Evaluate Initial AI Feasibility
Key Takeaways
- Initial feasibility is a Phase 1 screen measured in days to weeks; it decides whether to fund Phase 2 Data Understanding, not whether the model will ultimately work.
- Screen across four axes — technical viability, data availability and quality, computational resources, organizational readiness — plus an honest comparison against non-AI alternatives.
- If deterministic rules, a process redesign, better reporting, or an off-the-shelf product solve the problem reliably, AI is the wrong tool.
- The Phase 1 compute question is only whether the needed capacity is obtainable, lawfully located, and affordable; detailed sizing waits until the AI workspace is provisioned.
- The deliverable is a recommendation carrying a confidence level and a named list of unknowns — and "not feasible now" is a legitimate, valuable outcome.
What "Initial" Feasibility Means
Domain II Task 2 is evaluate initial AI feasibility, and the word initial carries the weight. This is a Phase 1 screen — days to a couple of weeks of structured questioning — not the deep profiling of Phase 2, Data Understanding, and not the experimentation of Phase 4, Model Development. Its job is to decide whether the organization should spend Phase 2 money at all, and to state plainly what nobody yet knows. A feasibility screen that consumes two months has become the project.
The First Test: Is This Even an AI Problem?
Before scoring anything, apply the blunt screen. If deterministic business rules, a process redesign, better reporting, or an off-the-shelf product solve the problem reliably, do that instead. AI earns its cost only where the decision is genuinely probabilistic — patterns too numerous or too fluid to encode as rules — or where the input is unstructured (documents, speech, images) and rules cannot read it. A payment-terms check governed by 14 documented conditions is a rules engine. Predicting which of 400,000 invoices will be disputed is a model. Getting this backwards is one of the most heavily punished judgments on the exam.
Compare AI Approaches Against Traditional Solution Alternatives
Every feasibility memo should show the alternatives you considered and why they lost:
- Rules engine or robotic process automation — deterministic, auditable, cheap to explain; fails where patterns shift.
- Process redesign — removing the step often beats automating it.
- Better reporting or analytics — sometimes the human just needed the data in front of them.
- Buy a vendor product — fastest to value; you inherit the vendor's roadmap, pricing, and data handling.
- Do nothing — the honest baseline against which return is measured.
Inside the AI path the same comparison runs as buy versus fine-tune versus build, and at this gate you need only its rough shape: buying is fastest but sends data outside, building keeps control but takes longest, and fine-tuning sits between them. Foundational-level questions ask you to justify that choice on time-to-value, data sensitivity, and total cost rather than on architecture; the full sourcing comparison is drawn up with the solution draft.
Assess Technical Viability of Proposed AI Solutions
With the problem already mapped to one of the Seven Patterns of AI, ask whether a proven approach exists for that pattern at your scale, what systems it must integrate with, what latency the persona needs, and whether anyone must be told why a decision was made. An explainability obligation — a customer, an auditor, or a regulator entitled to a reason — rules out some approaches before cost is even discussed, so surface it here rather than in Phase 5.
Analyze Data Availability and Quality for Model Training
This is the early sniff test that precedes Phase 2, not a substitute for it. Six questions:
- Does the data exist, and who owns it?
- Can we lawfully use it for this purpose — consent, contract, and privacy rules included?
- Is there enough history to cover seasonality and rare events?
- Are outcomes labeled, or does someone have to label them?
- How many systems must be joined, and is there a reliable key to join them on?
- Is the data still being produced the same way it was three years ago?
Rare-event problems deserve special attention: at a 0.2% fraud rate, a million records yield only about two thousand positive examples. You are not profiling distributions yet — you are confirming that Phase 2 has something worth profiling.
Evaluate Computational Resource Requirements and Constraints
At this gate the question is deliberately narrow: is the compute this solution would need obtainable and affordable at all? Confirm that capacity of the right kind can be had within your budget and your existing cloud or hosting agreements, that data-residency rules do not forbid processing where that capacity lives, and that no required capability is simply unavailable to you. Count the labeling effort here too, because it is paid in people rather than machines. Detailed sizing happens later, when the AI workspace is provisioned in Phase 2. What Phase 1 has to catch is the order-of-magnitude blocker — a use case whose running cost is plainly beyond anything the benefit could justify, or one that cannot lawfully run anywhere you have capacity.
Review Organizational Readiness for AI Implementation
Read four signals: a named executive sponsor who holds budget and will defend the project; a data culture in which decisions already get made on evidence; the skills to run and monitor a model after go-live; and appetite for probabilistic answers. That last one decides projects. An organization that expects software to be right every time will reject a model that is right 92% of the time, however good that is against the human baseline. Ask also who must change their workflow — and whether anyone has asked them.
| Feasibility dimension | Question to answer | Red flag | Evidence needed to clear it |
|---|---|---|---|
| Technical viability | Is there a proven approach for this pattern, at this scale, with this latency? | The use case needs novel research, or a required explanation cannot be produced | A comparable reference implementation plus named integration points |
| Data availability | Does usable history exist, lawfully, in a joinable form? | Key outcomes were never recorded, or consent does not cover this purpose | Named systems, owners, record counts, date range, and a join key |
| Data quality and labels | Are outcomes labeled and consistently captured over time? | Labels are inferred, disputed, or a collection method changed mid-history | A sample review by the data SME and an estimated labeling effort |
| Computational resources | Can the compute this would need be obtained, lawfully located, and afforded? | No capacity of the needed kind is available to you, or run cost plainly exceeds any plausible benefit | An order-of-magnitude cost range plus confirmation that capacity is obtainable |
| Organizational readiness | Who sponsors it, who operates it, who accepts probabilistic output? | No named sponsor, or users are measured in a way that punishes model use | Signed sponsorship, an operating owner, and user-group engagement |
| Alternatives compared | Would rules, redesign, reporting, or a vendor product do this? | The AI option was never compared to anything | A short written comparison with the reason each alternative was rejected |
Deliver a Recommendation, Not a Verdict
The artifact is a short feasibility memo carrying three things: a recommendation (proceed to Phase 2; proceed with stated conditions; do not proceed now; solve this without AI), a confidence level, and a named list of unknowns with how each would be resolved and what resolving it would cost. A screen that says "feasible" and lists no unknowns is not credible at this stage — Phase 1 cannot know what Phase 2 exists to discover.
"Not feasible now" is a legitimate and valuable outcome. It returns budget to the portfolio and it usually names the precondition: instrument the process, start capturing outcome labels for two quarters, resolve consent, consolidate the two source systems. High-value candidates that fail the screen are not rejected — they are sequenced behind the data-engineering work that would make them viable, which is exactly the judgment the exam rewards.
An operations director asks for a machine learning model to decide which purchase orders require manager approval. The current process applies 14 documented conditions consistently and is correct about 98% of the time. What should the project manager recommend?
A demand forecasting initiative has three years of sales history, but promotions and stockouts were never recorded. The sponsor wants a feasibility answer this week. What should the project manager deliver?
A claims triage model is technically proven, the data is strong, and the compute cost is modest. However, claims handlers are measured individually on decision accuracy, and leadership has said publicly that claims decisions must be right every time. Which feasibility concern is most significant?