Picking the right first three use cases

Picking the wrong first AI use case will set the program back 12-18 months. Picking the right one compounds into broader success. Here's the selection framework.

The five selection criteria

For each candidate use case, score:

1. Value density. How valuable is the task, per person per week?
2. Error tolerance. What happens when AI gets it wrong?
3. Data availability. Is the data to make this work already in usable shape?
4. User enthusiasm. Does the target team actually want this?
5. Replicability. Would success here inform other use cases?

Prioritize high scores across all five.

Value density

Rule of thumb: save ≥ 30 minutes per user per week, to ≥ 20 users, for ≥ 6 months. Below that, the deployment cost doesn't amortize.

High-density candidates:

• Customer support triage (high volume, clear task).
• Report generation from structured data.
• Meeting notes and action items.
• Knowledge lookup in large content bases.

Low-density candidates:

• Tasks 2-3 people do rarely.
• Tasks already nearly optimal (marginal gain small).
• Tasks where humans can't accept AI output without heavy review.

Error tolerance

AI produces wrong answers sometimes. Good first use cases have one of these properties:

• Errors are cheap to catch and fix (human review is part of the workflow).
• Errors are low-impact (drafts, suggestions, not decisions).
• Errors are rare and surfaced (confidence thresholds exclude low-confidence outputs).

Bad first use cases have:

• Errors that propagate (AI result feeds another AI result).
• Silent errors (plausible wrong output, no verification).
• High-stakes errors (legal, financial, health — wrong answer causes real harm).

High-stakes uses are eventual destinations, not first use cases.

Data availability

Ask: Is the data this use case needs already:

• Digital and searchable?
• Well-structured?
• Accessible with appropriate permissions?
• Reasonably current?

If no to multiple, prioritize data work first. AI can't compensate for missing or rotten data.

User enthusiasm

Teams that want AI outperform teams that are asked to adopt AI, every time.

Early pilots should go where:

• Someone's advocating for it.
• There's a specific pain point they'd AI-solve.
• The team is technically capable and curious.

Don't start in the most skeptical or disengaged team. You'll spend all your capital on change management and get lukewarm results.

Replicability

Does success here teach you something usable elsewhere?

High replicability: deploying an AI coding assistant to one engineering team — learnings directly apply to other engineering teams.

Low replicability: bespoke agent for one unique workflow — success is valuable but doesn't inform next steps.

For a first use case, prefer replicable.

The "3 x 3" exercise

List 3 candidate use cases. Score each 1-5 on the five criteria. Total out of 25.

If the top two are close, go with the one that has the highest user enthusiasm score. That single factor predicts success more than any other.

Common bad first picks

• Legal document review. High stakes, low error tolerance, expensive to get wrong.
• Executive briefings. Low-volume, not replicable, exposes AI to high-scrutiny eyes.
• Customer-facing chatbot. Visible to customers; failures damage brand; requires mature infra.
• HR decision automation. Biased AI could produce discrimination risk; heavy review burden.

These are all valuable eventual targets. They're terrible first deployments.

Common good first picks

• Internal knowledge search for a specific org.
• Code assistance for engineering teams.
• Meeting summaries and action items.
• Draft generation for routine communications.
• Support ticket triage (drafting responses, classifying, not auto-sending).

Notice the pattern: high volume, human-in-loop, clear task, enthusiastic users.

The 3-use-case initial strategy

Don't pick just one. Pick three:

• One clear win — high-confidence, moderate scope.
• One moderate challenge — real value, meaningful change management.
• One exploration — worth trying, acceptable if it fails.

The portfolio hedges against any single failure. Also teaches your team how to run AI rollouts.

Budgeting for the three

Rough allocation:

• Clear win: 50% of resources.
• Moderate: 35%.
• Exploration: 15%.

Don't under-fund the win because it "should be easy." It still requires good execution.