AI Use Case prioritisation
Decide which AI ideas to build, which to enable and which to stop.
Most organisations have more AI ideas than people to build them. A shared method turns a queue of competing requests into decisions that leaders can explain and teams can trust.
Without a shared method, the loudest request wins.
Every use case feels urgent
Each function believes its idea matters most. Without common criteria, there is no fair way to compare them.
Engineering time is finite
Engineers can take on only a few use cases each sprint. Every yes is also a no to something else.
Requests arrive as solutions
People ask for an agent before they have named the outcome. The better question is what they need to achieve.
Six steps, each with one owner. Only build and pilot work uses engineers.
Submit the request
States the outcome, one KPI and a baseline taken before any build starts.
Function owner
Pass the intake gate
Four must-pass checks. A request that fails goes back with the reason.
AI enablement lead, with CIO and CISO on new tools
Score it
Five weighted criteria, scored 1 to 5 and visible to every requester.
AI enablement lead, with the function owner
Decide the route
Each sprint, using the score and the engineering capacity available.
Head of AI and AI programme lead
Deliver by the right route
Build, pilot, self-serve, or park. Not every use case needs an engineer.
AI engineers or AI enablement lead
Review the value
At 30, 60 and 90 days against the baseline. Then scale, redesign or stop.
Function owner and AI enablement lead
Nothing is scored until it passes four checks.
The gate is where low-value requests stop, before they take anyone’s time.
- A named business owner who will own adoption after go-live, not just the request.
- An outcome and one KPI, with a baseline. Written as a hypothesis: we believe this use case will achieve this outcome for this team, and we will know when the KPI moves by an agreed amount.
- Tools and data within the approved catalogue, or reviewed and cleared by the CIO and CISO before scoring.
- Not already solvable by an existing feature, a prompt pattern or a simple automation.
Five criteria, weighted so that measured value leads.
Each criterion is scored 1 to 5, then weighted to a total out of 5. The weights are working defaults, calibrated in the first scoring round.
| Criterion | Weight | Question it answers | Scores 5 when | Scores 1 when |
|---|---|---|---|---|
| Value | 35% | How much measurable benefit, and for how many people? | A large, measured gain in time, cost, revenue or risk across a whole function | A small or unmeasured gain for a handful of people |
| Feasibility | 20% | Can it be delivered with the data, tools and engineer time available? | Data is accessible, tooling is approved and it fits in one sprint | It needs new data, a new integration or several sprints |
| Strategic fit | 15% | Does it advance a stated company priority? | It directly supports a named priority in the organisation’s AI direction | It has no link to a current priority |
| Risk and governance | 15% | How safely can it run? | Internal, low-sensitivity data with a person reviewing outputs | Customer or regulated data, or decisions made without human review |
| Adoption readiness | 15% | Will people actually use it? | A committed owner, named champions and a small change to the workflow | No owner time and a large change to how the team works |
Value carries the most weight because the aim is measured benefit, not activity. Feasibility comes second because engineer time is the scarce resource. The other three are checks: a high-value idea with weak fit, high risk or no committed owner should not jump the queue.
The score decides the route, and not every route needs an engineer.
Build now
High value and high feasibility. Engineers take it into the next sprint.
Pilot
High value but uncertain. A time-boxed prototype with exit criteria agreed before it starts.
Self-serve
Met through enablement: an approved tool, a prompt library or a guided agent. No engineering time.
Park or stop
Low score. The reason is recorded and the case is revisited next quarter.
Five rules keep the pipeline honest.
- A capacity limit. A fixed number of use cases in build at any time. A new item enters only when one finishes or stops.
- One visible backlog. Every request, its score and its route in one place that stakeholders can see.
- Value reviews at 30, 60 and 90 days. Active users, repeat use and KPI movement against the baseline. Anything below its exit criteria is redesigned or stopped.
- The same gate for senior requests. A sponsor can move a request up only by trading another item out or by funding extra capacity.
- A quarterly reset. Weights, capacity and the backlog are reviewed against company priorities and changes in available AI capability.
The framework can be running within four weeks, using requests and tools that already exist.
Week 1
One backlog
Collect every open request and every AI tool in use, with a named owner for each.
Week 2
Agreed model
Agree weights, thresholds and the capacity limit. Confirm the review path with the CIO and CISO.
Week 3
First scoring round
Score with function owners. Publish a ranked backlog that every requester can see.
Week 4
First decisions
Make the first route decisions and take baselines for build and pilot work.
The direction is clear. The order matters.
If your AI backlog is growing faster than your capacity to deliver it, a short conversation is a good place to start.