Manual workloads
Repetitive tasks staff do by hand: routing, classifying, summarising, extracting numbers from documents.
→ Try: text analysis, digitisation, chatbotsAI projects fail when they start with the technology. Start with the problem instead. This step helps you brainstorm real problems your team faces, then narrow down to one or two worth piloting.
Run these as quick rounds of discussion, around five minutes for each. Capture every idea on sticky notes to extract key themes later.
Example: manually checking documents against a checklist, copying information from one system into another, or chasing people for missing details.
Example: re-entering the same data into several systems, compiling the same report by hand every month, or answering the same routine query over and over.
Example: printing and postage for letters that could be sent digitally, or staff overtime during predictable peaks that could be planned for.
Example: queuing at a counter, waiting weeks for a decision letter, or never hearing back after submitting a form.
Example: how many people will show up at a service centre next week, or how many applications will arrive before a deadline.
Example: one person who knows how a regulation is applied in practice, or a process that only works because someone remembers an exception from years ago.
Example: how long it actually takes to process a case end to end, or whether a service is reaching the people who need it most.
Example: the part of a process that causes the most complaints, or the task staff find most demoralising to do every day.
Tick everything that rings true for your organisation. The more you tick, the more raw material you have for Step 2.2.
0 ticked. Carry the ticked items into Step 2.2 as candidate problems.
Drag your ideas onto the matrix and aim to pilot something in the top-left quadrant, where impact is high and complexity is low. Avoid the top-right unless you already have very strong technical support in place.
Use the descriptions below to judge where each idea sits on the two axes. They run from low to high - anything you'd call medium or above leans toward the high end of that axis when you place it on the matrix.
Once you have a candidate problem, classify it. The category decides which AI approaches are worth investigating in Step 3.
Repetitive tasks staff do by hand: routing, classifying, summarising, extracting numbers from documents.
→ Try: text analysis, digitisation, chatbotsMaking sense of financial, operational, or administrative data: spotting patterns, trends, and outliers in what you already collect.
→ Try: anomaly detection, dashboards, predictionAnticipating future values: tax revenue, crop yields, demand, weather, outbreaks, electricity load.
→ Try: time-series forecasting, predictionDeciding where to focus limited resources: audits, inspections, social-programme targeting, fraud risk.
→ Try: supervised ML, anomaly detection, optimisationAssessing whether a programme or intervention actually worked, for whom, and at what cost, so you can decide what to keep or change.
→ Try: causal analysis, prediction, dashboardsHandling questions, requests, and complaints from the public faster, supporting more communities with fewer dropped tickets.
→ Try: chatbots, text classificationInformation locked in documents, people’s heads, or unsearchable filing systems. Index it and make it searchable.
→ Try: AI-powered search, text generationRecords that exist on paper, in different systems, or in formats nobody can analyse. Get the data ready first.
→ Try: digitisation, dashboardsPull it all together. Capture the top one or two problems you want to take forward: describe each one, match it to a type, and record its impact score. This is what you carry into Step 3.
Bring your top problem (and its category) to Step 3 to find candidate AI approaches.
Continue to Step 3 →