Finding high-ROI use cases inside your company

Most AI projects fail because the team built the wrong thing, not because the tech didn't work. A systematic use-case discovery process is the cheapest insurance.

The four lenses

Apply all four to each candidate use case. Anything that scores weakly on any lens is a "not yet" rather than a "not ever."

1. Volume × time-saved. How many people do this task, how often, and how long does it take? Saving 30 min/week for 200 people = $1M/year of capacity at $100/hr loaded cost. Saving 30 seconds once a month for 5 people doesn't clear implementation cost.
2. Tolerance for imperfect output. If a 5% error rate is acceptable (with review), AI is a fit. If 5% errors cause compliance or reputation risk, you need heavy review infrastructure — budget for it.
3. Availability of the data. Does the AI need access to documents / databases / context that exists and is well-structured? Use cases where the answer is "yes, this is all in Confluence/Salesforce/S3" ship faster than ones where you'd have to build data infrastructure first.
4. Change management load. Does using this require people to change workflows? Tools that fit into existing workflows adopt fast. Tools that require people to change routines need deliberate change management.

The high-ROI shapes

Top categories that consistently deliver ROI in 2026:

• Customer support triage and drafting. AI suggests a response, human sends. 30-40% handle-time reduction is common.
• Meeting and call intelligence. Summaries, action items, searchable transcripts. Low risk, high volume.
• Internal knowledge access. "Where's our policy on X?" A well-deployed enterprise search tool saves hours per person weekly.
• Coding productivity. IDE assistants deliver measurable speedup on common tasks.
• Routine document generation. Quarterly reports, standard contracts, RFP responses. AI draft, human polish.

The false-high-ROI mirage

Watch for projects that sound high-ROI but aren't:

• "AI for sales calls." The model transcribes and summarizes fine. Adoption flops because salespeople don't want to use new tools mid-call.
• "AI for code review." Works technically. Culturally, it either annoys reviewers (noise) or gets ignored.
• "AI for executive decision support." Executives trust their own judgment; AI summaries often don't change decisions.

The lesson: technical feasibility ≠ organizational adoption. Always check: would the people who'd use this adopt it willingly?

The pilot framework

For any candidate project:

1. Pick one team of 5-20 people as the pilot.
2. Define one success metric upfront (e.g., "reduce ticket handle time by 20%").
3. Run for six weeks.
4. Evaluate honestly: measure the metric, survey the users, decide.
5. If it worked, roll out to adjacent teams one at a time.

Most failures happen when pilots are too big, too short, or lack a defined success metric.