The 95% problem
Quick one this week, because the point is simple.
MIT just put a number on what most of us have been feeling. Their "GenAI Divide" report looked at 300 enterprise AI deployments plus interviews and surveys with hundreds of executives.
The finding: 95% of generative AI pilots deliver no measurable P&L impact — and that's after $30–40 billion in enterprise spending.
McKinsey's numbers are in line with it: roughly 8 in 10 companies now use generative AI, and roughly 8 in 10 report no significant bottom-line impact from it.
Ninety-five percent. Not "underperforming." Zero measurable impact.
The popular explanations are model quality and regulation. I don't buy either, and neither do the MIT researchers — they point at integration. I'll go one step further and name the thing I keep seeing inside real companies:
FACT: We're over-engineering
AI agents made building cheap, so coders and product people are building too much. Vibe-coded dashboards on top of dashboards. Agents that generate reports nobody asked for. Pipelines that turn one pile of information into three new piles of information, each with its own interface to check. From the front end all the way through the backend, teams are using AI to manufacture more surface area — more things to read, more places to look, more stuff to maintain.
That's backwards. AI's job is to simplify the intelligence layer, not expand it.
Trust the model to do the thing it's actually good at: distilling. The right output of an AI system is almost never another feed. It's one curated answer. A three-sentence brief. A single visual that replaces a meeting. If your AI initiative adds to what your people have to read and check every day, it's a cost center — no matter how smart the model underneath is.
The 5% of projects that work share a trait: they remove work. Measure your AI by what it deletes from someone's day, not by what it generates.
What I'm building: exactly this idea, as a product
TextMyAgent is an AI assistant that lives where you already are: your text messages. No app. No dashboard. No new interface to check.
It watches your email and calendar, texts you when something actually matters, and answers the questions you'd otherwise dig for — "what needs me today?", "did she ever respond?", "what's on deck tomorrow?" You ask in plain English. You get one distilled answer back.
Text 737-377-1927 and try it. Your first two weeks are free.
That's the whole pitch. Send it one text and see what it does with your Monday.
Does this resonate?
If you're seeing the over-engineering problem inside your company — or you think I'm wrong — hit reply and tell me. Email me back, let's talk. I read every response, and the best conversations in this newsletter's history started as replies.
— Keith