An AI project is worth doing when it improves a specific, measurable piece of work by more than it costs to build, run and supervise, after you account for risk. That sounds obvious, yet many business cases skip the running costs, the review time and the downside. This framework keeps the numbers yours and the claims honest. It contains no benchmark figures on purpose, because the right ones come from your own baseline.
Step 1: Define the job, not the technology
Start from a task that already exists, such as answering routine enquiries, drafting first versions of reports or sorting incoming leads. Write down:
- Who does it today and how often.
- What a good result looks like.
- What goes wrong when it is done badly.
If you cannot describe the task without mentioning AI, you do not yet have a business problem. Our overview of AI automation for business operations lists where automation commonly fits.
Step 2: Measure the baseline
You cannot show improvement without a “before”. Record the current effort, error rate, delay and cost for the task, using your own records. A sample over a few weeks is better than a guess. If the baseline is unmeasured, the first output of the project may simply be measurement, which is a legitimate result.
Step 3: Count every cost
| Cost type | Examples to include |
|---|---|
| Build | Design, configuration, integration, testing |
| Run | Subscriptions, usage charges, hosting, monitoring |
| People | Staff time to review outputs, handle exceptions, train users |
| Change | Process updates, documentation, communication |
| Risk | Cost of an error, privacy handling, correcting mistakes |
Review time is the item most often forgotten. If every output needs checking, the saving is the difference between creating and checking, not the whole task.
Step 4: Capture benefits in two columns
Quantitative: time saved, reduced rework, faster response, additional capacity. Use your baseline, and state assumptions explicitly.
Qualitative: consistency, staff morale, better customer experience, lower dependence on one person. These matter but are harder to price. Record them separately instead of inventing a dollar figure.
Be careful with benefits that only exist if staff hours are actually redeployed. Saved time is only a saving if it is used or released.
Step 5: Assess risk using a recognised structure
NIST’s AI Risk Management Framework groups risk work into four functions: Govern, Map, Measure and Manage (NIST). The framework’s own materials describe Govern as building policies, roles and culture, Map as establishing purpose and context, Measure as testing and monitoring risks, and Manage as prioritising and responding (NIST AI RMF Core). It is voluntary and aimed at making AI systems more trustworthy.
Applied to a small project, that translates into questions:
- Govern: who is accountable for this system?
- Map: who could be affected, and what could go wrong?
- Measure: how will we test quality, and how often?
- Manage: what is the fallback, and who can switch it off?
If personal information is involved, add a privacy check under Australian law before you start. This is general information, not legal advice.
Step 6: Pilot small, then decide with rules
Set the decision rules before the pilot so you cannot move the goalposts:
- Scope: one task, one team, a fixed period.
- Success threshold: what improvement, against which baseline, justifies continuing.
- Stop condition: what result or incident ends the pilot.
- Review point: a scheduled date to decide, expand or stop.
Then compare results with the baseline and the full cost list. Our guide to building a digital marketing strategy sets the objectives to measure against, and our article on measuring creative campaign performance applies a similar logic to marketing, where the metric hierarchy matters as much as the number.
Common traps
- Counting the demo, not the daily reality. An impressive sample does not show performance on messy real inputs.
- Ignoring maintenance. Systems drift as data and processes change.
- Treating vendor claims as your results. Treat them as hypotheses.
- No exit plan. Know how you would revert to the manual process.
Where Radiance Impact fits
Radiance Impact works across creative production, AI and digital technology, and can help scope a small, measurable first workflow. See our About page or get in touch to talk it through.
Sources
- AI Risk Management Framework, National Institute of Standards and Technology
- AI RMF Core, NIST AI Resource Center
