Why Does AI Strategy Stall in the Executive Meeting?

AI strategy stalls when executive support is not converted into explicit ownership, authority, evidence, value, and a review date.

Executive sponsorship creates value only when it becomes an operating decision

Standfirst and thesis

AI strategies often lose momentum across executive meetings. Leaders endorse artificial intelligence and approve pilots, yet leave consequential questions unresolved: Which business decision should change? Who owns the result? What may the system do? What evidence is sufficient? How will value and harm be measured? When will the initiative be scaled or stopped?

That pattern is leadership drift. My thesis is more precise: it is not an abstract cultural weakness but a governance and decision-design defect. Without explicit commitments, innovation teams keep experimenting, risk teams add reviews, and business units preserve existing workflows. No one has authority to redesign the system or accountability for its economic outcome.

The remedy is not louder sponsorship. It is a compact management instrument that makes the missing decisions visible. I call it the AI Decision Contract.

What changed: AI moved from content to consequential work

During the first wave of generative AI, many enterprise uses were advisory: summarize a document, draft an email, find information, or explain a dataset. Agentic systems change the stakes because they can use tools, navigate applications, update records, and initiate downstream work.

OpenAI’s Enterprise Signals reported that, as of June 2026, Codex produced 64% of the combined Codex and ChatGPT output tokens among OpenAI enterprise customers. The company cautions that token volume is an imperfect value proxy because longer agentic tasks generate more output. The figure does not mean 64% of enterprise work is autonomous or valuable; it shows that delegated activity can expand faster than authority and accountability.

A June 2026 survey of 750 executives, published by Gartner C-Level Communities, found that 44% of organizations were exploring agentic AI, 30% were piloting it, and only 9% reported operationalization in business processes. Governance was the leading selected scale barrier at 17%, followed by data quality and skills at 15% each. These self-reports do not prove causality, but they describe an execution gap technology alone cannot close.

The mechanism: how drift becomes organizational behavior

Leadership drift begins when a strategic choice becomes an adoption problem. Executives ask whether employees are ready before agreeing what the company intends to change. Responsibility moves to the CIO, an innovation lab, or a vendor, while process owners retain existing targets, budgets, incentives, and decision rights.

The portfolio fills with low-conflict pilots. A demonstration can impress without forcing decisions about pricing, staffing, production priorities, or risk appetite. Committees count use cases and users while avoiding which approval should disappear, who may override the agent, or which P&L line must improve.

Research by Morgan Blangeois and Thomas Roulet gives this dynamic a useful name. Their Harvard Business Review article, based on 23 interviews and three workshops across 11 European IT services firms, suggests that the brake may sit within leadership. The study is small, qualitative, sector-specific, and available through a paid subscription, so it should support—not anchor—a broad claim. Its diagnostic value is showing that apparent agreement can coexist with no executable decision.

The AI Decision Contract converts strategic ambiguity into six reviewable commitments before production authority is granted.

The AI Decision Contract

A use case describes what AI might do. A Decision Contract states what the organization will do differently if the system works. It is not a legal document and should not become a thirty-page compliance form. It is a one-page agreement completed before a material pilot receives production access, operational authority, or significant funding.

Contract field	Executive commitment
1. Strategic question	Define the consequential decision the initiative must improve. Replace “increase productivity” with a question that specifies the business condition and unacceptable trade-off.
2. Named owner	Assign one business executive who owns the outcome and the scale, redesign, or stop decision. Contributors can be many; accountability must remain singular.
3. Workflow and authority change	State which work step changes, what the AI may observe, recommend, draft, approve, execute, or escalate, and which actions remain human-only.
4. Evidence and assumptions	Record approved sources, data limitations, model conditions, unresolved gaps, dependencies, and assumptions that must remain true for action to be justified.
5. Value metric	Establish the baseline, operational KPI, financial value path, and counter-metrics for safety, quality, compliance, or customer harm. Adoption is an input, not the outcome.
6. Review and stop date	Set the date and thresholds for scaling, revising, pausing, or retiring the initiative before enthusiasm becomes an indefinite budget.

This design reflects the evidence discipline in NIST SP 1353, an August 19, 2026 initial public draft. Its three notional cybersecurity use cases map artifacts, assumptions, evidence gaps, current states, and target outcomes. NIST says they are not assessment or assurance methods. Structured evidence makes a recommendation inspectable; it does not transfer accountability.

Manufacturing example: predictive maintenance meets MES scheduling

Consider a predictive-maintenance agent combining sensor history, failure records, work orders, spare-parts data, and the current schedule. A pilot may optimize prediction accuracy and display a risk score. Drift begins when nobody decides what that score may change.

Can the agent create a maintenance order, reserve a technician, or move a lot in the MES? Who bears a false intervention or an ignored warning? Without answers, the model may perform well while the workflow remains inert or teams improvise inconsistent actions.

Under an AI Decision Contract, the strategic question could be: Under which evidence and operating conditions should predicted failure risk justify intervention? The equipment-reliability leader owns the outcome, while the production owner retains authority over material schedule disruption. The agent may gather approved evidence, draft a work order, check parts, and simulate schedule options. It may not stop the tool, release a purchase order, or move a committed customer order without named approval.

The value metric is avoided unplanned downtime net of intervention and false-alarm cost. Schedule stability, product quality, and safety remain counter-metrics. A ninety-day review requires leaders to scale, narrow, redesign, or stop the workflow against predefined thresholds. This is bounded autonomy: the agent can advance work, but authority and accountability do not disappear.

My perspective

Leadership drift persists because ambiguity is temporarily convenient. A broad AI ambition lets every executive support innovation without reallocating budget, changing a workflow, accepting a risk threshold, or owning a result. But once an AI system can take action, executive ambiguity becomes machine-speed operational behavior. The system either acts under an unclear mandate or waits in a review loop that destroys its value.

Training cannot repair a missing decision, and a stronger model cannot resolve disputed ownership. Leadership requires sharper, reversible commitments reviewed against evidence, not greater enthusiasm.

Three implications for enterprise leaders

1. Agentic AI raises the cost of indecision. A chatbot can produce an unused answer. An agent with system access can change records, create orders, or trigger downstream work. Unclear authority becomes a control risk as well as a strategy delay.

2. Govern the decision, not only the model. The same model may be low-risk when drafting a report and high-impact when changing a production schedule. Governance should follow the action, affected stakeholder, reversibility, evidence requirement, and escalation path.

3. Separate activity from value. Tokens, active users, automated tasks, and pilot counts indicate engagement. They do not establish economic value. Leaders need a defensible chain from evidence to decision, workflow change, operational KPI, and financial or risk outcome.

Limitations and counterargument

A reasonable objection is that Decision Contracts could slow experimentation. That will happen if every sandbox test receives production-level paperwork. The answer is proportionality. A reversible assistant that drafts meeting notes needs a lightweight record: owner, learning hypothesis, data boundary, success signal, and expiry date. Requirements should deepen as access, autonomy, and impact increase.

The evidence also has limits. The HBR study covers 11 European IT services firms. Gartner’s survey is self-reported. BCG’s July 2026 CEO research reports associations, not experimental causation; only 14% of respondents clearly defined P&L impact for every AI initiative, and fewer than one-third funded process and skills redesign. OpenAI observes its own customers, token use is not ROI, and NIST SP 1353 remains a draft. The sources reveal an execution problem, not a universal failure rate or guaranteed remedy.

Five actions for the next executive meeting

1. Replace one demonstration with one decision. Ask which business decision will change, why it matters now, and what current work will be removed if the new workflow succeeds.

2. Select a few end-to-end workflows. Concentrate attention on domains with measurable value, meaningful process change, and a real business owner rather than maintaining a long inventory of disconnected ideas.

3. Issue the AI Decision Contract. Require the six commitments before production access, material funding, or autonomous action. Keep it short enough to use and specific enough to expose disagreement.

4. Instrument the decision path. Log evidence, assumptions, approvals, exceptions, actions, reversals, and outcomes. Monitor business value and counter-risks, not only model accuracy or adoption.

5. Make the review date consequential. Scale initiatives that meet the contract, revise those that produce valuable learning, and stop those that cannot demonstrate a defensible route to value.

Conclusion

AI strategy stalls when executive discussion produces no durable allocation of ownership, authority, evidence, and value. Leadership drift is therefore not solved by another inspirational message or technology roadmap. It is solved when leaders make the consequential choices that allow an organization to redesign work safely.

The AI Decision Contract does not eliminate uncertainty. It makes uncertainty explicit, assigns ownership, and sets a date for the next decision. That bridges supporting AI in principle and governing it in practice.

FAQ 1 — What is AI leadership drift?

ANS: AI leadership drift is the gradual replacement of consequential strategic choices with supportive but noncommittal discussion. Leaders may endorse AI while leaving ownership, authority, workflow changes, value measures, and risk thresholds unresolved.

FAQ 2 — How is leadership drift different from employee resistance?

ANS: Employee resistance concerns willingness or ability to use a system. Leadership drift occurs when executives have not made the organizational decisions required for the system to create value. Both may coexist, but they require different interventions.

FAQ 3 — What is an AI Decision Contract?

ANS: It is a lightweight management agreement that records the strategic question, accountable owner, workflow authority, evidence and assumptions, value and guardrails, and the date for a decision to scale, redesign, pause, or stop.

FAQ 4 — How should manufacturers apply it to AI agents?

ANS: Manufacturers should tier authority by operational impact. Reversible, low-risk actions can run within defined limits; actions affecting safety, quality, schedules, inventory, or cost require evidence, named approval, auditability, and a recovery path.