How to Run AI Discovery: Baselines, Feasibility, and a Spec in Weeks

Most AI projects skip discovery or mistake a demo for it. They fund a build on an estimated benefit, discover in month two that the data is locked, and learn in month six that the accuracy users need was never reachable. Discovery is the short phase that answers whether and how before the expensive part: baseline measured, correctness defined, feasibility tested with evaluation, data and integration verified, cost modeled, specification written, decision made. This guide covers how to run it, drawing on FISTA Solutions' AI enablement practice. The specification it produces is in how to write an ai spec and the stage that follows in the ai pilot checklist.

What does discovery answer?

Question	Activity	Deliverable
How big is the problem?	Baseline measurement from operational data	Baseline report
What does correct mean?	Working sessions with domain experts on real cases	Correctness definitions; first golden dataset
Is it feasible?	Candidate approaches evaluated against the dataset	Evaluation results by category
Is the data usable?	Access, permissions, quality verified with samples	Data readiness report
Can it be integrated?	Target systems, APIs, and permissions inventoried	Integration inventory
What will it cost?	Build and run cost modeled at expected adoption	Cost model
What could go wrong?	Risks identified with controls	Risk register
Should we proceed?	Evidence assembled	Specification and decision

How do you measure the baseline?

From operational systems, not interviews: volumes, cycle times, cost per unit, error and rework rates, backlog, and satisfaction for the target process over a representative period. Segment by the categories the AI would handle differently. The baseline decides whether the problem is worth solving and becomes the reference for every later value claim. Case structure is in ai business case template.

How do you define correct?

In working sessions with the people who do the work today, using real cases: what a correct output is, what an acceptable error is, which categories matter, and what must never happen. Capture the results as labeled examples, which become the first golden dataset, and as written acceptance criteria by category. This is the step most often skipped and the one that determines whether the project can succeed. Dataset design is in what is a golden dataset.

How do you test feasibility?

Run candidate models and approaches, hosted and open, retrieval variants, prompt strategies, against the first golden dataset with calibrated grading, by category, and compare with the draft thresholds. The result shows which categories are reachable, which need narrowing, and what each approach costs per correct output. A demo shows possibility; evaluation shows reachability. Evaluation practice is in the ai evaluation checklist.

How do you verify data and integration?

Obtain sample data from each source with permissions confirmed, assess quality on the fields the system needs, and document lineage and residency constraints. Inventory the target systems, verify API access and rate limits, and identify where actions will need gates. Discovery that ends with samples in hand and API calls tested has removed the two most common causes of stalled builds. Readiness practice is in the ai data readiness checklist.

How do you model cost and risk?

Model build cost from the specification and run cost at expected adoption including model usage, human review, infrastructure, and maintenance, with routing and caching assumptions stated. Identify risks with controls and owners in a register. Both feed the decision and the staged funding plan. Budgeting is in the ai budget planning checklist and controls in how to de-risk an ai project.

What is the decision at the end?

Proceed to pilot with the specification and funding for the next stage; narrow scope to the categories evaluation showed reachable; or stop because the baseline is too small, the data is unavailable, or the thresholds are unreachable at acceptable cost. Narrowing is the most common outcome and the most valuable; stopping is a success that saved a build. Pilot agreements that follow are in how to structure an ai pilot agreement.

How do you run discovery in a few weeks?

• Week one: baseline extraction; correctness sessions; data access requests; integration inventory.
• Week two: first golden dataset labeled; candidate approaches run; sample data assessed.
• Week three: evaluation results by category; cost model; risk register; specification draft.
• Week four: specification finalized; decision meeting with evidence; staged funding plan.

Timeboxes vary with workflow complexity, but discovery that runs for months has become a pilot without a decision. Specification discipline is in the spec-driven development whitepaper.

What mistakes undermine discovery?

Estimating the baseline instead of measuring it; letting engineers define correctness without domain experts; testing feasibility with a demo on hand-picked examples; assuming data access; skipping integration inventory; and treating stop or narrow as failure. Each reappears as a stalled build. Prioritization upstream of discovery is in how to prioritize ai use cases.

How FISTA Solutions runs discovery

FISTA Solutions runs discovery as a fixed-price phase: baseline measurement, correctness sessions with client domain experts, a first golden dataset, feasibility evaluation across candidate approaches, data and integration verification, cost model, risk register, and a specification with a proceed, narrow, or stop recommendation, all delivered in weeks. The AI enablement practice leads discovery, forward deployed engineers carry the specification into delivery, and AI agents supplies the systems. The record behind the approach is 150+ projects for 50+ companies.

To know whether and how before you fund the build, message FISTA on WhatsApp, or read how to write an ai spec for the document discovery produces.