The Day AI Couldn't See What We Could

The Day AI Couldn't See What We Could

Today we asked AI to explain one of our own retirement scenarios.

It got the answer wrong.

Not because the calculations were wrong.

Not because the maths was wrong.

Because it couldn't see what we'd typed.

That wasn't the result we expected.

So we started investigating.

The numbers were correct.

We changed the annual spending figure from £30,000 to £3,000.

The retirement planner updated immediately.

The calculations changed.

Everything worked exactly as it should.

But when Microsoft Copilot explained the scenario...

...it still believed annual spending was £30,000.

The AI was confidently reasoning from information that no longer existed on the screen.

We wanted to know why.

After opening the browser developer tools we discovered something interesting.

The live page knew the value had changed.

The original HTML still contained the default figure.

Humans were reading the live application.

The AI appeared to be reading a snapshot taken before anything had changed.

In other words...

We weren't both looking at the same information.

Nobody was wrong.

The retirement planner wasn't broken.

The browser wasn't broken.

The AI wasn't broken.

Interactive websites simply weren't designed with AI readers in mind.

People interact with applications.

Some AI assistants appear to inspect the underlying page rather than the application's current state.

That's a subtle difference.

But it changes everything.

It explained something we'd been thinking about.

For a while we've been describing Wayli as a place to prepare before you prompt.

Until today, that was mostly philosophy.

Now it's engineering.

If your AI starts from different information than you do, the quality of the conversation suffers before you've even asked your first question.

The prompt wasn't the problem.

The context was.

This changes our roadmap.

We've already planned a feature called Copy Wayli Summary.

Originally, it was about convenience.

Now it's about reliability.

Instead of asking AI to interpret an interactive application, Wayli will give you a structured summary containing your scenario, assumptions and conclusions.

You and your AI will start from exactly the same understanding.

That's a much better place to begin.

Building in public.

One of the advantages of building Wayli in public is that we get to share discoveries like this.

Some experiments confirm what we already believed.

Others change the direction of the product.

This one did both.

It reminded us that AI isn't just about asking better questions.

It's about making sure everyone is looking at the same facts before the conversation even begins.