MIT Says Your AI Stinks. Here’s Why. And how to fix.

MIT’s NANDA initiative just published what every executive already knows but won’t say publicly: 95% of enterprise AI pilots are failing.

I’m not surprised. I predicted this exact failure rate in 2022-2023.

Note: This article is a part of The Final Industrial Revolution series.

 
Here is the original article from MIT:  

Figure 1. Image from MIT’s MIT report: 95% of generative AI pilots at companies are failing published article.

 

MIT’s diagnosis is correct: 95% failure rate.
Their prescription? Buy vendor tools instead of building internally.

But here’s what they missed: You can’t integrate intelligence into chaos.

Your enterprise software wasn’t designed. It evolved.
Like a coral reef — layers upon layers of legacy systems, vendor solutions,
and 'temporary' workarounds that became permanent.

Vendor AI tools fail for the same reason internal builds do:
They’re trying to add intelligence to architectural anarchy.

In The AI Evolution Playbook: Why 80% Will Fail, July 2023, I wrote:
"80% of AI projects would fail because most enterprise software wasn’t designed — it evolved."

I was too optimistic. It’s 95%.

MIT says buy vendor tools.
I say: Fix your foundation first, or join the 95%, vendors and all.

Now, let’s take a look at the actual failure patterns to understand "Why".

Unlike previous technology transformations, AI cannot be approached as a "fast follower" strategy.^[1]
 
Cloud and digital transformations allowed companies to wait for packaged vendor solutions and catch up through commodity adoption.
 
AI transformation is fundamentally different because it directly affects business identity and competitive positioning.

The Three Patterns:

Every one of those 95% failures followed one of three patterns. I documented these in 2023. MIT just proved me right.

Pattern 1: The Wrapper Gambit:

What they promised: "Let’s build a chatbot that sits on top of everything!"
What happened: $50K/month to run. Nobody uses it. Why? Because wrapping chaos in a conversation interface doesn’t make it intelligent.

Pattern 2: The Big Bang Rewrite:

What they promised: "We’ll AI-enable your entire platform!"
What happened: $3M spent. Nothing shipped. Team laid off. You can’t revolutionize what you don’t understand.

Pattern 3: The Vendor Solution:

What they promised: "Our platform has AI built in now!"
What happened: 40% higher licensing costs. Works like expensive search. Even MIT fell for this one — recommending vendor tools with a 33% failure rate as "success."

And now, Pattern 4: The MIT Gambit (The "Just Buy It" Fantasy):

What MIT promises: "Purchase AI tools from vendors - 67% success rate!"
What they don’t tell you: That’s still a 33% failure rate for software that costs millions. And those "successes"? Many are just expensive search engines with chat interfaces.

It is all just for ONE trivial and generic Use Case!

MIT’s advice is like telling someone with a broken foundation to buy better furniture.
Sure, vendor tools work better than DIY disasters. But they still fail when your architecture is chaos.

The real tragedy? MIT identified the problem — a "learning gap" where tools don’t learn from or adapt to workflows  — but then recommended the same tools that can’t bridge that gap.

The math doesn’t lie: If 95% of internal builds fail, and 33% of vendor solutions fail, that means ~37% of ALL enterprise AI initiatives succeed. MIT calls 67% vendor success a win. I call 37% total success an industry crisis.

What we’re missing in the business world right now is understanding.

The Actual Validated Insight:

MIT says the core issue is a "learning gap" — that tools and organizations can’t learn from each other.

They’re half right.

The real problem? Tools can’t learn from chaos.

When your architecture is:

• 47 different systems that don’t talk "human";

• Data scattered across 12 databases;

• Business logic buried in stored procedures from 2003;

• "Temporary" Excel sheets running critical processes.

No AI — vendor or internal — can make sense of that!

The Model Context Protocol isn’t just another protocol.
It’s the architectural foundation that makes AI possible.
Without it, you’re asking AI to be a mind reader for a patient with multiple personality disorder.

Companies with clean architectural boundaries can add AI in weeks (and have).
Companies with architectural chaos will fail — whether they build or buy^[2][3].

MIT got one thing devastatingly right: Generic tools like ChatGPT excel for individuals because of their flexibility, but they stall in enterprise use since they don’t learn from or adapt to workflows.

But then they recommend…​ more generic vendor tools? That’s like diagnosing heart disease and prescribing aspirin.

Don’t Join the 95%!

You have three choices:

1. Keep trying vendor AI tools — 67% success rate sounds good until you realize that’s still 1 in 3 failures for expensive software;

2. Build it yourself — Join the 95% failure club with a homegrown disaster;

3. Fix your foundation first — Then AI actually works (in every way).

The executives who succeed aren’t smarter. They’re not luckier.
They just understood that you can’t build intelligence on chaos.

Before you become another MIT statistic, let’s diagnose your specific situation.
Confidentially. No public exposure of your challenges.

Schedule your confidential assessment
and understand exactly why your AI is struggling
   — and more importantly, how to fix it.

Architecture is trust.

Code is conversation.

Velocity comes from clarity.

— ASE Inc.
Vadim Kuhay

P.S. The MCP Craze:

In the last two weeks I had three MCP startups book a weird meeting with me. I met with two and ghosted a third one — unintentionally — apologies.

What I can already conclude is this:

1. The startups don’t have business foundations in place;

2. They see MCP as some sort of a magic bullet for AI-all;

3. They don’t interview prospects for the root cause - just push MCP.

One of these startups is actively recruiting a lead MCP engineer with 3 years of experience — dudes, MCP is 8-months old!

But here is the harsh truth:

1. MCP is just a convenience tool adding structure to ACLs;

2. Great integration happened before MCP was even imagined;

3. MCP doesn’t solve ANY of your business problems.

Please, don’t fall for hype — understand instead!

The real solution isn’t another tool, protocol, or vendor promise.
It’s understanding your architecture well enough to make any tool work.

That’s what we do at ASE Inc.

Toodles!

---

1. This article The AI Evolution Playbook: Why 80% Will Fail explains HOW to avoid the TRAP described here American AI Integration Trap — your playbook on integrating AI and staying ahead.

2. This is the original article that exposes the root cause for American AI Integration difficulties The AI Evolution Playbook: Why 80% Will Fail.

3. See AI Integration Architecture for Enterprise for more on this story. It’s a part of The Final Industrial Revolution series.