [ ] |
| Exception owner | None | Team only | Role assigned | Person assigned | [ ] |
| Data sensitivity | Restricted | Special category | Personal data | Non-personal | [ ] |
Hard stops override the score
Do not approve a production pilot while any answer below is “no.” A workflow cannot compensate for an uncontrolled risk simply by happening more often.
- A named workflow owner can approve the scope and target.
- A named exception owner can review failures within an agreed time.
- The team can test outputs against a written acceptance rule.
- Access to required systems and data is authorised and technically available.
- Personal or restricted data has a documented basis, retention rule, and access boundary.
- The workflow has a safe manual fallback and a measurable rollback trigger.
Decision
19–24Production candidate. Scope monitoring and fallback before build.
13–18Run a two-week discovery. Resolve the lowest-scoring constraints.
0–12Do not automate yet. Fix ownership, inputs, or acceptance rules first.
Value case
Use measured inputs. Treat model output review, exception handling, licences, integration maintenance, and monitoring as costs—not free time.
Annual cases[weekly volume × working weeks]
Current annual effort[annual cases × minutes per case ÷ 60]
Expected hours removed[current effort × realistic automation rate]
Annual operating cost[tools + models + review + maintenance]
Net first-year value[hours removed × loaded hourly cost − build and operating cost]
Success measure after 60 days[cycle time, hours, error rate, cost, or revenue measure]
Production guardrails
Fill in before you start the build. These are your rollback and review conditions.
Workflow owner[fill in]
Exception owner[fill in]
Baseline hours or cost[fill in]
Target after 60 days[fill in]
Human-review threshold[fill in]
Maximum monthly model / tool spend[fill in]
Rollback trigger[fill in]
Personal-data basis and retention[fill in]
Built from the workflow checks used in Some Tech Work automation discovery. See AI automation delivery and pricing or the AI automation cost guide.