What You’ll Find This Week

95% of enterprise AI pilots are failing to scale. That number has been out there for almost a year now, and the response from most of the people covering it has been either "the technology isn't ready" or "your organization isn't ready."

Neither answer tells you what to actually do differently. Neither answer tells you where the money is going wrong. The data has an answer. It's not the one most organizations want to hear.

Here’s what you’ll find:

Don’t Miss Our Latest Podcast

This Week’s Article

95% of generative AI pilots are failing to scale. MIT put that number out last summer, and two camps immediately made it about themselves.

Investors read it as a technology story: AI isn't ready yet.
Consulting firms read it as a market opportunity: your organization isn't ready yet.

What neither bothered to ask was this: if 95% of pilots are failing, and the technology demonstrably works, where is the failure?

BCG's research on enterprise AI transformation identifies a consistent allocation pattern among organizations that achieve tangible results for their AI initiatives:

They call it the 10/20/70 rule, and it's framed around where your effort should land. But effort is hard to measure directly. Budget is what tells you where organizational priorities actually show up. And the spend data on enterprise AI is consistent. Across every major survey of enterprise AI investment, the technology layer (software licenses, infrastructure, integration) absorbs the overwhelming majority of every AI dollar. The organizational layer (training, change management, workflow redesign) gets what's left. What's left is not 70%.

The AI budget pyramid isn't slightly off. It's completely inverted.

MIT's NANDA report published another statistic worth pausing on: 61% of enterprise AI projects were approved on projected ROI that was never formally measured after deployment. Not "failed to achieve ROI."

Never even measured.

An organization that doesn't measure whether its AI investment worked has already revealed the unstated question: what was the purpose of this AI initiative? The answer is simple: to be able to claim AI readiness and AI implementation in front of a world that increasingly expects it. The initiative was the deliverable. Launching it was the goal. The business case got the budget. The proof never did.

The organizational work that determines whether AI actually gets used was never in the plan: no measurement infrastructure, no accountability structures, no feedback loops to tell you whether the deployment changed how work gets done. The ROI slide was in the deck for the budget approval. Nobody bothered to build the slide for the post-deployment review.

A deployed tool is not an adopted tool. SAP WalkMe surveyed 3,750 executives and workers and found that more than half had abandoned enterprise AI tools, and 37% skip AI entirely on any given workday (SAP WalkMe, April 2026). The executives surveyed estimated their organizations were running 35 AI applications. The actual number was 661.

Dan Adika, CEO of WalkMe, described what happens when expensive tools land without workflow integration: they get used for whatever requires the least behavior change. He called it "the most expensive spell checker ever built." The tools shipped. The workflows didn't change.

The work that would have connected them was never in the budget.

When an organization buys an AI tool, the purchase is the event. There's a contract, a go-live date, a deployment announcement. The tool is live. The job feels done.

Deloitte surveyed organizations across industries in 2025 and found that 37% are deploying AI with little or no change to existing processes. The tool lands. The workflow stays the same. The people expected to use it never had a reason to change how they work.

The organizations that avoid this pattern invest differently. Gartner's April 2026 analysis found that organizations achieving meaningful AI outcomes invest up to four times more in data quality, governance, talent, and change management than those that don't. Four times more. On the organizational layer. The layer that doesn't show up in the deployment announcement.

The misallocation isn't limited to the technology/people split. 50% of GenAI budgets flow to sales and marketing (Menlo Ventures, 2025). The visible end. The end that makes for a good earnings narrative. The clearest, most consistently documented ROI from AI is in back-office automation: finance, operations, BPO cost reduction. The money and the returns are pointing in opposite directions.

In BCG's 10/20/70 framework, people and process change account for 70% of the effort allocation: the work that determines whether the technology layer ever gets used. Specific organizational infrastructure. Not a training budget. Not a change management retainer you sign and forget about. The actual redesign of how work gets done around a new capability.

Stanford's Digital Economy Lab studied 51 actual AI deployments across 41 organizations, spanning 7 countries and more than a million employees. One of them was a professional services company that tried to automate its recruiting function with AI. It failed.

The same technology, applied to the same function at the same company the following year, delivered an 83% efficiency gain. What changed? The CEO took ownership instead of delegating to the CTO. They mapped and fixed the entire recruiting workflow before touching the technology again. Same tool. Different approach. Different outcome. Stanford's summary of the first attempt:

That distinction is what the organizational investment actually buys: fix the process first, then apply the tool. Stanford's research identified four factors that consistently separated the organizations that scaled from those that didn't.

Workflow mapping before technology selection.
You understand the process before you choose the tool. The organizations that failed skipped this and assumed the tool would reveal the workflow. It doesn't.

Governance embedded from day one.
The question of who owns AI outcomes, who measures them, and who is accountable for the gap between deployment and actual use gets answered before the vendor contract is signed. Organizations that added governance after the fact found there was nothing meaningful left to govern.

Observability before production.
The measurement infrastructure (what gets tracked, what constitutes success, how performance gets evaluated) gets built before the tool goes live. Not retrofitted later, when the data to build it no longer exists.

Leadership continuity through early setbacks.
The organizations that scaled kept the same executive sponsor through the first cycle of failures. The organizations that didn't made leadership changes when early results disappointed. The initiative lost the organizational permission to fail and learn.

Every one of those is an effort decision. None are procurement decisions.

The organizations finding real ROI on AI have stopped treating the tool purchase as the main event.

The 10/20/70 rule asks two questions:

Pull your actual AI spend. Not the projected allocation, the actual one. Separate the technology layer (tools, licenses, infrastructure, integration) from the organizational layer (training, workflow redesign, change management, governance, internal champions, measurement infrastructure). In most organizations, the ratio will not look like 10/20/70. The gap between what you find and what BCG says is your organization's honest statement about AI. The budget doesn't lie. The question is whether anyone's been reading it.

Share This