Service / Assessment
Decide where AI belongs.
An AI opportunity assessment examines whether a particular task would benefit from artificial intelligence. Start with the work, not a product demonstration. The aim is a decision that explains the expected benefit, the practical constraints and the reasons to proceed, change direction or stop.
1. Describe the existing task.
Write down the trigger, inputs, actions and output. For an incoming supplier document, distinguish extracting the purchase order reference from deciding whether to authorise payment. These are different tasks with different consequences. Record who completes each action, which system holds the authoritative record and who resolves exceptions.
Use representative examples rather than only the cleanest documents. Include incomplete scans, ambiguous requests and cases that need judgement. Establish a baseline from your own records: time spent, correction work and missed deadlines. Keep assumptions separate from observations so the assessment does not mistake a hoped-for saving for an established benefit.
2. Compare simpler alternatives.
Generative AI produces new text or other content from a prompt. It can help with variable language, but it is not the default answer to every repetitive task. A form with mandatory fields may prevent missing information more reliably than extracting it afterwards. A database query may answer a structured question without generating prose.
Compare rules-based automation, ordinary search, optical character recognition and a language model against the same requirement. Rules are easier to inspect when conditions are stable. Language models handle variation but can produce plausible mistakes. Select the simplest approach that meets the need, including a combination of deterministic checks and AI-assisted drafting.
3. Check the information available.
Identify the sources the system would receive and the permissions attached to them. A shared folder is not necessarily a usable knowledge base: duplicated files, superseded policies and unclear ownership can undermine an otherwise capable assistant. Record document formats, update responsibilities and whether important facts exist only in someone’s email.
Separate personal information, confidential business information and material that may be used freely. For UK organisations, the UK GDPR and Data Protection Act 2018 are relevant when personal data is processed. Consult the ICO’s AI and data protection guidance; obtain qualified advice where the proposed use raises legal questions.
4. Define acceptable and unacceptable results.
State what a useful result looks like before selecting a model. A draft email may be acceptable only after staff review. A extracted reference may need an exact match against an existing record. A recommendation affecting a person’s employment or access to a service requires a different level of scrutiny from an internal summary.
Describe failure handling as well as success. Specify when the system should refuse, ask for clarification or return the task to a person. Do not use a single overall accuracy figure to conceal serious errors. Separate routine formatting mistakes from failures that expose information or cause an unauthorised action.
5. Estimate the complete operating effort.
Include integration, review, monitoring, maintenance and supplier management alongside model charges. A small API charge does not make a process inexpensive if every output needs extensive correction. Account for document preparation and for the effort of changing procedures when systems or policies change.
Hosted services reduce some infrastructure responsibilities but introduce supplier and data-location questions. An internally operated model offers a different level of control, while adding deployment and security work. Ask suppliers for written information on retention, subprocessors and contractual terms rather than assuming that a product label establishes suitability.
6. Agree a bounded next step.
A pilot is a limited implementation used to examine a defined question. Choose a task owner, permitted data, review method and stopping conditions. Keep the pilot away from irreversible actions until its behaviour can be assessed. Record which assumptions the pilot must resolve and what evidence would justify a wider rollout.
An assessment brief can specify a process map, alternatives comparison, data constraints, risk register and pilot proposal. Agree the deliverables before commissioning the work. If the evidence points towards better forms or clearer documentation rather than AI, retain that conclusion: improving the process is the purpose of the assessment.