How to Find, Evaluate, and
Prioritize AI Use Cases
A practical, run-it-yourself methodology for finding where AI and automation actually belong in your organization - plus the scoring rubric, worked examples, and pitfalls that decide whether a use-case list turns into funded work or another abandoned whiteboard.
16-minute read - Applies to any department or team size
AI use case discovery, defined
AI use case discovery is the structured process of finding, describing, and validating specific, bounded opportunities for AI or automation to improve a business outcome - starting from real operational signals (volume, cost, error rate, exceptions) rather than from a technology looking for a home.
Most organizations don't lack AI ambition; they lack a reliable way to turn that ambition into a ranked list of specific, fundable candidates. McKinsey's 2025 State of AI research found that while most enterprises have deployed generative AI somewhere, roughly 80% report no material earnings impact yet - a gap consistent with use cases proliferating faster than they're systematically identified, evaluated, and scaled. McKinsey, The State of AI: How Organizations Are Rewiring to Capture Value (2025).
This guide is a methodology, not a pitch: every step below can be run manually with a spreadsheet, a calendar, and a rubric. Where a step of the methodology is exactly what IntakeOS automates, that's called out directly - not as a substitute for understanding the method, but as the answer to "how do I do this without it consuming a quarter of my team's time?"
“The bottleneck in enterprise AI adoption is rarely the technology. It's reliably finding, describing, and ranking the right problems to point that technology at.”
Six steps, run in order
Each step works whether you're running it with a spreadsheet or letting a platform run it for you. Skipping a step doesn't remove the work it does - it just means that work resurfaces later as rework.
Build the candidate list from real signals, not guesses
Run it yourself
Pull the signals that point at pain before you talk to anyone: ticket and case volumes, cycle times, error and rework rates, overtime patterns, and headcount concentrated in manual steps. Ask every team lead for their top three time sinks. Treat every answer as a candidate, not a commitment.
How IntakeOS runs this step
VARA, IntakeOS's AI business analyst, interviews stakeholders across every department in parallel instead of one workshop at a time, so the candidate list forms in days rather than the weeks a serial workshop schedule takes.
Run structured interviews, not open-ended chats
Run it yourself
Use a fixed question set for every interview - volume, frequency, systems touched, exception rate, current pain, and what "done well" looks like - so answers from different teams stay comparable. An unstructured conversation produces a good story and an unusable process map.
How IntakeOS runs this step
VARA runs the same structured question set on every interview: it verifies claims against evidence, structures answers consistently, and probes exceptions the interviewee wouldn't think to volunteer - the same rigor, at any scale, with no interviewer-to-interviewer drift.
Map the process and its exceptions, not just the happy path
Run it yourself
Document the current-state flow end to end, including the handoffs between systems and departments where errors concentrate, and the exception paths employees handle from memory. Most manual process maps only capture the happy path because that's what gets described first.
How IntakeOS runs this step
Every validated intake generates a Process Design Document capturing the as-is and to-be flow, stakeholders, exception paths, and KPIs automatically - and it's routed to the business analyst lead for validation rather than left as raw interview notes.
Score every candidate against the same fixed rubric
Run it yourself
Evaluate each candidate on the same dimensions every time - volume, cost, complexity, exception rate, system dependencies, compliance sensitivity - and write the scores down. Comparing candidates on gut feel instead of a fixed rubric is the single most common reason a backlog ends up ranked by politics.
How IntakeOS runs this step
8-dimensional deterministic scoring evaluates every candidate against the same explicit rules and logs which rule fired on which field value, so the ranking is reproducible and auditable rather than a subjective call - and it also flags candidates that are really a broken system needing a fix, not an automation opportunity.
Quantify ROI in the organization's own numbers
Run it yourself
Build the ROI case from the volumes, handle times, and labor rates you actually collected, not an industry rule of thumb - a defensible business case has to survive a CFO asking where each number came from.
How IntakeOS runs this step
The ROI engine calculates annual value and payback from the same intake data the moment the interview completes - the inputs, assumptions, and formula are documented alongside the number, so it can be audited later, not just quoted.
Keep a living backlog instead of a one-time list
Run it yourself
Revisit the ranked list on a cadence, add new candidates as they surface, and re-score anything whose volume, systems, or exception rate has changed - a use-case list from a single workshop starts going stale the day it's finished.
How IntakeOS runs this step
Portfolio Intelligence keeps every candidate ranked continuously, not just at the moment it was scored, so leadership is always looking at a current backlog rather than a snapshot from the last review cycle.
A discovery question set and prioritization rubric
Use both together: the question set turns a conversation into structured facts, and the rubric turns structured facts into a comparable score. This is the same rubric IntakeOS's 8-dimensional deterministic scoring applies automatically to every intake.
Discovery interview questions
- How many times does this happen per day, week, or month?
- How long does one instance take, start to finish?
- What percentage of instances are exceptions to the standard path?
- Which systems does someone touch to complete this?
- What happens today when something goes wrong partway through?
- Who currently does this, and what else could they be doing instead?
- What would "done well" look like if this ran perfectly?
- Is the underlying problem the process itself, or a system that's broken?
Prioritization rubric - 8 dimensions
- 1Volume. How often does this happen? Low-volume work rarely justifies build cost.
- 2Cost per instance. Labor time and burdened rate multiplied across the volume.
- 3Error / rework rate. How often does it have to be redone, and what does that cost?
- 4Complexity. Number of decision points, systems, and conditional branches involved.
- 5Exception rate. What share of cases fall outside the standard path?
- 6System & data dependencies. How many systems does a solution have to integrate with?
- 7Compliance sensitivity. Regulatory, audit, or data-handling exposure if this goes wrong.
- 8Estimated ROI & payback. Annual value against estimated cost to build and maintain.
What this looks like across departments
Illustrative examples of how the same methodology surfaces and scores very different work when run through an IntakeOS-led intake. Figures are representative of typical intake findings, not audited customer results.
Finance — Accounts Payable
The signal: Invoice exceptions that don't match a purchase order pile up for manual review at month-end close.
What the intake surfaces: An AP lead's intake surfaces that 22% of invoices require manual matching, each taking roughly 18 minutes and touching three systems. The exception rate and system-dependency score push this candidate to the top of the finance backlog, flagged for intelligent document processing plus workflow.
HR — Employee Onboarding
The signal: New hires wait days for account provisioning because IT, facilities, and HR each work from a separate checklist.
What the intake surfaces: The intake captures three disconnected checklists and a 30% rate of missed steps that surface only when a new hire can't log in on day one. Cross-system handoff risk and compliance sensitivity (access provisioning) both score high, ranking it above lower-volume candidates from the same department.
Customer Support — Ticket Triage
The signal: Tier-1 agents spend the first few minutes of every ticket manually categorizing and routing it before work even starts.
What the intake surfaces: Volume is high (thousands of tickets a month) and per-instance cost is low, but multiplied across volume the annual cost is material. The intake scores this as high-volume, low-complexity - a strong fit for automated triage with a fast payback, even though no single ticket looks urgent on its own.
Supply Chain — Purchase Order Matching
The signal: Procurement staff manually reconcile purchase orders against vendor confirmations across a mix of email and spreadsheets.
What the intake surfaces: The interview surfaces that this work is currently untracked in any system of record - it would never appear in a process-mining scan. Its combination of moderate volume and high error cost still scores it into the top quartile of the backlog once quantified.
Where manual discovery quietly fails
Each of these is a common ad hoc failure mode - and what a structured, software-run process does differently to avoid it.
A single all-hands workshop produces a whiteboard of ideas that never gets revisited. Six months later nobody can say what happened to any of them.
A structured process keeps intake open on a standing cadence and re-scores the backlog automatically, so candidates don't quietly disappear between reviews.
Without a fixed rubric, the candidate championed by the most senior stakeholder gets funded first - regardless of actual volume, cost, or ROI.
Deterministic scoring applies the same eight dimensions to every candidate and logs which rule fired, so ranking reflects the numbers, not the org chart.
An analyst takes notes in one interview, moves to another meeting, and by the time they write up the process map, the nuance is gone - documented process capture accuracy drops sharply once notes go cold.
A structured intake generates the process document from the interview itself, so nothing depends on an analyst's memory a day later.
Process mining and system-log analysis miss the spreadsheet workaround, the verbal handoff, and the judgment call an experienced employee makes without thinking about it.
Structured stakeholder interviews explicitly probe for shadow processes and undocumented exceptions that never touch a system log.
A process gets automated on top of a system that's actually broken, so the automation just makes a bad process faster instead of fixing the root cause.
A root-cause gate run before scoring asks whether the underlying issue is a broken system rather than a manual workload, and routes system-fix candidates away from the automation pipeline instead of scoring them as opportunities.
None of these pitfalls require new technology to fix - they require running the methodology consistently. IntakeOS exists because consistency is exactly what's hard to sustain by hand across dozens of candidates and multiple departments.
Stop running discovery by hand. Let VARA run the methodology.
See VARA conduct a structured intake, apply the same 8-dimensional scoring described above, and add the result to a ranked, live backlog - end to end, in one conversation.
See how the platform runs this end to endHave a question? Talk to our sales team before you decide.
Contact Our SalesAI use case discovery, answered directly.
Short, direct answers to what people search for on this topic.
What is an AI use case?
An AI use case is a specific, bounded business problem - a task, decision, or process step - that AI or automation can measurably improve, paired with a defined outcome (time saved, error reduced, revenue protected) and a way to verify that outcome. "Use AI for customer service" is a direction, not a use case; "use AI to triage and draft first-response replies for tier-1 support tickets" is one.
What is AI use case discovery?
AI use case discovery is the structured process of finding, describing, and validating candidate AI and automation opportunities across an organization - as distinct from picking a technology first and looking for somewhere to apply it. It starts from real operational pain (volume, exceptions, cost, delay) rather than from a tool, and it ends with a documented, evaluated candidate rather than a slogan.
How do you identify AI use cases inside a company?
Combine three inputs: quantitative signals (ticket volumes, cycle times, error and rework rates, headcount spent on manual steps), qualitative signals (structured interviews with the people who do the work, since they see the exceptions no dashboard logs), and a standing intake channel so opportunities keep surfacing between formal review cycles instead of only during an annual workshop.
How do you prioritize AI use cases once you have a list?
Score every candidate against the same fixed set of dimensions - volume, cost per instance, error/rework rate, complexity, exception rate, system and data dependencies, compliance sensitivity, and estimated ROI - so candidates from different departments can be ranked on one list instead of compared informally. Rank by that score, not by whichever team asked first or loudest.
How is AI use case discovery different from process mining?
Process mining reconstructs a process from system event logs, which is strong for systems of record but blind to work that happens off-system - shadow spreadsheets, verbal handoffs, judgment calls. Structured discovery through direct interviews captures that missing layer. The two are complementary: logs show what the system recorded, interviews show what the system never saw.
How does IntakeOS help with AI use case discovery?
IntakeOS runs the methodology on this page as a product, not a manual exercise. VARA, its AI business analyst, conducts the structured interviews across departments in parallel; 8-dimensional deterministic scoring applies the same evaluation rubric to every candidate automatically; and Portfolio Intelligence keeps the resulting backlog ranked and current as conditions change, instead of going stale between annual reviews.
How long should AI use case discovery take?
A single structured interview - covering volume, exceptions, systems, and pain points for one process - runs about 45 minutes when it's guided, versus the 8-14 days a manual cycle of scheduling, interviewing, and documentation typically takes per candidate. The gap comes from removing scheduling friction and manual write-up, not from skipping any step of the methodology.
Who should own AI use case discovery - IT or the business?
Neither alone. IT and data teams know what's technically feasible and what data exists; business teams know where the actual pain, volume, and exceptions live. The methodology on this page is designed to be run by a business analyst, an automation Center of Excellence, or a consulting team - the roles change, the steps and rubric should not.