AI MVP · VALIDATION PLAN

AI MVP validation plan: what to prove before production

An AI MVP is not a demonstration of every future feature. It should retire the most expensive uncertainties by separating user value, data fitness, AI quality, end-to-end workflow behaviour, and operational readiness.

GUIDE · UPDATED August 5, 2026

01 · DIRECT ANSWER

Direct answer

Before production development, an AI MVP must answer four independent questions: does the user need the workflow, are the data fit for purpose, is quality acceptable given error severity, and does the full path work in the real process? If any answer remains an assumption, the next stage should test that risk rather than scale the product.

02 · CRITERIA

Criteria before development

Apply these criteria to the specific workflow, data, and cost of error. They are not a universal readiness checklist.

01

Falsifiable hypothesis

The user, action, and observable signal that would support or reject value are stated. A feature list is not a hypothesis.

02

Representative examples

The evaluation set reflects real workflows, rare difficult cases, and unacceptable errors; provenance and data rights are known.

03

One vertical path

A user can complete the core workflow from input to outcome. Manual operations are acceptable when explicitly labelled and included in the decision.

04

Comparable alternative

The current process or a simpler option is recorded. Without it, the team cannot tell whether the MVP improves the task or merely looks novel.

05

Continuation rule

Before the pilot, define which observation will lead to further investment, a changed hypothesis, or a stop.

03 · PROCESS

Validation sequence

  1. 01

    Separate the risks

    Map uncertainty across user value, data, model quality, UX, integrations, security, and economics.

  2. 02

    Test the critical component

    If model quality is unknown, start with an evaluation set; if the workflow is unknown, start with a prototype or manual simulation.

  3. 03

    Build a thin vertical slice

    Connect interface, data, AI, and outcome in the smallest scope while logging errors and human intervention.

  4. 04

    Run a bounded pilot

    Choose a defined audience and observation period, and avoid silently changing the workflow version during the test.

  5. 05

    Decide from evidence

    Compare results with the original hypothesis and decide separately what to scale, redesign, or stop.

04 · DECISION

Decision table

Choose the validation format from the largest uncertainty, not from the desired completeness of the product.

What is unknownMinimum testWhat it reveals
Whether users need the workflowInterviews, prototype, or manual serviceReveals behaviour and value before a complex system is built.
Whether usable data and rights existData audit and sampleSeparates available material from assumptions about a future dataset.
Whether AI can perform the taskOffline evaluation on a held-out setSupports comparison and analysis by error type.
Whether the end-to-end process worksLogged vertical MVPExposes UX, integration, and operational failures.
Whether production investment is justifiedBounded pilot against the baselineConnects product quality with an actual work outcome.

05 · MEASUREMENT

What to measure

Record the baseline workflow and collection method first. Then compare equivalent scenarios without presenting a planned target as measured impact.

01

Workflow value

Target-task completion, repeat use, abandonment, and qualitative reasons for behaviour within the selected user group.

02

Quality and error severity

Evaluation-set results by category, including critical failures, abstentions, and cases requiring human intervention.

03

End-to-end operation

Latency, technical failures, integration success, and workflows that cannot complete without hidden manual work by the team.

04

Baseline comparison

Change in time, manual steps, quality, or throughput relative to the process recorded before the pilot.

05

Operating cost

Model, infrastructure, integrations, human review, support, and error correction are counted together, without claiming ROI before a real comparison.

06 · BOUNDARIES

Risks

01

The polished-demo trap

Curated examples hide the distribution of real requests and do not expose abstentions, latency, or manual corrections.

02

Evaluation leakage

If control examples shape tuning and are then reused as an independent test, the result becomes optimistic.

03

Unrepresentative pilot

A convenient early-user group or cleaned dataset may not reflect future load and workflow diversity.

04

Hidden manual operations

Manual data preparation and answer correction can aid learning, but must remain visible in cost and scaling plans.

WHAT CANNOT BE CLAIMED

Limitations of the conclusion

  • An MVP validates a bounded workflow; it does not prove that the architecture is ready for all future load.
  • A small pilot does not provide a universal forecast of business impact; results depend on the audience and process.
  • External-model quality can change after an update, so versioning and repeat evaluation are mandatory.
  • For high-impact decisions, an MVP does not remove security, access control, legal review, or human oversight.

07 · PORTFOLIO

Related projects and honest stage

The related projects are at technical or local MVP stage. These labels do not mean production operation or validated commercial impact.

Current stage: Technical MVP

Marketing Optimisation

The technical MVP implements one end-to-end CVM reactivation workflow and has been tested on synthetic data. The next step is a design-partner pilot with real CRM data; a production replacement for the department and commercial impact have not yet been validated.

Open project review

Current stage: Functional local MVP

Gadalnya

The local MVP implements the core flow, server API, validation, and a demo response in local demo mode. A public launch, real payments, and production LLM configuration have not yet been completed or verified.

Open project review

08 · METHODOLOGY

Primary sources

These sources support the methodology and definitions. They do not validate AI SaaS Solution project outcomes.

  1. Research · Google Research / IEEE Big Data

    The ML Test Score: A Rubric for ML Production Readiness and Technical Debt Reduction

    Open source
  2. Research · Google Research / NeurIPS

    Hidden Technical Debt in Machine Learning Systems

    Open source
  3. Official publication · NIST

    Artificial Intelligence Risk Management Framework (AI RMF 1.0)

    Open source