How to Find and Prioritize AI Use Cases by ROI (Framework + Scoring Template)

A step-by-step framework for transformation and AI leaders: where AI use cases come from, how to estimate the value of each one, how to score and rank them, and how to validate the impact before committing budget.

June 26, 2026Updated November 5, 202611 min read
how to prioritize AI use cases by ROIAI use case prioritization frameworkhow to identify AI use cases

The short answer

To prioritize AI use cases by ROI, map the current pain points and initiatives across every area from the people who do the work, estimate the value of each one in hours, cost and risk, score them on the same weighted criteria, and validate the top candidates with a small pilot before scaling. Horizon runs the first step with AI interviews across the organization; at Mercado Libre, that process reduced more than 600 overlapping initiatives to 24 with ROI, effort and a roadmap attached.

The hard part is rarely finding ideas. Enterprises usually have too many AI ideas with uneven evidence behind them. The framework below makes every use case compete on the same criteria, so the ones with real impact rise and the loudest requests stop setting the roadmap.

Key takeaways

Where do AI use cases come from?

Strong AI use cases start from a problem in the work, and the people closest to the work know where the problems are. Many portfolios start somewhere else: a vendor demo, an executive request or a team experimenting with a new tool. The result is a long list of ideas with little evidence and a lot of overlap.

A better source has three parts:

  1. Every current initiative, across every area. List what is already running or proposed, including pilots inside individual teams. Overlap shows up quickly once the list is complete.
  2. Pain points from the people doing the work. Ask each role where work waits, what gets done twice, which checks are manual and what they would automate. This is where the use cases with real hours behind them appear.
  3. The process context. For each pain point, the process step, the systems involved and the teams affected. A use case without a process attached cannot be sized.

Collecting that evidence through workshops reaches a small sample. AI interviews reach everyone in scope in days, so the list reflects how work actually runs. For techniques to surface candidates, see how to identify AI use cases, and for how the interviews work, AI employee interviews at scale.

How do you estimate the value of each AI use case?

Estimate value before effort, and write the assumptions next to the number. A simple formula covers most operational use cases:

Annual value = (hours per year × loaded hourly cost) + (errors or rework per year × cost per error) + revenue or risk impact

Worked example with public numbers. At Trafilea, AI interviews quantified 218 hours a month spent on duplicate status tracking across two project tools. That is 2,616 hours a year. At an illustrative loaded cost of $40 an hour, the annual value of removing the duplication is about $105,000, before counting the delays it causes downstream. With the company's own hourly cost, the same calculation gives the real figure.

Two cautions. Hours saved only become value when someone redirects the freed time, so name where it goes. And the use case's value is the part the AI actually removes; a step that still needs human review keeps part of its cost.

How do you score and rank AI use cases?

Step 1: apply the six gates

A use case enters scoring only when it passes all six:

  1. Strategic fit with a named business priority.
  2. A business owner who will be accountable for the result.
  3. Process visibility: someone can describe how the work runs today.
  4. Governed data and system access for the data the use case needs.
  5. Acceptable risk and compliance profile.
  6. A plausible path from pilot to rollout to measurement.

A use case that fails a gate goes to redesign or discovery before scoring.

Step 2: score each criterion from 1 to 5

CriterionWeightWhat a 5 looks like
Business value25%Large, quantified annual value with the assumptions written down
Evidence strength20%The problem was described by many people across teams, with hours attached
Data and technical readiness15%The data exists, is governed and can be accessed
Adoption readiness15%The affected team wants the change and has an owner for it
Execution feasibility15%Low effort and few dependencies (effort is scored in reverse)
Strategic fit10%Directly tied to a top business priority

Multiply each score by its weight and add them up for a total out of 100.

Step 3: rank the list

An example with six use cases from the same organization:

Use caseValueEvidenceDataAdoptionFeasibilityFitScoreCategory
Process discovery for operational bottlenecks54445488Do now
Duplicate status tracking consolidation45444382Do now
Frontline knowledge assistant43443473Prove next
Invoice exception triage assistant44333471Prove next
Automated executive reporting22435258Defer
Customer churn prediction42222454Redesign

Scores are illustrative. Weighted total = (Value × 25 + Evidence × 20 + Data × 15 + Adoption × 15 + Feasibility × 15 + Fit × 10) ÷ 5.

The point of the table is the order it produces. Automated reporting is easy to build and lands low because the business impact is small. Churn prediction has high potential value and weak evidence and data, so it goes back for redesign before anyone builds it.

How do you validate impact before investing?

A high score is still an estimate. Two checks turn it into evidence before budget is committed.

Check whether the same problem appears in several places. If the pain point shows up in several teams, subsidiaries or countries, the value multiplies and the case gets stronger. If only one team describes it, the use case may be local, and the value estimate should shrink to match. Interviews across areas make this check possible, because the same problem described in different words gets grouped into one finding with a count behind it.

Run a small pilot in one area. Pick the team where the problem is clearest, set a baseline for the metric the use case should move before the pilot starts, and measure after a fixed period. A pilot that cannot show movement on its own baseline should not scale, however good the demo looked.

Customers who come to this question usually say the same thing: finding initiatives is easy, and proving there is real impact behind each one is the hard part. The two checks above are how that proof gets built.

Should you redesign the process before automating it?

Often, yes. Automating a step that should not exist makes a bad process faster and harder to change. Before scoring a use case, ask whether the step is still necessary, whether the approval reduces real risk, and whether the root cause sits upstream. When the answer points to a broken process, the use case belongs in redesign, and the fix may need no AI at all. For more, see automating a broken process.

How do you go from a ranked list to a roadmap and a board business case?

The four categories become the roadmap:

For each "do now" item, the board business case needs five things: the problem with the evidence behind it, the annual value with its assumptions, the cost and effort, the risks and how they are handled, and the metric that will prove it worked. For a full template, see building a business case for AI transformation.

Review the portfolio on a fixed rhythm, quarterly at minimum. New evidence, finished initiatives and strategy changes move use cases between categories, and a portfolio without a way to stop things only grows.

Where Horizon fits

Horizon is an AI-powered continuous discovery platform. It covers the two steps that decide whether the rest of the framework works: finding the use cases in the work, and building the evidence that each one is real.

Discovery Cycles run AI-led interviews across every area at once. The Insights Dashboard groups what people described by process, system and cause, so several teams reporting the same problem appear as one finding with its hours attached. The Initiatives Dashboard turns the top findings into business cases with owners, effort and expected impact.

At Mercado Libre, the company was carrying more than 600 initiatives, many of them overlapping or misaligned, with a business objective that every initiative deliver ROI above 150%. Horizon interviewed 2,000 employees in four days across five countries. Cross-country benchmarking surfaced duplications that were invisible from inside any single market, and the portfolio resolved to 24 initiatives with ROI, effort and a roadmap attached, all clearing the 150% threshold. The result was $2.3M in projected annual savings and 76 FTE equivalents avoided, with first implementations live in under a week.

That is one engagement under specific conditions. The bulk of the reduction came from finding which initiatives described the same underlying problem, which reading titles alone cannot establish.

Find your highest-ROI AI use cases in one week. Run Horizon's one-week challenge on one area: AI interviews with the people who do the work and a ranked list of use cases with the evidence behind each one. Start the one-week challenge

Checklist before a use case enters the roadmap

  1. Is the business outcome named and measurable?
  2. Was the pain point described by the people who do the work?
  3. Is the annual value estimated, with assumptions written down?
  4. Does the use case pass all six gates?
  5. Does the same problem appear in more than one team or market?
  6. Is there a baseline for the metric it should move?
  7. Is there a named owner on the business side?
  8. Would redesigning the process remove the need for AI?
  9. Which of the four categories does it fall into?
  10. When will it be reviewed again?

FAQ

How many AI use cases should we start with?

Three to five in "do now", with one pilot per "prove next" item. A small number of well-evidenced use cases is easier to fund, staff and measure than a broad portfolio of partial efforts.

How do we measure AI ROI?

Set a baseline for the business metric the use case should move before it launches, then compare after a fixed period: hours, cycle time, error rate, cost per case or revenue. Track whether the freed time was redirected, since hours saved only become value when someone uses them.

Who should own AI use case prioritization?

A transformation or AI office runs the process and the scoring; each use case has a business owner accountable for its result. Technology teams score data readiness and feasibility; the business scores value and adoption.

How long does it take to prioritize AI use cases across a company?

The discovery step usually takes longest. With AI interviews, it can run in days: Mercado Libre interviewed 2,000 employees in four days. Scoring and ranking then take one or two working sessions with the owners.

What data do we need to prioritize AI use cases?

For each use case: the hours spent today, the volume, the error or rework rate, the systems and data involved, and the people affected. The people doing the work supply the bulk of it; system data confirms volumes where it exists.

Evidence first, then budget

AI portfolios get stronger when every idea has to prove itself on the same terms. Start from the work, put a number on each use case, score them together, and validate the best ones in one area before scaling. The roadmap that comes out of it is shorter, and each item on it has a reason to be there.

See it. Fix it. Lead it.

Ready to transform?

See Horizon in Action

Discover how AI-powered organizational discovery can uncover hidden opportunities in days, not months.

Get Started

Related Resources