Everyone’s talking about generative AI. Your board wants a strategy. Your engineers are building prototypes. But here is the uncomfortable truth: most of those projects will fail to deliver real money. Why? Because companies start with the technology, not the problem. They find a cool model, then hunt for a place to shove it in. That’s backwards.
Use case discovery is the discipline of finding specific business problems where generative AI creates measurable, incremental value. It isn’t about brainstorming every possible idea. It’s about filtering thousands of potential applications down to the few that actually move the needle on profit, efficiency, or customer satisfaction. As of late 2026, we’ve moved past the hype cycle. Stanford’s 2025 AI Index shows that while 71% of organizations now use generative AI in at least one function, only a tiny fraction see significant EBIT impact. The gap between adoption and value is massive. This guide walks you through how to close it.
Start With Business Pain, Not Model Power
If you ask an executive what they want from AI, they’ll say "efficiency." That’s too vague. You need to map your end-to-end workflows. Look for friction. Where do employees spend hours copying text from PDFs into spreadsheets? Where do support agents repeat the same answers? Where does legal review slow down sales?
Microsoft’s guidance on cloud adoption stresses this exact point: define the activity and the expected result first. Don’t just add a chatbot to a broken process. Redesign the workflow so the AI fits naturally. For example, instead of "improve customer service," try "reduce average handle time for tier-1 support tickets by allowing agents to retrieve policy snippets instantly." See the difference? One is a wish; the other is a testable hypothesis.
You should interview frontline workers. Process mining tools can show you where delays happen. Analyze service transcripts. Look for high-volume, repetitive tasks involving unstructured data-text, code, images. These are the sweet spots for generative models. If the task requires exact, deterministic answers with no room for error, generative AI might be the wrong tool. Save those for later, or don’t touch them at all.
The Opportunity Sizing Formula
How big is the prize? You can’t just guess. You need a defensible number. The basic formula for annual gross value looks like this:
- Addressable Volume: How many times does this task happen per year?
- Baseline Cost or Revenue: What does it cost today (labor, errors, lost sales)?
- Achievable Improvement: What percentage can AI realistically improve?
- Adoption Rate: Will people actually use it?
Let’s make this concrete. Say you have 10,000 invoices processed annually. Each takes 30 minutes. Loaded labor cost is $50/hour. That’s $250,000 in labor costs. If AI cuts processing time by 50%, you save $125,000. But wait. Do you fire staff? Or do they take on more work? If you don’t change staffing, you release capacity, not cash. This distinction trips up many finance teams. Capacity release is valuable, but it doesn’t hit the P&L immediately unless you grow revenue or reduce headcount.
Real-world data helps calibrate these estimates. A famous NBER study analyzed 5,179 customer-support agents using an AI assistant. It found a 14% increase in issues resolved per hour on average. But here’s the kicker: novice agents improved by 34%, while experts barely saw any gain. If your team is mostly senior, your savings estimate should be lower. If you’re hiring junior staff, the upside is higher. Always build conservative, base, and upside scenarios. Never rely on a single point estimate.
Prioritization: Beyond Impact vs. Effort
Most companies use a simple 2x2 matrix: High Impact/Low Effort go first. Easy, right? Wrong. Generative AI adds complexity. A high-impact idea might require perfect data you don’t have. It might carry huge regulatory risk. It might need a human to check every output, erasing the speed gains.
You need a multi-dimensional scorecard. Here are the criteria that matter in 2026:
| Criterion | Why It Matters | Red Flag |
|---|---|---|
| Data Readiness | Models need clean, accessible sources. Garbage in, garbage out. | Data is siloed, unstructured, or legally restricted. |
| Validation Burden | Humans must verify outputs. If checking takes as long as doing, you lose. | Output is hard to fact-check quickly. |
| Risk Exposure | Hallucinations can cause legal or reputational damage. | High-stakes decisions (medical, legal) with low tolerance for error. |
| Reusability | Does this build infrastructure for future use cases? | One-off solution with no shared components. |
| User Desirability | If employees hate it, they won’t use it. | Disruptive workflow changes without clear benefit to the user. |
Deloitte suggests looking at the "generation effort" versus the "validation effort." Ideally, you want tasks where AI generates content easily, and humans can validate it quickly. A summary of a meeting transcript is easy to generate and easy to scan for accuracy. A legal contract draft is harder to validate because missing clauses matter. Weigh these factors heavily.
Portfolio Strategy: Quick Wins vs. Moonshots
Don’t put all your eggs in one basket. Structure your portfolio into three waves:
- Quick Wins: Low risk, high feasibility, immediate value. Think internal knowledge search, drafting marketing emails, or summarizing support tickets. These build trust and fund larger projects.
- Strategic Transformations: Higher effort, higher reward. These might involve redesigning a core workflow, like underwriting loans or coding assistance. They require better data governance and integration.
- Platform Enablers: Investments in retrieval systems, evaluation frameworks, and security controls. These don’t always show direct ROI but make every subsequent use case cheaper and faster to deploy.
McKinsey cites a banking example where 23 potential domains were narrowed down to two: customer contact centers and know-your-customer processes. Why? Because they were measurable, scalable, and reusable. The chatbot wasn’t just a chatbot; it was a gateway to synthesizing complex customer data. Choose your pilots for their strategic leverage, not just their novelty.
Governance and Risk Management
You can’t ignore risk. The NIST AI Risk Management Framework provides the standard here. In 2026, regulators expect you to document your approach. Key risks include hallucinations, bias, privacy leaks, and prompt injection attacks.
Score risk as probability times severity. A hallucination in a marketing blog post is annoying. A hallucination in a medical diagnosis is dangerous. High-consequence use cases need strict controls: human-in-the-loop approval, audit logs, red-team testing, and defined rollback procedures. Microsoft’s responsible-AI guidance recommends mapping content risks and measuring their frequency with explicit test sets. Don’t just hope it works. Test it until it breaks, then fix it.
Also, consider intellectual property. Who owns the generated code? Can you train on proprietary data without leaking it? Legal teams need to be involved early, not after the pilot launches.
From Pilot to Production
A proof of concept (PoC) is not a demo. It’s a controlled experiment. Define success criteria before you start. Is it 20% faster resolution? 95% accuracy on citations? Reduce scope. Pick one department, one workflow, one metric. Run an A/B test if possible. Compare the AI-assisted group against a control group.
Measure realized net value, not forecasted value. Subtract the cost of inference, hosting, integration, monitoring, and human review. Many projects look profitable on paper but bleed money in operations. If the pilot fails, kill it fast. Pivot or stop. Don’t fall for the sunk cost fallacy.
Scaling is where most initiatives die. BCG reports that only 26% of companies have the capabilities to move beyond PoCs. Why? Lack of change management. Training users is critical. If agents don’t trust the AI, they’ll ignore it. If they over-rely on it, errors slip through. Design the workflow so the AI augments the human, rather than replacing them entirely-at least initially.
What is the biggest mistake in generative AI use case discovery?
Starting with the technology instead of the business problem. Companies often buy a large language model and then scramble to find uses for it. This leads to low-value pilots that never scale. Always begin with a specific, painful workflow problem and ask if generative AI is the best tool to solve it.
How do I calculate the ROI of a generative AI project?
Calculate the baseline cost of the current process (labor, errors, delays). Estimate the reduction in time or cost due to AI. Multiply by the volume and adoption rate. Then, subtract all new costs: model inference, integration, maintenance, human validation, and training. Focus on net value, not gross savings. Remember that capacity release is different from cash savings.
Which jobs are most affected by generative AI productivity gains?
Research shows that generative AI tends to help novice and lower-skilled workers more than experts. A study of customer support agents found a 34% productivity boost for novices, compared to minimal gains for veterans. This suggests that AI can flatten the skill curve, making junior staff more effective, but may not drastically speed up already highly optimized expert workflows.
When should I avoid using generative AI?
Avoid generative AI for tasks requiring exact, deterministic answers with zero tolerance for error, such as financial calculations or safety-critical controls. Also, avoid it if the validation effort required to check the output equals or exceeds the time saved by generating it. If data quality is poor or access rights are unclear, pause until governance is established.
How long should a generative AI pilot last?
Keep it short-typically 8 to 12 weeks. The goal is to validate technical feasibility, measure user adoption, and assess initial ROI. Longer pilots often drift into feature creep. Set clear go/no-go decision points based on predefined metrics like accuracy, latency, and user satisfaction scores.
