Most AI Tools Are Not Actually Smart

Every AI product on the market calls itself smart right now. Smart assistant. Smart search. Smart workflow. Smart agent. Say a word enough times in enough pitch decks and it stops meaning anything.

Here's what "smart" usually means in practice: the product found the thing you asked about and handed it back to you in a sentence. That's useful. Whether it's smart is a different question.

A tool that retrieves is not a tool that thinks. A tool that answers your question is not a tool that notices what you didn't ask about.

So this sprint we argued about a deceptively simple question: what would actually have to be true for an AI tool to earn the word "smart"?

The first answer is obviously data. Your tool should have access to the actual, trustworthy, complete data source. But Wikipedia with its search is not a smart assistant for sure. So what else? We think it comes down to the two words themselves.

Which means it couldn't be universal. It has to be user-specific.

In other words, we need to achieve something close to the quality of targeted advertising algorithms on Instagram. Which, if you read our last issue, you already know comes with its own complications. (Also, it reminds me of an episode of Black Mirror.)

By that definition, almost nothing on the market today fully qualifies as a smart assistant. Including, honestly, ours.

Leebry today is mostly retrieval: citation-backed answers, permission-aware access, humans in the loop on actions that matter. The noticing layer, flagging stale content, surfacing contradictions between documents, is maturing rapidly. The personal layer, the one that learns how you specifically work over weeks, hasn't started.

Even Gartner is Calling It Out Now

Gartner published their first-ever Hype Cycle for Agentic AI, and the most useful thing in it is not the curve, it's in the language.

For the first time, they have named "agent-washing" as an explicit market problem, defined as vendors rebranding legacy automation tools as AI agent platforms without substantial agentic capabilities.

The numbers underneath the trend back it up.

A separate enterprise AI report from Sinequa, surveying 740 senior executives, found that 51.3% claim to have AI agents in live production, but only 10% have actually deployed true multi-agent systems. The other 70.7% are operating "assistive AI" or below: sophisticated knowledge-retrieval tools that cannot independently take actions.

The reason this matters for IT leaders: if the language is broken at the analyst level, the procurement and evaluation language inside your own org is probably broken too. The question worth asking on any vendor call right now is what specifically the product does that can make it earn this language.

If you want the longer version of this argument, we hosted a panel at #BostonTechWeek on the same idea last month.

It is the cleanest framing we have heard for what "smart" should actually mean.

That's Issue #4.

If you have a working definition of what "smart" should mean in enterprise AI, reply and tell us. We are still arguing about it on our end.

See you in two weeks.

— The Leebry Team