01 · DIAGNOSTIC FRAMEWORK

Turn the symptom into a testable decision.

A demo proves that a model can produce an answer. A PoC must prove that a specific decision improves under real constraints. The sequence below is designed to preserve definitions, expose alternative explanations and lead to a decision that can be validated.

  1. Name the decisionState who acts, what choice changes, when it is made and what happens when the system abstains.
  2. Freeze acceptance criteriaSet the primary outcome, minimum lift, error limits, cost, latency and review burden before the experiment.
  3. Build a representative evaluation setSample routine, difficult, rare and adversarial cases from the intended environment; preserve a hidden holdout.
  4. Compare real baselinesInclude current human process, simple deterministic rules and the cheapest credible model.
  5. Test operations, not just outputsExercise permissions, recovery, monitoring, escalation and version rollback.
02 · EVIDENCE

Ask for the minimum data that can change the answer.

Begin with read-only access and a field-level purpose. Reconcile samples before scaling extraction, preserve event time and source provenance, and record missingness rather than silently filling it.

Task evidenceRepresentative inputs, expected decisions, supporting evidence and explicit ambiguity.
Operational contextUsers, access boundaries, latency budget, frequency, integrations, review and exception path.
Evaluation labelsExpert rubric, pairwise preference, objective outcome, disagreement and confidence.
EconomicsModel and tool cost, human review time, avoided work, error cost, implementation and maintenance.
03 · PROOF OF CONCEPT

Validate the claim before changing the operation.

Use a promotion gate, not a presentation

Create a test harness that logs model version, prompt, retrieved evidence, tool calls, output, reviewer decision, latency and cost. Evaluate a locked candidate on a hidden set, review failures by severity and segment, and run a limited shadow workflow. The promotion decision should be reproducible from stored evidence and reversible if later monitoring detects regression.

04 · FAILURE MODES

What makes the diagnosis look right and still fail.

  • Choosing the model firstThe experiment becomes a search for a use case that flatters the selected technology.
  • Evaluating fluent textStyle can hide unsupported claims. Score the decision, evidence and downstream consequence.
  • Testing only clean examplesProduction volume concentrates in edge cases, missing fields and conflicting instructions.
  • Changing the rubric mid-testMoving success criteria makes every result look promising.
  • Skipping the abstain pathA safe system must know when to request evidence or human review.
05 · SOURCE TRAIL

Primary and official references

These sources define the measurement, control or operating context. They do not replace validation on the company’s own data.

  1. NIST, AI Risk Management Framework
  2. NIST, AI RMF Playbook
  3. UK ICO, Guidance on AI and Data Protection
06 · FAQ

Questions enterprise teams ask first.

How large should an evaluation set be?

Large enough to estimate the decision by important segment and failure severity. Start with power or precision requirements, not an arbitrary count.

What is the difference between a pilot and a PoC?

A PoC validates a causal or technical claim in a controlled scope. A pilot validates the operating workflow with real users and live consequences.

Should human reviewers know which model produced the answer?

Blind review reduces brand and novelty bias when the task allows it.

When is a PoC ready to scale?

When it passes pre-set outcome and guardrail thresholds, has a monitored failure path and the operating owner accepts the residual risk.