There Is No Standalone AI Strategy

Start with the single sentence this entire module rests on:

"There is no standalone generative AI strategy. There is only business strategy, strengthened by generative AI."

AI is not a strategy. It is a capability — like cloud, like mobile, like the internet before it — that must be aimed at the priorities your business already has: growth, cost, customer experience, and risk. The moment "doing AI" becomes the objective, the organization optimizes for activity (pilots launched, seats bought, demos shown) instead of outcomes (revenue moved, cost removed, a metric improved).

This is the Wharton / BCG Henderson Institute framing, and it is the most important strategic correction a leader can make. Generative AI won't create value on its own (Knowledge@Wharton, 2025). Value comes from pointing it at a specific problem you were already trying to solve — and then measuring whether the number moved.

The test for any proposed AI initiative is therefore not "is this AI?" but "which existing priority does this advance, and by how much?" If you cannot answer in one sentence, you do not yet have an initiative — you have a science project.

If the reframe were obvious, the data wouldn't look the way it does. The defining story of 2025–2026 is a widening gap between AI adoption (near-universal) and AI value (concentrated in a tiny minority).

Note what these findings agree on: the bottleneck is not model quality. MIT NANDA traced the 95% failure rate to a "learning gap" — workflow, integration, and adoption — not to weak technology. The technology is broadly available to everyone; it is commoditizing. Winners differentiate on how people, processes, data, and governance adapt around the technology — which is precisely what a business strategy governs and a "technology strategy" ignores.

The headline to carry into the boardroom: "Everyone has AI; almost nobody has AI value yet." The companies closing that gap did so by refusing to treat AI as the goal.

When the objective is "do AI," the predictable result is a sprawl of small, technology-led experiments that look like progress and deliver nothing measurable. McKinsey named the mechanism: the gen AI paradox — roughly 80% of firms have deployed generative AI, yet roughly the same share report no material earnings impact.

Why? Because activity-led AI gravitates toward shallow, horizontal copilots (a chatbot here, a drafting assistant there) that scale fast but deliver diffuse, unmeasurable gains spread thin across the organization. The transformative value sits in deep, vertical workflows in core functions — and those stay stuck: by McKinsey's account, ~90% of vertical use cases remain in pilot mode.

The pattern of a value-less pilot is recognizable from across the boardroom table:

• It started from a tool or a demo ("we should try this model"), not a problem.
• It has no business owner with P&L accountability — only a technical sponsor.
• It has no target number and no baseline — so "success" is undefinable and therefore unfalsifiable.
• It measures vanity metrics — seats, logins, prompts run — that can be inversely correlated with value.

These pilots are not failures of technology. They are failures of framing. The cure is not a better model; it is a better question, asked before a dollar is spent.

The discipline that separates the 5% from the 95% is almost boringly simple: start from a business problem and a target number, not from the technology.

Invert the usual order. The losing sequence is technology → search for a use case → hope for value. The winning sequence is:

1. Name the business problem in plain language, tied to a strategic priority. "First-contact resolution in customer support is too low and is dragging CSAT."
2. Set a target number and a baseline. "Raise first-contact resolution from 62% to 75% within two quarters." No baseline, no initiative.
3. Name the accountable business owner — a P&L leader, not the AI team.
4. Only then ask whether AI is the right capability — and if so, whether to buy, partner, or build it.

This is the "business problem, not 'do AI'" move at the heart of every credible 90-day plan: pick a problem, attach a number, and assign an owner before you touch a tool. It is also why the strongest programs scope tightly. JPMorgan runs 450+ AI use cases in production with benefits growing ~30–40% year over year — not because it "did AI," but because each use case is scoped to a specific operational outcome with a named owner. Scoped, KPI-anchored deployment is the through-line of every program that actually moves the P&L.

A business problem framed this way naturally survives contact with reality: if the number doesn't move, you reshape or kill it — fast, and without ego, because the goal was never "do AI."

Once the problem is named, anchor the initiative to a single specific operational or financial KPI in one of your four strategic buckets. "Improve productivity" is not a KPI. "Cut average claims-handling time from 9 days to 5" is.

Measure value in three tiers, and don't stop at the first one (the canonical trap):

1. Adoption / usage — action counts. Necessary, never sufficient.
2. Workflow efficiency — time saved, throughput, error reduction.
3. Business / P&L impact — revenue, cost, and ultimately % of EBIT attributable to AI, the line McKinsey found separates high performers.

And track quality, not just volume — the hard-won lesson of the Klarna case. Klarna's AI assistant looked like a triumph on volume metrics (it handled roughly two-thirds of customer-service chats and cut resolution time dramatically), but CSAT deteriorated and the company rehired humans by mid-2025. The CEO's verdict: "We focused too much on efficiency and cost… lower quality… not sustainable." Volume metrics masked a quality collapse. The anchor KPI must include the quality dimension of the priority you chose — or you will optimize yourself off a cliff and call it success along the way.

You don't need to evaluate models to govern AI well. You need to ask the questions that force every proposal back onto business ground. Use these in any AI review:

• On framing: "Which strategic priority does this advance — growth, cost, CX, or risk — and what is the target number?" If there's no number, there's no initiative.
• On ownership: "Who is the accountable business owner, and is it a P&L leader or the AI team?" If only a technologist owns it, it's a science project.
• On the baseline: "What's the current value of that KPI today, and how will we know we moved it?" No baseline, no proof.
• On the choice: "Are we redesigning a workflow, or just bolting a tool onto a broken one?" Bolting AI onto legacy processes is how the 95% got there.
• On a vendor claim: "You say it's '90% accurate' — accurate in what way, what do the failing 10% look like, and who owns them?" Never accept a headline number at face value.
• On the kill criterion: "If the number hasn't moved by [date], do we reshape or stop?" Pre-committing to a kill date is what prevents zombie pilots.

These questions cost nothing and do most of the strategic work. They convert "doing AI" into "running a business initiative that happens to use AI" — which is the only kind that reaches the P&L. Re-asking them quarterly, and reallocating away from initiatives that aren't moving their number, is the cadence that compounds value.