The 95% Aren’t Failing. They’re Clearing the Runway.

What the failure rate is really measuring

The pilots failing in 2025 were built to answer a 2023 question: can a language model do a discrete task faster than a person? Write the email. Summarize the document. Draft the first pass of the memo. And the answer, demonstrably, is yes. The shadow AI economy proves it. Your employees are already paying out of pocket for ChatGPT and using it to get real work done, often without telling anyone. The technology works. They know what good feels like.

The problem is that no enterprise runs on discrete tasks. It runs on workflows. A customer support resolution is not “write a reply.” It is intake, triage, lookup across three systems, a judgment call, a follow-up, a status change, and a closed loop. A first-generation tool that nails one step of that chain and ignores the other six does not transform the workflow. It adds a thirty-second savings to a process that has forty other points of friction, and then asks the organization to absorb the cost of change management, integration, and governance to capture that thirty seconds.

So of course it doesn’t move P&L. We instrumented a sports car and bolted it to a horse cart.

The 95% is the market rejecting tool-based AI. It is not rejecting AI. Those are different sentences, and the companies that confuse them are going to spend 2026 cutting the wrong budget.

The number that actually predicts success

Here is the finding from the report I keep coming back to, because it reorganizes everything.

The companies that are winning, the roughly five percent generating compounding value, do not have better models. They have a different spending ratio. The losers put as much as ninety percent of their AI budget into algorithms.

Sit with that for a second. The single best predictor of return is not which frontier model you license. It is whether you treated the project as a business transformation or as a software purchase. McKinsey found the same pattern from a different angle: the firms reporting real financial returns were twice as likely to have redesigned the workflow before they ever selected the modeling technique. They started with the work, not the tool.

This is uncomfortable for a lot of people, because “buy the best model” is a procurement decision a single VP can make in an afternoon. “Redesign the workflow” is organizational surgery. It requires the CEO, not the CIO. And only twenty-eight percent of companies have CEO-level AI governance, even though it is the strongest single predictor of returns in the entire dataset. We delegated the most consequential technology shift of the decade to the IT department and then acted surprised when it produced IT-department-sized results.

The benchmark trap

There is a related mistake hiding in how we evaluate the models in the first place.

The traditional benchmarks are saturating. On the Chatbot Arena, the Elo gap between the top model and the tenth-ranked model shrank from nearly twelve percent to five and a half percent in about two years. Leading Chinese open models now sit at near-parity with the frontier on MMLU and HumanEval. When everyone scores in the high eighties, the score stops telling you anything useful about who will win.

Raw intelligence is commoditizing. That is not a tragedy. It is a signal about where the value moved. It moved off the model and into the system around the model: the data pipelines, the orchestration, the feedback loops, the proprietary workflow knowledge that no one else can replicate. The company with the highest GPQA score does not win. The company that builds the tightest learning loop around a perfectly adequate model wins, and keeps winning, because the loop compounds and the score does not.

If your AI strategy is a spreadsheet comparing benchmark numbers, you are optimizing the one variable that has already stopped mattering.

From intelligence to agency

The reason I am not bearish on any of this is that the next paradigm is already arriving, and it is built for the exact problem the first generation failed to solve.

A generative model writes the email. An agent is told “set up the meeting with the marketing team next week” and then checks calendars, proposes times, negotiates, and sends the invite. The unit of work is no longer the task. It is the goal. That is the whole shift in one sentence, and it maps directly onto the workflow problem that broke the pilots.

The ecosystem is reorganizing around this. I think of it as a three-layer stack. The Foundation layer is compute and models, and it is commoditizing fast. The Interface layer is how humans hand goals to agents. The interesting layer, the one where the next decade of enterprise value and lock-in gets decided, is the Orchestration layer in the middle: the systems that handle planning, memory, tool use, and multi-agent coordination. That is where your proprietary process logic gets encoded, and process logic is the one thing a competitor cannot buy off a shelf.

Agents also break the business model. Per-seat SaaS made sense when you were selling access to a tool a human operated. It makes no sense when the software does the work itself. Pricing is already migrating from access (subscriptions) to usage (consumption) to outcomes (a fee per resolved ticket, a share of the revenue an agent generates). The vendor stops being a tool provider and becomes a performance partner. That is a far more demanding relationship, and a far more valuable one.

What I would actually do

If I were running a company through this, I would do five things, and none of them start with a model. I would fix the data plumbing first, because forty-three percent of organizations name data quality as their top obstacle and an agent acting on bad data does not hallucinate a sentence, it hallucinates an action. I would map and redesign the workflow before selecting a single tool, and I would let the people who actually do the work surface the problems worth solving. Top-down strategy, bottom-up sourcing.

I would start with the unsexy wins. Not the customer-facing moonshot. The high-volume, low-risk back office: procurement, compliance reporting, internal IT. Boring, measurable, safe places to build confidence and bank the ROI that funds the ambitious stuff.

I would build the governance for autonomy now, while the stakes are low. An agent with API access is a new attack surface and a new failure mode. You want behavioral monitoring, audit trails over the full decision chain, and a clear human-on-the-loop posture before an agent is anywhere near a production system, not after.

And I would treat the whole thing as building a learning organization, not buying a capability. When a user overrides the AI, that override is not noise to discard. It is the most valuable training signal you have. Capture it.

This is a leadership problem before it is a technology problem

Read that list again and notice what it is not. It is not a list of tools. It is the job description for a kind of leader most companies do not have yet.

Every failure mode in this piece traces back to the same root cause: a prototype-first reflex running an organization that needed an architecture-first approach. Prototype-first asks what can we demo by Friday. Architecture-first asks what has to be true for an agent to act inside this enterprise without breaking trust. The first reflex produced the hundred-pilot portfolios and the 95% number. The second is the only thing that survives contact with production.

I have argued for a while that this is why a new executive archetype is becoming unavoidable. Not the CIO as infrastructure manager, not the CTO as product technologist, not the CDO as data custodian, not even the CAIO as AI evangelist. The AI-native enterprise needs a SuperCIO: the executive architect who connects the whole stack into a system the business can actually trust to act on its behalf. (I made the full case for this in From Generative AI Reinvention to the SuperCIO Enterprise.)

That stack is six layers, not the three I sketched above for the vendor market.

The first three are where prototype-first stops. The last three are where enterprise value actually lives, and they cannot be bolted on after the fact. They have to be designed in.

So here is the call to action, and it is not “buy a platform.” It is this: name the person who owns all six layers. If no one in your C-suite can connect infrastructure to information to intelligence to decision to governance to reinvention as a single operating system, then your AI strategy does not have an architect. It has a collection of pilots waiting to join the 95%. The companies that get to the upswing first will be the ones that treated this as a leadership appointment, not a procurement cycle.

The shape of the curve

The honest way to read 2025 is as the bottom of a J-curve. Heavy investment, steep learning, painful restructuring, and modest visible returns. That dip is not the technology failing. It is the cost of rewiring how organizations work, paid up front, before the payoff.

The upswing will not come from the next model being two points smarter on some exam. It will come from agentic systems and embedded AI maturing to the point where entire industries get re-architected rather than lightly augmented. That is where the non-linear gains live.

The gap between the companies that understand this and the ones still comparing benchmark scores is already widening. BCG puts roughly five percent of firms in the “future-built” category and sixty percent stuck in experimentation. That chasm is the real story of the next decade, not the model leaderboard.

The 95% failure rate is the runway clearing. The question worth asking is not whether AI works. It plainly does. The question is whether your organization is built to use it, and whether anyone owns the architecture that would make that possible. Most companies cannot answer the second part yet. The ones that get to the upswing first are not going to be the ones with the best model. They are going to be the ones who put an architect in charge of all six layers and then changed everything except the part everyone else obsessed over.