The AI
Decision Gap

The decision to build is being made without evidence

AI adoption is accelerating across every industry. But the decision that precedes adoption, which AI initiatives are actually worth committing resources to, is being made informally, politically, and without a systematic signal framework.

By the time an AI initiative formally enters a delivery structure, large parts of the admissibility decision have already been socially and operationally committed. Governance arrives too late to be useful. Resources are locked before the question "should we build this?" has ever been rigorously answered.

I have spent over 10 years leading technology adoptions and governance initiatives inside enterprise and mid-size organizations, and what I am watching happen in AI is the same pattern repeating, faster and at greater scale. This paper is the result of that experience, combined with research across seven practitioner profiles, institutional sources including Harvard Data Science Review and the Alan Turing Institute, four major industry frameworks, and direct conversations with decision leaders building in this space. It is not a conclusion. It is a hypothesis under active validation.

This is not a technology problem. It is a decision intelligence problem. And the data shows it is systemic.

These numbers share a common root cause. Organizations are committing to AI initiatives before answering the question that determines whether commitment is warranted. The decision layer is missing.

This became a decision problem, not a delivery problem

The dominant narrative in enterprise AI is a delivery narrative: how do we build faster, deploy more reliably, scale more efficiently? Tooling, platforms, and methodologies have proliferated to answer this question.

But a quieter, more consequential problem has emerged upstream of delivery entirely. Organizations now have more AI ideas than they have capacity to evaluate. Portfolio pressure, mandate anxiety, and FOMO-driven initiative creation are producing a backlog of AI experiments that nobody has the signal framework to sort.

This is the gap. Not the absence of delivery tooling. There is an abundance of it. What is missing is a pre-project decision intelligence layer that answers the question before resources are committed.

The governance frameworks that exist (PMI's AI governance guidance, AWS CAF-AI, Google PAIR, NIST AI RMF) all assume the decision to build has already been made. They govern execution. Nobody has built the systematic layer that governs the decision itself.

Governance is perceived as slow.
Signals are data.

The word "governance" carries organizational baggage. It implies process, audit, compliance, overhead. Decision leaders resist it, not because they reject rigor, but because past governance frameworks have behaved like brakes, not compasses.

The reframe is this: what organizations need before committing to an AI initiative is not a governance committee. It is a set of signals: data points that answer whether an idea has the definition, fit, and readiness to justify resource commitment.

Signal intelligence is fast. It is data-driven. It produces a defensible answer, not a committee recommendation. And it happens before the social and political commitment that makes course-correction expensive.

The decision layer is the missing layer. GreenfieldworkAI is building it as an open, data-driven research platform and a free tool available to any decision leader who needs it.

What we are testing

This is an active research project. We are not asserting conclusions. We are gathering evidence. Three hypotheses are being validated through practitioner interviews, podcast analysis, industry framework review, and platform data.

Voices from the field

The following validation has been gathered through direct practitioner conversations and published practitioner research. All contributors are cited by role until explicit permission to name them has been confirmed.

These are not practitioner opinions. They are the academic and policy establishment: the UK's national AI research institute in partnership with Harvard Data Science Review, naming the same failure this research documents from the field. The alignment between institutional framing and practitioner evidence is independent. Neither cites the other.

Four patterns repeated independently across practitioners

The following failure modes appear across multiple independent practitioner sources. I am treating repeated independent observation as pattern evidence, not proof, but signal.

• Governance arrives too late. Across four independent sources (an AI governance founder, an enterprise PM, a product educator, and AWS institutional guidance) the same failure is described: governance is either absent at commitment, retrofitted during delivery, or treated as a post-deployment audit. None describe governance as a pre-commitment gate. That is precisely the gap this research addresses.
• No framework for saying no. Three practitioners independently describe the absence of a structured mechanism for rejecting AI ideas. Each has built their own workaround: structured intake, hypothesis validation, compliance guardrail at intake. They are solving the same missing infrastructure: a formalized way to say "this AI idea should not proceed."
• The decision is made before governance is applied. Three sources describe the same sequencing failure: the AI commitment is made, often in a meeting, through executive enthusiasm, or through mandate, before any governance infrastructure exists to evaluate it. The decision and the governance run in the wrong order.
• The PM role is shifting from execution to governance design. Four independent practitioners describe the same structural shift: senior PM value is moving away from delivery execution and toward decision architecture, governance design, and strategic alignment. This validates the decision leader as the primary audience, someone whose role is becoming more consequential, not less, as AI reduces coordination overhead.

What the established frameworks tell us

The major AI governance frameworks (PMI, AWS CAF-AI, Google PAIR, NIST AI RMF) collectively confirm the problem without solving it. Each framework assumes a decision to build has been made. None provides a systematic method for making that decision in the first place.

The gap is consistent across all four. GreenfieldworkAI operates at the layer these frameworks assume is already handled, and isn't.

Five signals that determine whether an AI initiative is worth building

Based on practitioner research and framework synthesis, I have identified five signals that collectively determine whether an AI initiative has the definition and fit to justify resource commitment. These signals are captured through a structured intake conversation, not a form, not a committee review.

• Problem clearly articulated: Can the initiator describe the problem in plain language? Vague problems produce vague solutions. An AI initiative without a clearly stated problem is not ready.
• Target user identified: Who has this problem? If the answer is "everyone" or "the business," the signal fails. Specificity of user drives specificity of solution.
• AI justification credible: Is AI genuinely the right tool, or would a simpler solution achieve the same outcome? Mandate-driven AI has a place, but it must still demonstrate fit.
• Success metric specific: What does success look like, measured how, by when? Initiatives without a specific success metric cannot be evaluated or defended.
• No unmitigated risk flags: Data risk, compliance exposure, key-person dependency. If risk flags exist and are unaddressed, the initiative is not ready to advance.

These five signals produce a Readiness Score (1-5). Combined with AI Fit classification (Strong / Moderate / Weak) and Work Classification (Strategic / Operational / Productivity), the PMO Director has a defensible, evidence-based portfolio view, not a political consensus.

Free. Open. Built by practitioners.

GreenfieldworkAI is a pre-project decision intelligence platform I built to operationalize the signal model described in this paper. It is built by a practitioner who has lived inside the problem, not by engineers optimizing a feature roadmap.

The platform is available free. No sales process. No paywall. If it becomes useful to enough decision leaders, I will ask for voluntary contributions to sustain and scale it. The research and the tool will remain open regardless.