AI as the person sitting beside me

Why I’m finding AI most useful when it works with me, not instead of me

Over the last few weeks, I’ve found myself using AI in a slightly different way.

Not to write something for me. Not to generate an image. Not even simply to answer a question.

Instead, I’ve had ChatGPT open on one screen while I work on another.

It sounds simple, but I’m beginning to think this may be one of the most useful ways I’ve found to work with AI.

Recently, I was testing a Google Workspace setup I have been investigating for my wife's business and trying to understand some Gmail configuration, email aliases, and related settings. None of it was completely beyond me, but there were enough unfamiliar screens, buried options and bits of terminology to make it very easy to go wrong.

The traditional way to tackle something like this would probably be to find a YouTube video or start digging through Google’s documentation. Both have their place, but they also have limitations.

A YouTube video moves at somebody else’s pace. You pause it, rewind it, skip forwards, then realise the screen they are showing is slightly different from yours.

Documentation can be awkward in a different way. The information may all be there, but you still have to work out which part applies to your situation, in what order to follow it, and whether the instructions still match the system as it exists today.

So instead, I tried something more conversational. I put ChatGPT on one screen and Google Workspace on the other, then worked through the setup step by step. That turned out to be far more helpful than I expected.

One step at a time

The biggest difference was the pace. Rather than reading a whole page of instructions and trying to hold it all in my head, I could do one step, check the result, then move on to the next. That made the process feel much more natural.

More importantly, it helped the logic of what I was doing sink in. I wasn’t simply copying a sequence. I could see how each choice connected to the result I was trying to achieve.

For me, that difference matters.

There is a world of difference between getting through a task and actually learning something from it.

The screenshots made all the difference

The really useful part came when the instructions and the reality stopped matching. AI is often working from what it expects a system to look like. That may come from documentation, previous examples or knowledge of how a system usually works.

But software changes.

Menus move.

Buttons appear or disappear.

Names get altered.

Screens can look different depending on subscription, account type or what has already been configured.

That is often the point where an ordinary tutorial starts to fall apart.

You get to step five or six, cannot find what the instructions are talking about, and suddenly you're down a rabbit hole of searches trying to figure out what changed.

What made this different was that I could show ChatGPT what I was actually seeing.

I could send a screenshot and say, in effect:

This is the screen in front of me.

That changed the whole experience. Instead of me trying to interpret the problem alone, ChatGPT could look at the evidence, compare it with what it expected and adjust the guidance.

A small Gmail problem that proved the point

One particular example really brought this home.

I was trying to get Gmail to reply from the same alias that had originally received the message.

The setting looked correct, but when I tested it, Gmail still wanted to reply from my main address.

The first suggestion was reasonable enough: perhaps the setting needed saving. Except there was a problem.

There was no Save Changes button where ChatGPT expected one.

So I sent a screenshot. That immediately changed the conversation. Rather than carrying on as though the standard instructions must be right, ChatGPT acknowledged what I was seeing and suggested a different route through the settings.

That didn’t quite solve it either.

On another screen, there was a Save Changes button, but it was greyed out.

Another screenshot. Another rethink.

Eventually, the answer wasn't another hidden setting. The issue was how I was testing it.

We realised we needed to send a completely new incoming message to the alias after the setup was complete, rather than replying to an older test email.

That worked.

The interesting thing is not that AI somehow knew the answer all along. It didn’t.

What made the process useful was that it could respond when the evidence changed. It could reconsider its assumptions, adjust the next step and keep working through the problem with me until the result made sense.

That is much closer to how real troubleshooting actually works.

It reminded me of how things used to be learned

This stood out to me partly because I spent much of my working life in IT. Earlier in my career, learning something new often meant physical books, long manuals and trial and error.

Before Google became the obvious first stop, getting stuck in a system configuration usually left you with two choices: ring somebody who knew more than you did, or keep trying things until you worked it out yourself.

There was value in that. Trial and error teaches you a lot.

But it could also take a long time. Search engines changed things by making information easier to find.

AI feels like another step on from that.

The information is not just searchable. It is discussable. You can work through it. You can challenge it. You can tell it that what it expects is not what is on the screen, and the conversation can move on from there.

That is a very different experience from simply looking something up.

AI needs to be allowed to be wrong

This also made me think about something else. For AI to be genuinely useful, it has to be allowed to be wrong. That may sound like a criticism, but I mean the opposite.

The Gmail example worked only because the AI didn’t stubbornly insist the first answer must be correct.

It expected a button that wasn’t there.

It suggested one route, then another.

Some of those assumptions didn't match what I was actually seeing.

And that was fine.

The useful part was not that it never made a mistake. The useful part was that it could reconsider when new evidence appeared. I think that matters a great deal if AI is going to become genuinely helpful in everyday work.

Trust does not come from pretending that AI always knows best.

Trust comes from being able to test its advice against reality, challenge it when necessary and refine the answer together.

Humans are not infallible either

Of course, the same applies to us. Humans make mistakes too.

I can misread a screen, misunderstand what I am looking at or explain something badly.

So a proper partnership between AI and humans can't simply mean the human checks the machine and the machine quietly obeys. It has to be more balanced than that.

The AI may have access to a huge amount of information, but I have the real screen in front of me.

The AI may know how a system usually behaves, but I can tell it how it is behaving now.

It can suggest the next step. I can test the result.

Then both sides can adjust.

That kind of mutual checking feels much more realistic, and much more useful, than all the grand talk about AI replacing human thinking.

Where this has become useful for me

Since then, I’ve noticed the same pattern in other situations too. I used ChatGPT in a similar way while setting up a response form on my Wix website. I also used it while sorting out email authentication within a CRM, where getting the settings wrong could affect whether emails were trusted or pushed into spam.

In each case, the value was not just in being told an answer. It was in working through an unfamiliar system step by step, with the guidance adapting as the situation unfolded.

That feels much more like having a knowledgeable person sitting beside me than reading a fixed tutorial.

Partnership is the part that interests me most

There is a lot of talk about AI at the moment, much of it focused on big claims.

What it might replace.

What it might automate.

What jobs it might do.

What interests me more is something smaller and, for me at least, more useful. What happens when AI helps me do the job myself?

I don’t really want to trade understanding for convenience. If I am setting up a system, I still want to know what is going on. I still want to learn something from the process. I still want to recognise when something is not right.

Used this way, AI seems to support that rather than weaken it.

One screen contains the problem.

The other contains a patient assistant that can suggest a next step, respond to evidence and help me think my way through something without pretending to be infallible.

That feels like a much better model. Not the machine that knows everything. Not the human who blindly follows it.

A partnership where both sides bring something useful, both sides can be wrong, and both sides improve the result by checking and adjusting together.

If that is where AI is heading, I think it may become far more valuable than simply another place to look up answers.