KelenAI· Kevin Li · AI implementation · 7 min read
AI Opportunity Assessment: Choose Your First Workflow
Compare AI workflow opportunities using real cases, stop conditions, and a small proof that tells you what to do next.
An AI opportunity assessment helps you decide which business workflow deserves attention first—not which AI tool looks most impressive. Compare a few workflows against the same business priority, remove candidates without a safe or testable boundary, and choose the one that can produce useful evidence with the least irreversible change.
For example, a distributor might consider emailed purchase-order intake, weekly management reporting, and customer complaint replies. Order intake could be the first AI test if staff can check extracted fields before an order reaches the ERP. Reporting may need agreed definitions and ordinary integration instead. Complaint replies may wait until policies and approval authority are clear. The assessment’s output is those decisions and their next tests, not a single score.
See the decision before the method
The table below is illustrative, not a KelenAI client result. Assume the business priority is to reduce avoidable operations backlog. None of the ratings or outcomes is a measured claim about a real company.
| Candidate | What recent cases should show | Main obstacle | Decision and smallest useful test |
|---|---|---|---|
| Emailed purchase-order intake | Repeated manual extraction, corrections, and waiting before an accepted order | Customer and product matching; permission to write to the ERP | Test now: propose fields from representative orders, validate against master data, and have staff approve before entry |
| Weekly management report | Time spent reconciling the same figures and correcting definitions | Two systems disagree about what a metric means | Fix without AI first: agree on definitions, connect sources, and test a deterministic report |
| Customer complaint replies | Repeated requests for approved information and costly escalation | No agreed response policy or owner for concessions | Prepare, then revisit: define approved evidence and escalation before testing reviewer-approved drafts |
The order-intake test is not “let AI enter orders.” It asks whether a bounded extraction-and-review path reduces total handling work without increasing incorrect orders. If the same exceptions still require lengthy correction, the candidate may lose its place. This is the decision an owner needs; a high theoretical savings estimate alone cannot supply it.
What should an AI opportunity assessment produce?
A useful assessment compares workflows—each with a trigger, completed business result, owner, inputs, and exceptions. It ends with four short decisions:
- Now: a valuable, observable problem with a bounded, reviewable first test.
- Prepare: a promising problem blocked by a named data, process, ownership, or control prerequisite.
- Not AI: the problem is real but a policy change, native software feature, rule, or integration is the better first intervention.
- Defer or stop: no meaningful priority, evidence, owner, or safe test boundary.
This is different from an AI readiness checklist. Opportunity selection asks which workflow? Readiness asks whether the selected workflow’s data, controls, people, and fallback can support a pilot. An implementation roadmap comes after both.
The Australian National AI Centre’s opportunity guidance likewise considers frequency, repetition, data, error consequences, current cost, and simpler non-AI options. Those are helpful prompts, but a score should not override an unresolved stop condition.
Step 1: Fix the business priority and collect real cases
Start with one outcome that matters now: fewer orders waiting for correction, a shorter time from enquiry to approved quote, or less overtime closing the books. “Use AI” is not an outcome. Name the person who owns the workflow result and can act on the finding.
Choose a short list of candidate workflows serving that priority. For each, inspect recent completed and troublesome cases with the people doing the work. Record the arrival event, handoffs, software, source information, exceptions, active time, waiting time, and final state. If you cannot describe what “done correctly” looks like, audit the workflow before proposing a solution.
Use the same compact record for every candidate:
| Question | Record an answer or mark it unknown |
|---|---|
| What result must improve? | One observable business outcome and its owner |
| What work repeats? | Trigger, completed state, recent case examples, and approximate volume |
| What might AI do? | A bounded task such as extracting, classifying, matching, or drafting |
| What must stay exact? | Policies, IDs, calculations, system-of-record updates, and approvals |
| What could go wrong? | Common exceptions, consequence of error, review, and recovery |
| What would change the decision? | The largest unknown and a small test of it |
Label material inputs observed, estimated, or unknown. “Staff spend too long on orders” is an impression; a sample of recent orders with touch times and corrections is evidence. A preliminary estimate is fine if it remains visibly an estimate.
Step 2: Apply stop conditions before comparing value
Do not let a weighted score compensate for a missing owner or unsafe action. Pause or narrow a candidate when:
- no one can define or own the completed result;
- required source information is inaccessible, unreliable, or cannot be used with appropriate authority;
- errors have significant consequences and cannot be intercepted or reversed;
- the process changes so often that representative cases do not exist; or
- the business cannot observe whether the proposed change helped.
A stop condition may reshape the first test. For example, autonomous complaint replies may be inappropriate while an internal draft grounded in approved policy and reviewed by a service lead remains worth examining. NIST’s AI Risk Management Framework is voluntary guidance that emphasizes context, intended use, impacts, and human oversight; it does not certify an opportunity as ready. Use that discipline to define the boundary before choosing a tool.
Step 3: Compare the surviving paths, including non-AI paths
For each remaining candidate, ask four questions in this order:
- Does it matter? Is the current friction linked to the stated business priority, not just an annoying task?
- Can we see it? Are there enough representative cases, a baseline, and an outcome a reviewer can judge?
- Does it need AI? Is variable language or document interpretation the bottleneck, or would a required field, native feature, rule, or integration solve it more directly?
- Can we test it safely? Can the first path limit access and actions, route exceptions to an owner, and return to manual work if it fails?
Consider implementation effort and recurring operating work, not just build cost: source cleanup, integration, review, exception handling, and maintenance. OpenAI Academy’s prioritization matrix uses value and complexity as a first comparison. For a small business, the tie-breaker is often the shortest path to evidence that changes a real decision. The highest-upside idea may be an important second project rather than the right first one.
If the work is exact and structured, follow a native-first implementation path instead of adding a model. If AI only produces a draft that never reaches an approved business state, it is an AI feature, not yet an AI workflow.
Step 4: Write the test and decision rule before building
For the order-intake example, a useful first test might use representative purchase orders from a permitted customer group. The system proposes fields, rules validate required values and master-data matches, and an employee corrects or approves each order before ERP entry. Compare this path with the current one on total handling time, correction work, incorrect-order findings, and exception volume. Keep the manual path available.
Write down what evidence would cause you to proceed, change scope, or stop. If matching fails on common customer aliases, perhaps master-data cleanup is the next project. If extraction works but review takes longer than manual entry, the automation boundary may be wrong. If the path helps only one unusual customer format, the opportunity may be too narrow.
Avoid converting released minutes directly into promised cost savings. Time can create capacity, faster service, or lower overtime only when the business can actually use it that way; review and support costs remain in the equation. Use a fuller AI automation ROI model once baseline and operating costs are credible.
The final one-page decision should identify the leading workflow, why it outranks the alternatives, each candidate’s stop or preparation condition, the first proof question, the owner, and the next decision date. It should be possible for someone who missed the meeting to understand why one candidate advances and another does not.
When not to run this exercise yet
If leadership has not chosen a business priority, start there. If the team has no access to recent cases, gather a small sample. If no one can own the result, appoint an owner before scoring ideas. If the work is a one-off expert judgment rather than a repeatable workflow, a portfolio ranking may add ceremony without improving the decision.
Do not treat a short conversation as a full assessment of an unfamiliar business. KelenAI’s free workflow consultation request starts with a brief description of one recurring workflow. We review the submission and follow up to decide whether a focused conversation would help identify its boundary and next evidence step. You can submit the trigger, current tools, common exception, and result you want to improve—without sharing confidential records.