1. Describe the work in operational terms
Begin with the input, the expected output and the person accountable for the result. “Improve customer service” is too broad. Classifying incoming enquiries by subject, preparing a draft response from approved guidance or extracting a reference from a document gives an assessment something concrete to examine.
Our audit service focuses on these boundaries. A proposed scope can include process mapping, a source inventory, an options comparison and a decision brief. It should also record exclusions: which decisions remain with staff, which systems cannot be accessed and which data must stay outside a trial. Agree these deliverables before work starts.
2. Compare AI with a simpler baseline
A baseline is the existing or simpler method against which a new approach is compared. For fixed categories and predictable inputs, a rules engine may be easier to maintain than a language model. Regular expressions can identify a consistently formatted reference; a spreadsheet formula can apply an explicit calculation.
Language models are more relevant when input varies in wording or the task involves drafting and summarisation. They can also produce plausible but incorrect output. Compare the entire process, including review and exception handling, rather than comparing a fluent draft with a blank screen. An approach that generates more checking work may not improve the task overall.
3. Inspect the information and permissions
List the documents and records needed to perform the task. For a policy assistant, this might include approved SharePoint pages and their access groups. For document extraction, include both readable PDFs and scans with poor text recognition. Do not treat every file in a shared drive as an approved source.
Check ownership, confidentiality and retention requirements before copying material into a tool. For UK personal data, the ICO’s AI guidance is a useful starting point. A discovery exercise does not remove the need to establish a lawful basis or assess a supplier’s processing terms.
4. Define evidence for the decision
An evaluation set is a collection of inputs used to assess how a system behaves. Select examples across ordinary work and difficult cases: missing attachments, contradictory instructions, outdated policies and inputs outside the intended scope. Keep some examples separate from those used to adjust the system, so familiar inputs do not disguise weaknesses.
Specify acceptable behaviour for each task. Extraction requires correct fields and a clear signal when information is missing. Summarisation needs faithful coverage without invented statements. Drafting needs a reviewer who can inspect sources. Set acceptance conditions with the process owner; a generic accuracy score cannot describe every consequence of an error.
5. Examine the full operating cost
Model charges are only one part of the cost. Include source preparation, integration, access management, staff review, monitoring and changes to the surrounding process. Compare managed model APIs with self-hosted models against confidentiality, operational capability and support needs. Self-hosting changes responsibilities; it does not automatically establish privacy or reliability.
Ask who handles supplier outages, failed jobs and changes in model behaviour. Identify whether the process can continue manually. If the only person able to maintain the prototype is unavailable, the implementation has an ownership problem regardless of how persuasive its demonstrations appear.
6. Choose a bounded next step
The decision brief should recommend proceeding, changing the scope, gathering more evidence or not using AI for the task. A limited pilot is appropriate when uncertainty can be resolved safely. Keep its permissions narrow and its results separate from production decisions until the agreed checks have been completed.
For an enquiry, bring a short process description, the source locations and the main failure you need to avoid. Use anonymised examples initially. We can then discuss the audit boundary and proposed deliverables without asking you to send an unrestricted collection of internal records.
Discuss an AI audit