Honest AI

What AI can't do (yet): an honest list

May 12, 2026 · 6 min read

Here's my rule of thumb: anyone selling you AI who can't tell you where it fails is selling you something else. I build with these tools every day, so my list of limits comes from bruises, not blog posts. This is that list.

1. It can't know your business

AI has read most of the internet. It has read nothing about how your shop actually runs — which customers get flexibility, why that supplier gets one more chance, what that cryptic column in your spreadsheet really means. Every useful thing I've built worked because a human supplied that knowledge explicitly. The AI cooks; you fish. No model version will change whose knowledge that is.

2. It can't be trusted to state facts unsupervised

AI models can assert false things with total confidence — the industry calls it hallucination, and you should just call it "being wrong, confidently." It matters most exactly where it's most tempting to let AI run alone: facts, figures, citations, legal and medical specifics. The practical rule: AI drafts, a human verifies anything that has to be true. This is also why I push so much work toward deterministic scripts, which can't hallucinate — same answer every time, by construction.

3. It can't give you the same answer twice

Ask a model the same question twice and you may get two different answers. Fine for brainstorming; disqualifying for anything a business must do identically every time — invoice math, compliance steps, data transforms. Rule-based work belongs in rule-based code. AI can help write that code, and that's precisely the division of labor that works.

4. It can't own consequences

AI will happily draft the email that damages a twenty-year customer relationship. It bears none of the consequences; you bear all of them. That's not a temporary technical gap — accountability just isn't something software can hold. Anywhere a decision is expensive to reverse, AI belongs in the "prepare" seat, never the "decide" seat.

5. It can't tell you when it's out of its depth

A good employee says "I'm not sure about this one." AI mostly doesn't — it produces its best guess for the impossible request in the same confident tone as the easy one. The countermeasure is structural, not hopeful: design workflows so verification is built in — checkable outputs, small steps, a human at the judgment points. (Being able to specify checkable outputs is why asking precisely is the skill I teach first.)

6. It can't replace the doing that creates judgment

The subtlest one. If AI drafts everything, where does the judgment to evaluate drafts come from? The answer, for now: from humans who learned by doing. In a small business this cashes out simply — use AI to remove the repetitive work, and be deliberate about keeping the work that builds the judgment your business runs on.

What this list is for

None of this is a reason to sit out. Notice the shape of the failures: they cluster around facts, consistency, accountability, and judgment. Everything outside that cluster — the drafting, the extracting, the reformatting, the code-writing, the gigabyte-of-text reading — is enormous, real, and available to you today. Knowing the limits isn't pessimism; it's how you deploy the strengths with confidence.

Hype says AI can do everything. Cynicism says it's all smoke. Both save you the trouble of thinking. The money is in the specifics.