•
3 Silent Killers of Enterprise AI (And Why Your Pilot Will Probably Fail)
Here's a pattern I keep seeing. Smart teams at smart companies spin up an AI project. They pick a vendor. The demo looks great. Everyone's excited.
And then nothing happens.
The pilot sits there. Months pass. Eventually someone quietly kills it. Nobody talks about why.
I've watched this play out enough times now that I can almost predict the failure points before they happen. There are three of them. And weirdly, none of them are about the AI being dumb.
Pain #1: The Access Control Nightmare Nobody Talks About
Let's say you want to give your sales team an AI that can answer questions about pipeline, accounts, renewals. Simple enough, right?
Wrong.
The moment you try to deploy this to real users, you hit a wall: who gets access to what?
Your IT team can enforce access policies. They configure the rules, manage the permissions lifecycle, handle SSO. But they have no idea what the data actually means. They don't know which columns contain material non-public information. They don't know that one particular field shouldn't be visible to commercial sellers but is fine for enterprise reps.
Meanwhile, your data platform team understands all of that. They know exactly what's sensitive, what the compliance implications are. But they can't enforce anything. They don't own the permission systems.
Two teams who need to collaborate on something neither of them fully owns. And caught in the middle is some poor business analyst who just wants to ask the AI a question and can't understand why it takes three months of "security review" before anyone can use anything.
Here's the real issue: for twenty years, we've controlled data access at the application level. You build an app, you define who can see what within that app, done. But AI breaks that model completely. Suddenly you're not controlling access to screens and buttons - you're controlling access to data itself, across every possible question someone might ask.
That requires a different layer. Something that sits between the AI and the data. Something that knows about users AND data sensitivity AND can enforce policies at query time. Most organizations don't have that layer. They're trying to retrofit enterprise AI onto a permissions architecture that was never designed for it.
Pain #2: The Multiplayer Problem (Or: Why Scraping Your Confluence Won't Save You)
Okay so let's say you've solved the access control thing. Now you face an even trickier problem: context.
Everyone agrees that AI needs context to be accurate. The question is where that context comes from. And the answer that most vendors give you - "we'll scrape your docs, we'll index your Confluence, we'll RAG over everything" - fundamentally misunderstands the problem. (I wrote about why RAG approaches fail.)
Here's what I've learned: no single person in any organization knows everything. The implications are profound.
When you're deploying AI for business operations, you need context about how the business actually works. Not how it's documented - how it actually works. Those are different things. The documentation says inbound and outbound leads are handled the same way. But actually, in APAC, for historical reasons nobody quite remembers, the regional manager decided to handle them differently. It worked for them.
Until the AI comes along and gives wrong answers for APAC because nobody told it about the exception.
Who was supposed to tell it? The VP of Sales didn't know about the decision - it was made two levels down. The regional manager who made it left the company last year. The current team knows the process but doesn't realize it's different from everywhere else.
This is what I mean by multiplayer. The context isn't sitting in one place waiting to be scraped. It's distributed across dozens of people, many of whom don't even realize they're holding pieces of it. It emerges from decisions made in hallways and Slack threads that were never documented.
You cannot interview your way to this context. You cannot scrape your way to it. You definitely cannot have some vendor come in for a two-week discovery phase and expect them to capture it.
The only way to build shared context is incrementally, through actual use. Someone asks a question. The AI gets it wrong. That person - who has the specific context to know it's wrong - corrects it. That correction gets captured. Repeat this a thousand times across a hundred people and you start to build something that actually represents how your organization works.
Most enterprise AI projects fail because they treat context as a setup problem instead of an ongoing, multiplayer, organizational problem. Traditional semantic layers weren't built for this.
Pain #3: The Speed Trap
So let's say you've somehow solved both of those problems. Now you need to actually deploy this thing.
The enterprise AI buying cycle has trained everyone to expect a six-month implementation. Three months of "discovery." Two months of "integration." One month of "testing." Consultants and system integrators everywhere.
By the time you go live, the business requirements have changed. The executive sponsor has moved to a different role. The team that was excited about this has moved on. The pilot dies of old age before it ever becomes production.
But here's the thing: most of that complexity isn't inherent to AI. It's inherent to the assumption that you need to move data around, build ETL pipelines, stand up new infrastructure. If you can query data where it lives - without copying it, without transforming it, without six months of data engineering - the deployment timeline collapses.
POC equals production. That's not a marketing slogan. It's a statement about architecture. If your AI system is designed right, the thing you build in week three of evaluation should be the same thing that goes live. Not a demo. The actual system.
The Pattern
These three problems - access control, shared context, time to production - share something: they're all cases where the traditional enterprise software playbook doesn't work, and organizations keep trying to apply it anyway.
The 95% of enterprise AI projects that fail (per MIT's research) mostly fail here. Not because the AI isn't smart enough. Because the infrastructure assumptions were wrong from the start.
Want to understand how shared context actually works in practice? Read more here.
