Follow a real item through the business
Pick an invoice, enquiry, customer onboarding or report and follow it from arrival to completion. Record the people, systems, decisions and hand-offs. Include time waiting for information, not just time entering it. Ask the people doing the work where it regularly gets stuck.
Check whether the cause is missing information, an unclear owner or a disconnected system. Automating an unnecessary approval or a broken hand-off can make a poor process move faster without making it better. Simplifying the work may be the right first intervention.
Compare a small shortlist on practical criteria
Look at frequency, handling effort, rework, consistency of inputs and business importance. Then consider API access, data quality, process stability and how easily mistakes can be detected and reversed. Give preference to work with someone available to test and operate the result.
Use these questions to structure a discussion rather than manufacture a precise score from guesses. A frequent low-risk reporting task may be a better first pilot than a high-value financial decision with unclear authority and no reliable data.
- How often does this happen, and what is the current burden?
- Can we obtain representative inputs and verify a correct result?
- Who owns exceptions and can stop or reverse the workflow?
Decide whether AI is needed at all
Copying known fields, enforcing approvals and calculating totals are conventional software jobs. Reading varied emails, classifying text or preparing a draft may justify AI. Often the useful design combines both: AI interprets an input, software validates it and a person approves the consequential action.
Illustrative candidates include preparing CRM updates from enquiries or extracting document details into a review queue. These are possible designs, not promises of unattended accuracy. Do not begin by giving a model unrestricted access to customer records or payment actions.
Define success before the pilot starts
Record current volume, handling time, waiting time and common errors. Agree what counts as an accepted result, how much human review remains and what failure rate is tolerable for that task. Include the team’s experience: a fast workflow that creates extra checking elsewhere is not a clear win.
Use representative examples, including awkward inputs and known exceptions. Test access failures, repeated events and unavailable systems. Keep a manual route while the pilot is evaluated and name the person who decides whether it can expand.
Choose the next engagement to match the uncertainty
If you cannot yet identify the strongest opportunity, the AI & Automation Opportunity Review produces a written roadmap with priorities and indicative costs. The existing assessment can help structure your initial thinking; it is not a substitute for inspecting the real process.
If a workflow is already understood, discuss a 10-Day Automation Sprint or scoped implementation. Wider work may suit a 90-Day AI Transformation. Confirm dependencies and scope first: neither a programme name nor a promising pilot proves that every process should be automated.