Business Strategy & ROIFailure Modes

What AI actually cannot do

In brief

Four of the five limits this article originally listed are now false — Claude searches the web, remembers across sessions, reads your files through connectors, and runs code to compute. Here's the current list of what Claude genuinely cannot do, and why the remaining limits are structural rather than technical.

5 min read·Hallucination

Contents

♡Sign in to save

Most published lists of "what AI can't do" are quietly out of date, and this article used to be one of them.

When we first wrote it in April 2026, it said Claude could not search the web, could not remember anything between conversations, could not read your files unless you pasted them in, and could not learn from your corrections. All four of those are now false. Web search shipped. Memory shipped, and it carries across chat and Cowork. Connectors shipped, and there are now over 950 of them.

That is worth sitting with before you read the new list, because it is the most useful thing in this article: the limits move roughly every quarter, and the confident list you read six months ago is the thing most likely to be wrong. People who built workflows around "Claude can't look things up" spent a year pasting search results into a chat window for no reason.

So here is the current list, with the understanding that it too has a shelf life. What has changed is that the remaining limits are no longer capability gaps. They are structural.

The limits that fell

What we used to say	What is actually true now
Cannot look anything up in real time	Web search is available in Claude and on the API, with citations
Cannot remember previous conversations	Memory persists across chats and into Cowork tasks, as editable entries you control
Cannot read your files, emails, or documents	Connectors reach Drive, Gmail, Slack, Airtable, GitHub and ~950 more; custom remote MCP servers work on every plan
Cannot reliably produce accurate numbers	Claude runs code to compute, rather than predicting digits
Cannot learn from your corrections	Corrections you ask it to remember persist into later sessions

If your mental model of Claude is more than about two quarters old, check it against the model overview before you design a workflow around a limit.

What Claude genuinely cannot do

It cannot tell you how confident it should be

This is the one that has not moved and shows no sign of moving. Claude produces fluent, well-structured, confident prose whether or not the content is correct. The register does not change. A wrong answer and a right answer look identical.

Search helps less here than people expect. Retrieval means the claim now has a source attached — but Claude can still misread the source, cite a page that says something adjacent, or lean on a result that is itself wrong. A citation is evidence that a page exists. It is not evidence that the page supports the claim.

What to do: Calibrate on stakes, not on how confident the output sounds. For anything you will act on or ship, click through to the source. The failure mode is not "Claude made something up" — it is "Claude gave you something plausible and you had no signal to check it."

It cannot know what nobody wrote down

Connectors give Claude your documents. They do not give it the reasoning that never made it into the documents — why the deal was structured that way, which stakeholder will block this, what happened the last time someone tried it, which parts of the roadmap everyone privately knows are dead.

In most organisations, the load-bearing context is in people's heads and in conversations that were never recorded. Claude reads what exists. Nothing gives it what doesn't.

What to do: When output is technically fine but somehow wrong for your situation, the missing ingredient is almost always undocumented context. Say the quiet part explicitly in the prompt.

It cannot hold accountability

Claude can draft the analysis, run the numbers, and argue both sides. It cannot be the one who is responsible when the decision is wrong. That is not a capability gap that a better model closes — it is a category difference. Responsibility requires someone with something at stake.

This matters practically, not philosophically. The moment a workflow has no named human who owns the output, the quality of that output stops being checked by anyone, and it degrades without anyone noticing.

What to do: Every automated workflow needs a named owner and a review step with teeth. See when agents break for what this looks like in practice.

It cannot reliably report on itself

Ask Claude what it just did, which tools it called, why it made a choice, or what it is capable of, and you get a plausible reconstruction rather than a log. It does not have privileged access to its own processing, and its knowledge of its own product surface is bounded by a training cutoff that is months behind the release notes.

This is the specific reason "I asked Claude and it said it can't do that" is unreliable evidence about what Claude can do.

What to do: Check capability questions against the docs and the release notes, never against the model's self-description. For agent behaviour, read the actual trace, not the summary.

It cannot make the calls that require taste or stakes

What your product should be. Whether to fire someone. Which customer to disappoint. When to abandon a strategy you have publicly committed to. Claude will help you think through any of these more completely than you would alone, and it will not tell you what to believe, because conviction is not something it has.

What to do: Use it to pressure-test the decision — "what am I not considering, what would change my answer" — then make the decision yourself.

The practical stance

Treat Claude as a capable colleague whose work you read before it goes out, whose sources you spot-check, and whose confidence you ignore as a signal.

And re-check the limits every few months. The most expensive mistake with AI right now is not over-trusting it. It is building a careful workaround for a constraint that quietly stopped existing two releases ago.