Customer health report
High weekly effort, stable sources, one operating owner, reversible drafts. Strong first candidate: measure time-to-report and decision latency, then automate preparation with human approval.
AI opportunity selection
AI use case prioritization framework: how to choose your first project
Start with business pressure and a repeatable unit of work. The best first project has meaningful value, enough volume, usable data, a stable process, an accountable owner, and consequences you can control.
By Naman Kumar · Updated October 29, 2026
Short answer: score each candidate from 1–5 on the six dimensions below. A 5 always means more favorable conditions, including less change burden and lower risk. Favor high-value, high-frequency work with an owner and measurable baseline. Penalize unstable processes, inaccessible data, and high-consequence decisions even when the theoretical upside is large.
The matrix
How should you score AI opportunities?
| Dimension | Ask | 5 looks like | 1 looks like |
|---|---|---|---|
| Business value | Would this materially affect revenue, margin, capacity, quality, or risk? | Leadership KPI with a clear economic mechanism | Convenience with no owned outcome |
| Frequency and volume | How often does the unit repeat? | Daily, high-volume, similar cases | Rare, bespoke work |
| Process stability | Can operators describe the normal path and exceptions? | Documented, observable, bounded exceptions | Changes constantly or depends on heroics |
| Data readiness | Are required inputs accessible and trustworthy? | Structured, permissioned, representative | Missing, inaccessible, or contradictory |
| Change burden | How many roles, systems, approvals, and habits must change? | One owner, few systems, reversible rollout | Enterprise-wide behavior change |
| Risk and judgment | What happens when the system is wrong? | Low consequence, easy review and reversal | Material legal, financial, safety, or customer harm |
Read the scores as a profile, with 5 always favorable. Do not subtract the change-burden or risk scores: a 5 already means low burden or low risk. Before ranking, require a named owner, a measurable baseline, permitted data access and an agreed review or approval process. A missing requirement means resolve that gap before implementation. Then compare business value and frequency among candidates you can safely deliver; use stability, data readiness and change burden to choose what to start first. These scores organize judgment; they are not a validated prediction of returns.
Worked example
Which AI project should go first?
Contract negotiation
High value but low frequency and high judgment. Use AI for research and clause comparison, not autonomous decisions. Keep legal approval explicit.
Company-wide knowledge bot
Broad appeal but vague owner, permissions, success metric, and source quality. Narrow it to one decision and one audience before funding a build.
| Candidate | Value | Frequency | Stability | Data | Low change burden | Low risk | Decision |
|---|---|---|---|---|---|---|---|
| Customer health report | 4 | 4 | 4 | 4 | 5 | 4 | Start with reviewed drafts and measure preparation time. |
| Contract negotiation | 5 | 2 | 2 | 3 | 2 | 1 | Limit scope to research and comparison; legal approval stays with a person. |
| Company-wide knowledge bot | 3 | 5 | 2 | 1 | 1 | 2 | Resolve ownership and access, then narrow the first use case. |
The report goes first because it has a clear owner and a bounded, reviewable output. High potential value does not make contract negotiation ready for autonomous decisions, and frequent use does not repair missing permissions.
Deferred work
What if the work is not getting done today?
A time-savings calculation misses work the team has stopped doing. Examples include reviewing support conversations for recurring problems or checking whether onboarding instructions still match the product. The question is whether doing that work would change a decision or outcome.
Name the backlog, who would use the result, and what they would do differently. Then test a small sample with human review. Measure useful findings, errors, review time and the cost of running the process. Treat unmeasured value as a hypothesis; do not claim hours saved when nobody was doing the work.
Deferred work still needs an owner and a stop rule. If the output does not lead to useful action, cheaper execution has not made it worth funding.
Sequence
How do you turn the shortlist into a funded pilot?
1. List workflows, not tools
Name the unit of work, owner, volume, handoffs, and failure cost. “Use an agent” is not an opportunity definition.
2. Quantify the current state
Measure touch time, wait time, rework, loaded cost, quality, and downstream capacity or revenue effect.
3. Fund a bounded proof
Pick one baseline, one owner, one review window, one human-control policy, and one stop rule.
Find the constraint before choosing the tool
In earlier founder work, the constraint was the training-data collection process, not just the model. Read how that process was redesigned to produce more usable training data. Use the same question on your shortlist: what is actually stopping progress?