Let's talk
← Blog

What AI still can't do

Every tool is defined by its edges. Knowing where these ones are is the difference between using AI well and being disappointed by it.

An open notebook showing a handwritten two-column list of ticks and crosses

Most writing about AI is either selling it or dismissing it.

This is neither. It is a list of the things it cannot do, written by someone who uses it every day and expects to keep using it.

Knowing the edges is not a reason to stay away. It is the thing that makes the middle useful.

It does not know what matters

It can process a hundred customer messages and tell you what they say.

It cannot tell you that the short, polite one from your biggest client is the one to worry about.

Knowing what matters requires knowing the stakes, the history, and the person. It has none of those unless you supply them, and even then it is working from a description rather than a relationship.

This is the limit that will not close soon, and it is the one that keeps you necessary.

It does not know it is wrong

Knowing when to check is the habit that covers for this one.

There is no internal signal that says "I am unsure about this."

A confident wrong answer looks and reads exactly like a confident right one. Same fluency, same tone, no hesitation.

Everything else on this list is manageable. This one is why every other item needs care, because you cannot rely on the output to warn you.

It has no memory of you

Each conversation starts from nothing. It does not remember your prices, your policies, the client you discussed yesterday, or the correction you made an hour ago in a different window.

Tools are adding memory features, and they help. But the underlying situation is that context has to be provided rather than accumulated.

This is why people who use it well end up with a document they paste in: their business, their voice, their rules. Not because it is clever, but because the alternative is explaining yourself forever.

It cannot take responsibility

If it produces a wrong figure and you send it, that is your wrong figure.

There is nobody to escalate to and nothing to hold accountable. It cannot be blamed, cannot learn from the consequence, and cannot notice that it caused a problem.

Any task where accountability is the actual point stays yours. Not because the writing would be worse, but because delegating the words does not delegate the responsibility.

It cannot tell you what you should be doing

It will answer any question you ask.

It will not tell you that you are asking the wrong question, that this project is a bad idea, or that the reason you are avoiding a task is worth examining.

Ask for help writing a follow-up and you get one, even when the honest answer is that this client should have been let go two months ago.

A simple way to think about it is this. It is a very capable pair of hands with no opinion about where the business is going. The strategy stays with you, and it will not notice if the strategy is wrong.

It cannot do relationships

The apology that lands. The difficult conversation. The thank-you that meant something.

Being caught outsourcing any of those does more damage than the time it saved, because the entire value was that a person chose to spend attention on another person.

It does not learn from your corrections

You fix the same thing in its output ten times. The eleventh time, it does it again.

Within one conversation it will adapt. Across conversations, nothing carries over unless you carry it.

This surprises people more than anything else on this list, because it works nothing like training a new employee. The way to make a correction stick is to write it into the instructions you reuse. Corrections spoken once vanish.

It cannot verify itself

Ask whether its answer is correct and it will produce a confident assessment, which is generated the same way the original answer was.

Asking it to check its own work catches formatting problems and internal contradictions. It does not catch invented facts, because nothing in the process ever touched the world to find out.

Verification requires something outside the conversation. A source, a system, a person.

It does not know what happened recently

Its knowledge has a cutoff, and it does not know what changed after it. Some tools search the web to fill the gap, which helps and introduces a second problem: now it depends on whatever it happened to find.

For anything current, prices, policies, product features, laws, treat it as a starting point and confirm elsewhere.

What this leaves

A long list of limits, and yet the useful territory is still large.

It cannot decide what matters, but it can process everything so you decide faster.

It cannot take responsibility, but it can draft what you will be responsible for.

It cannot remember you, but it will work tirelessly from context you provide.

It cannot tell you what to do, but it will do a great deal of what you have decided.

None of that is small. It is just narrower and more specific than the way it gets sold to you, and knowing the shape of it is what separates people who get value from people who tried it twice and gave up.

The edges are not the disappointing part. They are the map.

Free resource

Start here

The free AI Readiness Auditor separates the parts of your workflow that hand over cleanly from the parts that never will.

Open the free AI Readiness Auditor

Keep reading

AI at WorkCan AI reply to my customers?

Yes, but not the way most people mean. Here is what is safe to automate, what should never be, and where the line sits.

8 min read
AI at WorkHow much time does AI actually save?

Less than the headlines claim, more than the skeptics say, and not always on the tasks you expect. How to measure it honestly in your own business.

7 min read

Custom build

Want this done for your business?

IdeaForgeLabs can turn this into real pages, structured data, and workflows that ship.

Work with me