Most "enterprise AI" products are just consumer AI with an SSO login and a higher price.
The actual enterprise requirements — retrieval-layer access control, data residency, audit completeness, compliance documentation — are missing or bolted on. If the vendor cannot explain how access control works at the retrieval layer, they built it for consumers and are charging enterprise prices.
The ROI numbers in vendor decks are fictional.
Not lying exactly. But they are based on ideal adoption, perfect data, and users who already know how to prompt well. Ask for numbers from customers who are 18 months in with normal adoption rates and real organizational data quality. That number will be 40-60% of what the deck says.
AI tools that feel impressive in a 30-minute demo are not the ones that prove reliable at month twelve.
Reliability at scale requires boring infrastructure work: document lifecycle management, access control, retrieval quality monitoring, prompt versioning. None of this is impressive in a demo. All of it determines whether the tool is trustworthy a year later.
"We have a zero training data policy" is not a data handling policy.
It is one specific commitment about one specific thing. It says nothing about inference logging, prompt caching, subprocessor chains, retention schedules, or what happens to your data during a security incident. Read the DPA. If you are not sure what to look for in a DPA, that is a different problem.
The access control failures in AI deployments are usually not the vendor's fault.
They are the customer's fault for not designing the deployment around access control from day one. Connecting AI to "all company documents" without thinking about which documents should be inaccessible to which users is a design choice that creates a predictable failure. The vendor did not make that choice. You did.
Free trials are almost always misleading.
The demo data is clean. The users are motivated. The vendor is on-site. None of that is what production looks like. The useful evaluation happens when you test with your actual messy data and your actual least-motivated users and no vendor support in the room.
Most AI tool sprawl is a leadership failure, not a technology failure.
If you have twelve AI tools and nobody can tell you which three are actually worth the spend, the problem is not the tools. Nobody made explicit decisions about which ones stay and which ones go because that kind of decision is uncomfortable to make. Uncomfortable decisions that don't get made turn into expensive drift.
None of this is popular to say when everyone is very excited about AI. All of it will seem obvious in retrospect.
Top comments (0)