AI NewsAUG 13, 20264 min read
The industry ran the experiment for us
MIT watched more than 300 enterprise AI deployments and 95 percent returned nothing. The reasons are specific, and so is what Bigge does about each one.
The most quoted statistic in enterprise AI right now comes from MIT: 95 percent of corporate generative AI pilots produce no measurable return. The number has spent a year getting repeated in vendor decks and boardrooms, usually without the part that matters, which is why the failures happen. The reasons turn out to be specific, and specific is useful. Here is what the three big reports actually say, and what Bigge does about each one.
The 95 percent problem
The study behind the number is The GenAI Divide from MIT's NANDA project: 52 executive interviews, 153 leaders surveyed, more than 300 public AI deployments analyzed. The 95 percent were not blocked by model quality. They were blocked by fit. The tools did not retain feedback, did not adapt to the actual workflow, and got quietly abandoned by the people they were supposed to help.
The 5 percent that succeeded share findable traits. They aimed AI at back-office operations, while everyone else put more than half their budget into sales and marketing pilots and the returns showed up where the budgets were not. They put tools in the hands of the people doing the work instead of central labs. They picked systems that learn from corrections instead of repeating the same mistake with confidence. And underneath the official failures, the report found a shadow economy: only 40 percent of companies provide official AI access, while workers at over 90 percent of them use personal AI tools anyway.
- Aim at the back office. Our four active builds sit in service, shop, parts, and warranty on purpose, and the flashier customer-facing category has no active build this year. The report found the returns exactly where the budgets were not going.
- Buy when buying wins. Bought tools succeeded about twice as often as internal builds in the study, 67 percent versus 33, and we take that seriously rather than personally. The parts initiative runs on an Egnyte account upgrade instead of a homegrown document store, and our operating model hands finished interfaces to enterprise software teams to maintain. We build where the workflow is ours alone.
- Ship systems that learn. MIT's core diagnosis is tools that forget corrections and repeat mistakes. Every build here ships with evaluation sets and regression checks, and a named owner who confirms it works in practice. A tool nobody grades is a tool nobody fixes.
- Treat workarounds as signal. Bigge is on the sanctioned side of the shadow AI line: office employees have Claude and ChatGPT, and Automation Scout exists to turn the quiet workarounds into intake. The Letter this issue makes the same point. A workaround is information.
Agents are arriving faster than the rules
Deloitte's 2026 State of AI in the Enterprise finds 74 percent of organizations expect to deploy agentic AI within two years, while only 21 percent rate their governance mature enough to manage autonomous agents. Analysts now call governance, rather than model capability, the main barrier to getting value out.
The takeaway for Bigge is visible elsewhere in this issue. The WorkPro connector shipped read-only, and write access gets earned one approval tier at a time with a person signing off. Every system we ship keeps a person in the loop, which means when warranty automation drafts a claim, a human files it. That discipline is what lets an agent stay in production after the novelty wears off.
2.59 trillion dollars of pressure
Gartner forecasts worldwide AI spending at 2.59 trillion dollars in 2026, up 47 percent in a year. The part of that number that reaches you is smaller and more specific: every software renewal and vendor demo Bigge sits through now arrives wearing an AI badge, and often an AI line item.
Hold us to the same standard. It is the exact question the Automation Scout brief asks, and it is the question every build in On the Board had to answer before it started.
The industry spent a year and 30 to 40 billion dollars documenting how AI projects fail: no owner, no baseline, no workflow fit, no learning loop, no review. The rules this department runs on are that failure list turned into gates. We would rather learn from other people's write-offs.
