Is This AI Strategy Real?

The easiest mistake in AI strategy is to take visible activity for progress. Companies launch internal chatbots, functions run pilots, vendors show agent demos and board decks list dozens of use cases, so everyone agrees the firm is moving. Sometimes it is. Often the way work actually gets done has not changed at all.

Fake AI strategy costs more than ordinary software theatre. Besides wasting budget, it teaches executives the wrong lesson about the technology: enthusiasm appears in pockets, and disappointment arrives later with no diagnosis attached. A company can describe itself as experimenting while leaving the hard work untouched, which means redesigning the workflow, naming an operating owner, accepting specific risks and agreeing in advance when to stop if the evidence stays weak.

Recent market evidence makes the problem hard to ignore. McKinsey's 2025 State of AI survey reported broad regular AI use, while many organizations were still in experimentation or pilot stages and only 39 percent reported any enterprise-level earnings impact. A 2025 CEO study from IBM reported that only 25 percent of AI initiatives had delivered expected return on investment and only 16 percent had scaled enterprise-wide. Boston Consulting Group (BCG) in 2026 found that only 5 percent of surveyed companies were generating measurable value from AI at scale, while 60 percent were not achieving what it called material value. Deloitte's enterprise generative-AI survey found enthusiasm still running ahead of full scaling, with many experiments unlikely to scale in the near term. Read together, the figures describe a technology spreading faster than the management practices around it.

The question for the board is which operating system the initiative changes, and what proof of that change will survive contact with the P&L.

Five conditions an AI initiative should meet

An initiative is more likely to be strategy than theatre when it meets all five.

1. It names the workflow it will change

The initiative should identify a recurring business workflow, rather than a capability, a model or a department's enthusiasm. "Use AI in sales" commits nobody to anything. "Reduce the time between qualified inbound lead and first tailored proposal from three days to six hours without lowering win-rate quality" makes the team specify inputs, handoffs, decision rights, data access, approval thresholds and how the result will be measured.

Workflow redesign keeps showing up in the better research on AI value. McKinsey's 2025 survey on organizations rewiring for value found workflow redesign to be the attribute with the biggest effect on reported earnings impact from generative AI. BCG's AI-at-work research made a similar argument: tools produce value once the work around them has been redesigned.

2. It has an accountable owner

Someone who controls the business process being changed has to own the result. A centre of excellence can advise, IT can enable, and legal and security can set control points, but if nobody accountable for the process owns the outcome, the project is structurally weak. These projects usually have a sponsor instead. A sponsor likes the idea; an owner can change staffing, metrics, handoffs, incentives and stop conditions. Without that authority, the project ends up demonstrating what might be possible in an organization that never has to move.

3. It defines its evidence before the pilot starts

The team should say in advance what would prove or disprove the thesis. Adoption, seat activation and prompt volume settle nothing on their own, and time saved is incomplete if quality, rework, customer trust, risk and revenue are ignored. A serious evidence plan has a baseline, a target, a measurement window, a quality check, a cost denominator and a decision rule. The cost denominator needs particular care because AI systems now bring governance, data, security, review, observability and change-management costs with them. A tool that saves time while creating a new review burden may still be worth having, provided the burden appears in the calculation.

4. It can be managed after launch

Running the system in production requires named permissions, logs, a review cadence, exception routing, rollback rules and some way to discover whether the agent or assistant is still working inside its intended scope. Demos rarely face any of this. In a demo the workflow is narrow, the data is curated and failures stay private; in production the system meets edge cases, stale records, user shortcuts, security limits and unclear accountability, and the real cost of adoption starts to appear.

AI Needs a Management Layer argued that agent registries, observability logs, ownership controls, provenance rules and transparency obligations are becoming an operating layer in their own right. If an initiative cannot be governed through that layer as ordinary work, it has not yet left the pilot stage.

5. It justifies its claim on attention

The initiative should explain why this workflow deserves scarce attention now. Much of AI strategy consists of turning down tempting use cases, and the strongest roadmaps are short, concentrated on workflows where the upside is large enough to justify moving the organization. The cost of saying yes may be executive time, change fatigue, security review, data cleanup, vendor dependence or integration work. It may also be the right to ask employees to alter a process that already functions. A mediocre AI project can crowd out a boring operational fix that would have produced more value.

A screen to use before approval

Executives can apply a simple screen before approving the next AI initiative.

Why plausible projects survive

The most dangerous theatre is the plausible project in which every function sees its own agenda, from savings in finance to content velocity in marketing, while vendors see expansion and the board sees responsiveness to the market. Obviously silly demos get caught; this kind does not, because nobody wants to be the person who slows it down and each initiative is individually defensible. An explicit pass-fail screen changes the question in front of the approving executive from "is this interesting?" to "does this meet the conditions for managed operating change?"

What to do this week

Skip the grand AI strategy refresh. Take the ten most visible AI initiatives in the company and score each against the five conditions, using three labels only: pass, revise or stop. The scoring should ignore how much anyone believes in AI and ask only whether the company will change how the work is done, measured, owned and governed.

Pass means the initiative has a named workflow, an accountable owner, an evidence plan, a management layer and a case for its opportunity cost. Fund it properly and protect it from distraction.

Revise means the initiative is plausible but missing one or two conditions. Rewrite the charter before spending more money.

Stop means the initiative is mostly signalling. Kill it politely, record the reason, and move the freed attention to a workflow with a better claim.