Most organisations do not have a shortage of AI ideas. The problem is working out which ones deserve attention, which can be solved simply and which are worth investing in.
Start with business problems, not AI features. Prioritise work that is frequent, costly, slow or difficult, then test the smallest workable solution against a clear baseline before committing to a larger build.
Instead of “What can we build with AI?”, ask: “What outcome matters, what is getting in the way, and what is the lightest intervention that could improve it?”
A long use case list can create the impression of progress while making decisions harder. “Use AI for reporting” tells us nothing about who does the work, what takes time, what good looks like or why the current process fails.
A better use case describes:
Practical discovery works best around real work. Training, process reviews and conversations with subject matter experts surface the details that generic innovation workshops miss. Recent programmes have used cohort preparation and follow-up sessions to capture opportunities in a use case register for later assessment.
Not every valuable AI use case needs to become an application.
It helps to divide opportunities into three levels:
This prevents teams from overengineering straightforward tasks. A prompt and reliable source may solve the problem. A shared agent may be enough for repeated team work. Custom development becomes relevant when a process crosses systems, operates at scale or needs stronger control and integration.
In one recent programme, the possible response was deliberately broad: a shared prompt, agent, automation, process change, platform build or waiting for the platform to improve. AI products are changing fast, and a costly custom solution can be overtaken by a standard feature before it pays for itself.
Assess each candidate against six questions:
Estimate time saved, faster turnaround, increased capacity, reduced cost, better consistency or lower risk. Be specific about who benefits.
Saving two hours once a year is unlikely to justify a build. Saving 20 minutes across hundreds of tasks may do. Frequency and reach can turn a modest improvement into a strong business case.
AI will not rescue a process nobody can explain. Strong candidates have recognisable inputs, an intended output and someone who understands the exceptions.
One investment team’s due diligence process was suitable because its information hierarchy could be mapped, the repeated effort was material and human review would remain mandatory. The aim was to complete straightforward questions reliably so specialists could focus on nuanced ones.
Check where information sits, whether it is current, who may access it and whether sensitive data is involved. The safe first version might use approved documents, anonymised data or a narrower scope.
Prefer use cases that can be tested in days or weeks. A controlled pilot should show whether the output is useful before you invest in integration or production engineering.
Someone must own the result, sources, checks, adoption and maintenance. Without an operational owner, a technically impressive tool usually becomes an unsupported experiment.
For time-based benefits, start with a conservative calculation:
Annual benefit = time saved per task × task frequency × number of users × realistic adoption rate × appropriate employment cost
Then compare that with the full cost of licences, design, configuration, integration, testing, training, support and ongoing review.
Do not stop at time saved. Quality and risk may matter more. An AI-assisted review might improve consistency, identify missing information or create a clearer audit trail. Those benefits still need evidence.
One organisation had already scored use cases by complexity, implementation difficulty and data sensitivity. Rather than repeat discovery, it could select contrasting pilots, build them with subject matter experts and test reliable use.
Another technology team produced a list covering email summaries, knowledge articles, proposal reviews, scripting, troubleshooting and system connections. Because it came from real activity it was useful, but it still needed sorting into personal productivity, reusable team practices and higher-control integrations.
Before the pilot starts, record the baseline:
Test with representative users and approved examples, then measure the same things again. Record failures, user overrides and checking effort. A pilot that saves drafting time but doubles review time has not proved its value.
The decision at the end should be explicit: stop, improve the prompt, change the process, configure a shared tool, integrate with another system, build a controlled solution or wait.
The strongest AI portfolios are not the ones with the most ideas. They are the ones that connect a genuine business problem to a proportionate response, a named owner and measurable evidence.
Start small enough to learn, but with a problem important enough that success will matter. That is how you find the AI use cases worth building.
Whether you know exactly what you want to use AI for or are unsure, maybe it's both, we can help you find the best foot forward. We've worked with many Financial Service firms from use case development and implementation to Copilot and AI training. To see where you could go with AI contact us using the form below.
Or keep up with our bi-weekly LinkedIn Newsletter all to do with practical Copilot and M365 tips for everyday work.