AI audit checklist: find opportunities worth testing

What should an AI audit cover?

An AI audit should cover process volume, time and cost, data readiness, output quality, implementation effort, ongoing costs and ownership. Its output should be a prioritised decision about what to test, what to defer and what to leave alone.

1. Map a process before choosing a tool

Choose one recurring task and trace it from the initial request to the finished result. Record who does each step, which systems they use and where work waits or needs correcting.

• What starts the task, and what counts as a completed result?
• How often does it happen, and how much does volume vary?
• How many minutes are spent doing, checking and correcting the work?
• What happens when the result is late or wrong?

Use actual samples where possible. If the numbers are estimates, label them and validate them before making an investment decision.

2. Decide whether AI is necessary

Some problems need a clearer process or a straightforward connection between systems. A deterministic task, such as copying an approved field or sending a scheduled reminder, may be better served by conventional automation.

AI may be useful where the task involves interpreting variable documents, drafting text or finding relevant information. These tasks still need an appropriate way to check the result. Avoid making an uncertain model output the sole basis for a consequential decision.

3. Check the data and permissions

• Is the information accessible, current and sufficiently complete?
• Do you have the right to use it for the proposed purpose?
• Can access be limited to the people and systems that need it?
• What would the provider store, retain or use for training?
• Can you test with redacted or synthetic examples first?

Agree data handling before connecting a tool to live customer or business information. Requirements depend on the information and the systems involved.

4. Separate capacity from cash savings

Calculate time saved after human review and exception handling. Then identify whether that time translates into a genuine reduction in spending or simply creates capacity for other work.

For a cash-based payback calculation, divide the one-off implementation cost by the positive monthly cash saving after ongoing costs. If no cash spending is removed, do not present a cash payback period.

5. Prioritise benefits against effort and risk

For each opportunity, document the expected benefit, confidence in the estimate, dependencies, implementation cost, ongoing cost and impact of mistakes. A modest improvement in a well-understood process may be a better first project than a large but uncertain transformation.

Name an owner and define a measurable result. Useful measures can include processing time, rework, response quality or actual spend. Keep time savings and financial outcomes separate.

6. Run a bounded pilot

Test representative examples, including unusual and difficult cases. Compare the result with the current process and record both quality and total effort. Agree the conditions for stopping, revising or expanding the pilot before it starts.

• Who checks the output and handles exceptions?
• What result would justify wider use?
• How will failures be detected and corrected?
• Can the team return to the previous process?

What should you receive at the end?

A useful review gives you a shortlist of opportunities, a baseline, indicative benefits and costs, important assumptions and a recommended next step for each item. It should also explain where AI is unlikely to be worthwhile.

Turing&Co’s AI Opportunity Audit is a separate advisory service for businesses that want this assessment before committing to implementation.