Key takeaways
- Start from frequency and pain, not from technology
- Score candidates on volume, rule clarity, error cost, and access
- Pilot one process before building a program
Start with a list, not a tool
Most teams begin by asking what AI can do. A better starting point is a plain list of repetitive tasks: who does it, how often, and how long it takes each time. Include the small things. Five-minute tasks done fifty times a week add up faster than one big monthly report.
Ask the people who do the work, not only their managers. They know where the copy-paste steps and the workarounds are.
Score each candidate on four factors
Volume: how many times a week does it happen? Rule clarity: can you write down how a person decides what to do? Error cost: what happens if it goes wrong, and can someone catch it? Access: do the systems involved offer an API, a webhook, or an export?
A task that is frequent, rule-based, low-risk, and well connected is a strong candidate. A task that is rare, subjective, or locked inside software with no access is not, however exciting the AI behind it sounds.
Separate the AI part from the plumbing
Many 'AI' projects are mostly plumbing: moving data between tools reliably. The AI step might only be classification or extraction in the middle. Naming this honestly helps you budget correctly and choose simpler tools where they will do.
Where judgement is involved, plan a review step. A person approving a drafted reply costs seconds and prevents the expensive mistakes.
Check the data and the tools
Look at what the process reads and writes. Clean, structured inputs automate easily. Scanned PDFs, free-text notes, and several versions of the same spreadsheet need extra work first.
Also check permissions. An automation needs access to act. If getting that access will take months of approvals, factor that into the order you tackle things.
Pilot one, measure, then expand
Pick the highest-scoring process with the lowest risk. Define what 'better' means before you start: time per task, errors caught, turnaround. Run it alongside the manual process for a short period, compare, and only then widen it.
A readiness assessment or a short consulting engagement can do this prioritisation with you if you would rather not guess.