Created: November 20, 2026

AI Opportunity Assessment: How to Choose Which AI Use Case to Build First

Pavel Zverev.

Pavel Zverev

CTO

ML
AI Opportunity Assessment: How to Choose Which AI Use Case to Build First

Leadership wants AI on the product roadmap. Customers have suggested a copilot, support wants less manual work, and engineering has ideas for searching internal knowledge. Each proposal sounds useful, yet the team cannot explain which problem deserves the next budget or who would own the result.

An AI opportunity assessment gives those ideas a common basis for comparison before development begins. The immediate decision is which uncertainty to resolve next and whether resolving it could justify an investment. This guide shows how to define candidates, examine evidence, compare alternatives, and select a practical next step. You will leave with a way to choose what to investigate, validate, defer, or reject.

Key takeaways

What should an AI opportunity assessment help you decide?

An AI opportunity assessment should identify which opportunities deserve further investment and explain the evidence behind that recommendation. Its output is a ranked shortlist with conditions attached: what appears valuable, what remains unknown, and which next action would reduce uncertainty.

Keep the decision distinct from adjacent activities. AI strategy sets broader priorities and investment direction. An organizational readiness assessment examines prerequisites such as capabilities, governance, and infrastructure. Opportunity assessment considers those prerequisites where they affect particular candidates; it does not need to audit the entire company.

Discovery develops the scope of a selected opportunity, including users, requirements, dependencies, and a delivery approach. A proof of concept tests a bounded technical assumption. Neither activity should become the automatic destination for every idea entering the shortlist.

Agree on the assessment's decision authority before collecting proposals. A product leader may own the funding recommendation, while a workflow owner confirms the problem and engineering checks feasibility. Specify what they can approve and what requires another decision. Otherwise, a useful workshop can produce an attractive ranking that nobody has authority to act on. The assessment is complete when its audience can explain the next investment decision, including a decision to wait.

Start with the business problem, not an AI feature

Describe the work that needs to improve before naming a solution. “Add a copilot” leaves the user, task, and business consequence unresolved. A usable problem statement identifies a role, a recurring situation, the current obstacle, and an observable outcome.

Consider the hypothetical B2B SaaS support workflow developed later in this guide. Support engineers need to locate relevant incident history while investigating a ticket. Information is distributed across resolved tickets and documentation, so engineers search several places before deciding what to try. The desired outcome is more efficient investigation with accurate evidence available to the engineer.

The support operations lead should own the outcome. Before proposing AI, that owner can establish a baseline from representative tickets: how investigation time is recorded, where searches stall, and when an escalation is necessary. Separate time spent locating information from time spent resolving the underlying issue. A tool aimed at retrieval cannot reasonably take credit for every improvement in ticket resolution.

Write the hypothesis so it can fail: providing relevant, permission-appropriate evidence during investigation may improve the engineer's ability to choose a useful next action. Record what would count as improvement and what must remain acceptable, such as the accuracy of cited information or escalation behavior. This creates a basis for comparing solutions without assuming that generating text is the right intervention.

Build a shortlist from real product and workflow friction

Look for candidates where people can show the work, the obstacle, and its consequences. Review customer journeys, recurring internal operations, delayed decisions, content handling, support conversations, and customer requests. Ask process owners to bring recent examples rather than a list of desired AI features.

Customer requests are useful signals, but they still need interpretation. Establish what the customer was trying to accomplish, how they handle the task today, and whether the difficulty recurs across relevant users. Distinguish a contractual requirement from an isolated preference and from a broader workflow problem.

Capture every candidate in the same format: user or role; situation; problem; desired outcome; available data; affected systems; current workaround; and accountable owner. Include the evidence location and the unanswered question that could change the recommendation. This makes proposals comparable even when one arrives as a polished presentation and another as a frontline complaint.

Merge proposals that address the same underlying friction. Keep genuinely different interventions separate when their risks or validation paths differ. Limit the shortlist to candidates the team can investigate with available decision makers and evidence.

For a narrower application of selection principles, see Mad Devs' guide to choosing an AI use case for test automation. Here, the comparison remains at the product and workflow investment level.

Compare candidate use cases across the same criteria

Use a shared matrix to make assumptions visible. The matrix below is a working aid for this article, rather than a Mad Devs standard. Set weights, thresholds, and mandatory conditions according to your company's priorities and tolerance for error before debating favorites.

CriterionQuestion for the teamWhat counts as evidence
Business valueWhich KPI or user outcome should change?Baseline, KPI owner, expected direction or magnitude
User and workflow fitIs there a specific role with a recurring problem?Workflow map, user evidence, process owner's confirmation
Feasibility and dataAre the necessary data, access, and quality available?Sample data, access confirmation, limitations, integration checks
Total effort and operating costWhat will be required beyond the model?Integration, review, support, vendor/API, monitoring, change management estimates
Risk and ownershipWho is responsible for errors and the decision?Impact analysis, controls, approver, escalation and rollback responsibilities
Time to evidenceHow can a bounded test check the hypothesis?Testable assumption, pilot boundary, success and stop criteria

For each row, record a judgment, its supporting evidence, and the confidence in that evidence. “Data available” should point to an inspected sample and confirmed access. A stakeholder's expectation belongs in the record as an assumption until somebody verifies it.

Discuss business importance separately from how readily a candidate can be tested. An important opportunity may need substantial investigation before a useful experiment is possible. A smaller opportunity may offer a clearer test. This distinction helps leaders choose between near-term learning and preparation for a strategic dependency.

Avoid mechanically adding subjective ratings into a universal ROI score. A high value rating cannot compensate for unavailable data rights or an unacceptable consequence of error. Apply those conditions as gates, then compare the remaining candidates using your agreed priorities.

Resolve disagreements by identifying the missing evidence. If product expects frequent use while the workflow owner expects occasional use, inspect actual work records before adjusting a score. Where estimates remain uncertain, show the plausible range and explain whether the recommendation changes across it. The most useful ranking makes its reasons inspectable and specifies what new information would overturn the order.

Check feasibility, data readiness, and risk before ranking too high

Inspect a representative data sample before treating a dataset's existence as evidence of feasibility. Confirm who can authorize its use, which records are covered, how current they are, and whether important exceptions are missing. Historical records may document an outcome without capturing the context needed to support a new decision.

Check the interfaces to affected systems and the permissions the proposed workflow requires. Identify whether data can be read where the user works and whether any write action would need additional approval. Detailed AI agent integration design belongs after selection; at this stage, establish whether a critical dependency is accessible and who controls it.

Examine privacy, security, and applicable compliance requirements with the responsible specialists. NIST's AI Risk Management Framework provides a primary reference for considering AI risks in context. An opportunity assessment should flag relevant constraints and owners without claiming to constitute a compliance determination.

Agree on acceptable response times, reliability expectations, and how a person would verify an output. Include that review work in the operating estimate, alongside support, monitoring, maintenance, and vendor charges.

Separate fixable gaps from blockers. Missing documentation may require investigation with an owner and a defined deliverable. Unavailable rights to essential data may prevent a meaningful test altogether. Record the condition for reconsideration so a deferred candidate does not repeatedly return to the shortlist unchanged.

Ask whether AI is the right solution at all

Compare each candidate with the simplest credible way to achieve the outcome. Alternatives include rules-based automation, conventional search or analytics, an existing SaaS tool, a process change, and keeping the current workflow. Give these alternatives the same outcome and constraints so the comparison is fair.

Inspect the inputs and exceptions. Stable fields and explicit decision rules may support deterministic automation. Variable language or unstructured material may justify investigating AI, particularly when interpretation is necessary. However, an ambiguous input does not establish that a probabilistic output is acceptable.

Examine the cost of a wrong result and the user's ability to detect it. Human review is useful only when the reviewer has the context, time, and evidence to judge the output. A polished answer can still impose substantial verification work.

Test whether the problem comes from a missing capability or a broken process. Unclear ownership, outdated documentation, and unnecessary handoffs can undermine any new tool. Compare the cost of addressing that friction directly with the cost of adding another system.

A no-AI recommendation can be the strongest investment decision. Explain what alternative meets the need and what evidence, if any, would justify revisiting AI later.

Validate the strongest candidate before funding a full build

Choose the smallest investigation that can change the funding decision. State the critical assumption in operational terms, then identify evidence that would support or contradict it. Avoid a demonstration whose only success condition is that the system produces an output.

Select a representative sample that includes routine work, difficult cases, and relevant exceptions. Agree on the current baseline, intended outcome measures, and unacceptable trade-offs before examining results. Confirm that the sample can be used appropriately and that the person reviewing outputs understands the workflow.

Define the test boundary, reviewer, decision owner, and stop/go criteria. Specify what would lead to revision or abandonment as well as continuation. Record how findings will be compared with the current process and with a credible non-AI alternative. Keep the review effort visible when interpreting the outcome.

Use discovery when uncertainty concerns scope, user needs, dependencies, or technical direction. Broader product discovery may be appropriate when the AI proposal sits inside an unresolved product problem. Use a limited POC when the decisive uncertainty is technical and can be isolated, such as whether available records can support sufficiently relevant retrieval.

Sometimes interviews, a data inspection, or a manual workflow exercise will answer the question without a POC. End the validation with a decision record that explains what was learned, what remains uncertain, and which commitment the evidence now supports. Successful validation supports the next decision; it does not automatically justify full development.

From a ranked shortlist to a clear next step

Produce a short prioritized opportunity map that decision makers can revisit. For each candidate, include the rationale, evidence confidence, open questions, critical dependencies, owner, and recommended next action. Preserve enough context to explain why a promising idea was placed behind another.

Use explicit action labels. Investigate means gathering missing business or data evidence. Discovery means defining the selected opportunity's scope and direction. POC means testing an isolated technical assumption. Build requires sufficient evidence and the company's normal delivery approval. Defer means a stated condition must change; reject means the current proposal does not justify further investment.

Give unresolved questions an owner and a review point. A shortlist without these commitments can become a backlog of permanently promising ideas.

For teams seeking external help with this decision, Mad Devs offers AI Opportunity Consultation with a prioritized opportunity map and a recommended next step. Once a use case is selected, its AI Discovery Sprint can develop solution scope, technical direction, a prioritized backlog, and a POC/MVP plan. These formats support different decisions; selection should follow the uncertainty your team needs to resolve.

Worked example: choosing an AI use case for a SaaS support workflow

This hypothetical example continues the SaaS support workflow introduced earlier. It illustrates a decision process and contains no client results or measured gains. The support operations lead owns the outcome; a support engineer will review test outputs, and engineering will confirm access to the relevant systems.

The first candidate drafts responses to recurring support tickets. Its potential value is helping agents prepare accurate replies with less drafting work. Available inputs include incoming tickets, approved response guidance, and historical resolutions. Before testing, the team needs to determine whether those resolutions remain correct and whether account-specific context is available. Reviewers must check every proposed answer for unsupported instructions and inappropriate disclosure.

The second candidate recommends similar incidents and documentation to support engineers investigating a ticket. It targets time spent locating evidence. Its inputs overlap with the first candidate's, but the intended output is a set of relevant sources for investigation. Engineers can inspect those sources before choosing an action. Permissions and outdated material remain significant concerns.

The team provisionally selects the retrieval hypothesis for the first bounded test because engineers can judge source relevance directly during investigation. It will compare recommendations with existing search on representative tickets, examining the share of useful sources, permission compliance, and the time engineers spend finding and verifying evidence. The baseline comes from the current search workflow. The owner and reviewer agree on acceptance and stop criteria before results are reviewed.

This choice is conditional. If access controls cannot be preserved, records lack useful context, or existing search performs adequately, the team will defer the AI approach and address those issues first. Response drafting remains a separate candidate whose validation would require stronger evidence about answer correctness and reviewer workload.

Conclusion

Bring the decision to the people who own the workflow and its budget. Ask them to select a candidate whose problem is measurable, whose evidence can be inspected, and whose remaining uncertainty is worth resolving. Establish the baseline before changing the process, then commission the investigation, discovery work, or bounded POC that answers the next consequential question. Record the conditions for continuing and stopping. That decision record gives the team a practical starting point and a defensible reason to commit resources, choose a simpler alternative, or wait.

AI & Machine Learning
Choose an AI use case worth developing
Mad Devs can assess value and feasibility, then scope a discovery sprint or proof of concept.
Discuss your AI use case