The Other 5%

The Other 5%

MIT says 95% of enterprise AI pilots fail. The stat is everywhere right now, mostly as a sales pitch: the AI labs are now selling billions of dollars of on-site engineers to fix it for you. Less noticed is what the other 5% sound like when they explain themselves. Three companies recently did exactly that, in their own words, at three different altitudes: a whole company, a single department, a single pipeline. The stories are different. The arc is identical. Let people experiment, hit the wall, then institutionalize.

Checkr: the whole company, then the correction

Checkr pushed AI to everyone early. Every employee got a Claude license and a Lovable license, and the company ran hack days for non-engineers. COO Lindsey Scrase is candid about where that led: everyone building things for their own productivity, which was fine, and expensive, and not compounding. The correction is underway: "We're moving toward more centralization." Checkr now defines AI fluency levels and assesses people against them, while pulling individual experiments into shared, centralized agents. Her line for the new phase: "we don't need every single ops agent building an app." The production wins are real, an AI resolution rate in support and operations with goals above 90%, customer satisfaction up roughly 3x since generative AI took over those interactions, support scaled without added headcount. And the next product bet, fraud detection, came from a distinctly human pipeline: it surfaced in customer conversations before it showed up in any dashboard.

Plaid: one department, run like a program

Plaid's CFO Seun Sodipo watched the bottoms-up phase help individuals and stall at the seams: one person gets faster, the multi-hand workflow does not. Her response was to borrow from Plaid's engineering org and run finance AI adoption as a program. A declared four-month journey. Projects scoped at eight weeks, formed from a survey of what the team would build if they could recruit help, ending at a demo day that decides what goes live. Recurring Finance AI Days where the whole team steps out of the day-to-day to build. The builds are unglamorous and real: a mail scanner that routes misdirected tax notices to the right person among 1,100 employees, a third-party Salesforce tool rebuilt in-house in a day. Her bar is the sharpest quality standard we heard all month: "Failure mode in AI is like increasing volume without really increasing impact. For me, it needs to be like net better than what any human on my team could do." And the reason for the top-down layer: "the top down expectation, it raises the floor on the team."

Adobe: one pipeline, engineered like it matters

The window between a vulnerability being announced and exploited in the wild has collapsed to under 24 hours. Adobe runs about seven different web application firewalls, a legacy of acquisitions, and human rule-writers were barely keeping up. So Ammar Alim's product security team built an agentic pipeline: deterministic code polls CVE feeds every 15 minutes and checks whether the vulnerability exists in Adobe's environment; a research agent gathers exploit history; a model generates candidate firewall rules for every WAF; a judge model from a different vendor scores them against a rubric; then a testing ladder that ends with deploying the actual vulnerable app in a private environment and attacking it. Rules go to production in non-blocking shadow mode, get watched for half an hour, then flip on. A virtual patch ships in about ten minutes. The team even generates rules for vulnerabilities that don't affect Adobe, purely so the system learns. Alim's framing to product leadership is the whole business case in one sentence: "when there is a vulnerability, I will mitigate the emergency, so you do not have to go in emergency mode."

What the 5% have in common

None of these are pilots. That is the point. Each one started loose, hit the same wall, individual gains that didn't compound, and responded by building institutions: fluency levels and centralized agents at Checkr, demo-day gates and AI Days at Plaid, judge models and testing ladders at Adobe. The 95% stall because a pilot is designed to be abandonable. The 5% look almost boring up close. Governance, rubrics, gates, staged rollouts. Transformation, it turns out, is not the demo. It is what you build around the demo.

Sources: Lindsey Scrase (Checkr), In Depth, "What startups get wrong about enterprise," Jul 30, 2026. Seun Sodipo (Plaid), Run the Numbers, "Building an AI-Native Finance Team with Plaid CFO Seun Sodipo," Aug 3, 2026. Ammar Alim (Adobe), Cloud Security Podcast, "How Adobe Uses AI Agents for building a WAF Pipeline?," Aug 4, 2026. MIT pilot-failure statistic via Ray Rike and Peter Buchanan, Metrics that Measure Up, "Forward-Deployed Engineers (FDEs) - AI's New ROI Battleground," Aug 4, 2026.