The Capability Was Never the Problem. The Operating Model Is.

Madelaine Yue, put it as cleanly as I’ve seen it put: “AI doesn’t fix broken systems. It executes them faster.” The line has stayed with me because it names something the enterprise AI conversation has been circling for two years without quite landing on. Capability isn’t the problem in 2026. Capability is everywhere. What enterprises are still missing is the operating model that turns the capability they already own into outcomes their customers can feel.

I’ve started calling that operating model Applied Intelligence, because the gap it closes is the gap between having AI and using it. The distinction matters because the second one is the only one that compounds.

The capability ceiling has been hit

If you lead AI or digital strategy at an enterprise of any size, you’ve likely bought more AI than you’ve organized. A copilot for sales, a summarizer for support, a triage agent for operations, an analytics engine for marketing, sometimes all from different vendors. The capability inventory looks impressive in a board deck. The operator inside the workflow still walks the seams between those systems, and so does the customer.

When MIT NANDA’s 2025 finding that 95 percent of generative AI pilots produce no measurable P&L impact traveled, the headline outran the diagnosis. The pilots didn’t fail because the models were weak. They failed because nothing connected a model’s output to a business process that was already accountable for an outcome. The integration capacity — the people, accountability, data plumbing, operating discipline that turns five components into one experience — was almost never the line item that got funded.

VimalRaj Sampathkumar made an adjacent point  in April: enterprises stall when AI is treated as a series of discrete purchases rather than a capability that has to accumulate. Funding the accumulation is the unglamorous half of the work, and it’s half of what Applied Intelligence is actually about.

The strategy-to-execution handoff is where most AI dies

Most enterprise AI failure modes have been blamed on something. Data hygiene. Governance. Change management. Talent. Each of those is real, and none of them is the primary failure mode I see week to week. It’s the handoff between the team that wrote the strategy and the team that has to ship the system.

In the traditional consulting model, that handoff is a deliverable. A roadmap, a target operating model, a vendor selection matrix. The strategy firm hands the artifact to the systems integrator or the internal engineering team, and the strategy firm leaves. By the time the engineering work begins, half the assumptions in the artifact are stale, and the only people with judgment on whether to honor or revise them are no longer in the room.

The handoff is also where most cross-industry observers locate the gap. In a relay race, the race is won and lost at the baton pass, not on the straightaway. In enterprise AI, the same shape holds: the slowest part of the race is the moment when context, judgment, or accountability is supposed to move across a boundary, and nobody owns whether it actually does.

Applied Intelligence as I’m defining it has three operating-model components, each aimed at a different seam:

• Strategy and engineering under one roof. Recommendations get built by the people who will live with them, and engineering judgment shapes the strategy before it’s published, not after.
• Dual engagement with leadership and operators. Executive sponsorship and operator-level adoption run as concurrent workstreams, not sequenced ones. Adoption isn’t the last mile; it’s the parallel mile.
• Ninety-day measurable wins, not eighteen-month transformations. Each cycle picks one consequential handoff, fixes it as a single accountable system, measures what moves at the customer level, and then picks the next one.

None of this is novel as an aspiration. What’s novel is running it as one operating model rather than three procurement decisions.

What the discipline looks like in production

The clearest evidence that the components themselves work has nothing to do with the size of the model. It comes down to whether someone owned the seam.

In financial services, large institutions have spent a decade re-asking new customers for information the institution already collected during the application, because the KYC workflow and the onboarding workflow were optimized by different teams against different metrics. The institutions that quietly fixed that handoff, usually in twelve to fourteen weeks of focused work rather than a multi-year transformation program, saw application abandonment drop and time-to-funded-account compress materially. The AI didn’t change. The seam did.

In hospitality, the same lesson surfaced almost a decade earlier. Guests stopped tolerating a check-in that didn’t know they’d booked the room, and the operators who closed that loop did it by making one team accountable for the experience across the booking system, the property management system, and the loyalty platform. The operating model was the intervention. The platform decisions were downstream.

Healthcare is the most public current case because the consequences are immediate. At NYU Langone, a deployed BERT model triaging patient portal messages, studied across 396,466 messages and published in JAMIA Open in August 2024, cut clinician read time on high-acuity messages by 44 to 67 percent. That’s an at-scale, peer-reviewed result inside a single integrated workflow. The component is real. The institutions that will compound that result are the ones building the orchestration capacity around it, not waiting for the next vendor demo.

Practice, not project

The healthiest enterprise AI programs I see are run more like a clinical practice than a transformation program. Ninety-day learning cycles, one handoff at a time, with a single accountable owner for the experience across the channels that touch it. The data is instrumented to follow the customer rather than the channel, and the metrics live at the handoff, not at the dashboard for any individual system.

This is what Deloitte’s State of Generative AI in the Enterprise 2026 calls the “untapped edge” — the operating gap between organizations that have expanded AI access roughly 50 percent year over year and the roughly one quarter that have moved more than 40 percent of their experiments into production. The expansion of access has become almost free. The discipline of production hasn’t.

That discipline is the work. It’s also, in the end, where the next decade of enterprise AI value will be sorted.

The capacity you build now is the value you compound later

Yue is right that AI executes broken systems faster. The corollary I’d offer is that AI also compounds disciplined ones faster, and the choice between those two outcomes is an operating-model choice, not a tooling choice. The capability is already on the shelf. The enterprises building Applied Intelligence as a practice now — the strategy-engineering integration, the dual engagement, the small honest cycles — are the ones whose AI investments will earn their next budget cycle, and whose customers will stop feeling every seam.