AI Fluency: How Much Do the Rest of Us Actually Need to Understand?

Over the past few weeks, I’ve found myself having versions of the same conversation with people across business, finance, wealth management, technology and education — as well as with fellow parents.

Different conversations, but I keep coming back to the same question:

How much do the rest of us actually need to understand about AI?

I switched on a light this morning without thinking about electrons moving through a wire.

I’ve already used the internet countless times today without thinking about what TCP/IP is doing underneath it.

That’s one of the triumphs of technology: complexity disappears beneath the interface.

So perhaps AI should be the same.

I don’t think it is.

This time, the machine is participating in the thinking

Electricity transformed what became possible in the physical world. The internet connected people and information at extraordinary scale.

AI adds something different.

The resource becoming increasingly accessible and affordable is intelligence.

Think about how much of professional life ultimately depends on cognitive work.

Lawyers interpret evidence. Accountants analyse financial information. Consultants diagnose problems. Investment professionals evaluate companies and markets. Doctors interpret symptoms and scans.

The domains are very different. But sitting upstream of all of them is something

remarkably similar: the ability to interpret information, reason about it and exercise judgement.

And that is precisely the territory AI is beginning to enter.

Perhaps a useful way to think about AI in the workplace is as an extraordinarily capable assistant to whom we can increasingly outsource parts of our thinking.

It can read, research, analyse, interpret, reason, create, challenge and recommend — doing in seconds or minutes what might once have taken teams days or weeks.

But there’s one rather important catch.

The assistant can do the work. It can’t take the accountability.

And that changes the bargain.

Today’s AI systems are largely probabilistic. They can reason impressively without reasoning infallibly. They can infer when we wanted certainty, generate when we wanted retrieval, and produce something completely plausible that is nevertheless wrong.

As AI takes on more of this work, we can no longer simply throw technology over the wall to IT.

Business leaders don’t need to become AI engineers. But they increasingly need to understand enough about these systems to bring their own judgement into the decisions surrounding them.

That’s where AI Fluency begins for me.

AI is entering the language of business

Throughout my career, I’ve seen organisations wrestle with bridging the gap between business and technology.

AI is making that boundary harder to maintain.

Imagine a finance team redesigning its forecasting workflow.

The technology lead says:

“We’ll use RAG over the finance knowledge base, retrieve the latest actuals and forecasts into the LLM’s context window, then generate the variance narrative.”

The response?

“Sounds good. When can we try it?”

Sound familiar?

An AI-fluent finance leader might instead say:

“Fine for the narrative. But the numbers need to come from the system of record, and the variance calculation should be deterministic.”

They haven’t become an AI engineer. They simply understand enough to combine their finance expertise with an understanding of where probabilistic AI is useful — and where deterministic controls still belong.

A recent conversation about wealth management drove this point home for me.

Consider an adviser using AI across an end-to-end client workflow: speech-to-text captures the conversation, an LLM interprets it alongside client information, an ML model forecasts markets, optimisation helps construct the portfolio and another AI evaluates the recommendation.

In the latest conversation, the client talks about helping their daughter buy a house in the next few years. The ML model is forecasting weaker equity performance over the same period.

It would seem perfectly reasonable for the system to connect that forecast to the client’s short-term liquidity need.

But what if the portfolio decision it’s making is actually about the client’s 20-year retirement objective?

Suddenly, the question isn’t whether the market prediction is good.

It’s whether that prediction belongs in this decision at all.

That’s not simply an AI question. It’s a wealth-management question.

The engineer understands what the system can do. The wealth professional understands what it should do.

AI Fluency gives them enough common language to turn that professional judgement into system design.

Domain judgement starts shaping system design.

And once the system is operating, another question appears.

If the adviser doesn’t understand enough about how the system reached its recommendation to interrogate it meaningfully, does clicking an “Approve” button really constitute human oversight?

Putting a human in the loop isn’t the same as putting human judgement in the loop.

AI Fluency isn’t a substitute for good system design, governance or regulation. But if humans are expected to shape, challenge or ultimately take responsibility for AI-assisted decisions, understanding enough to exercise that judgement starts to matter.

And what about parents?

The same question follows me home.

What should our children learn about AI when we can’t possibly know which tools they’ll be using by the time they enter the workforce?

As parents, it’s easy to focus on today’s questions. Should they use AI for homework?

Will it weaken their ability to think for themselves? Which jobs will still exist?

But perhaps our more important job is to help them build mental models that outlast today’s tools.

Products will change. Interfaces will change. Models will change.

A good mental model travels.

AI-fluent parents won’t have all the answers. But they may be better equipped to help their children ask the right questions about whatever comes next.

So how much is enough?

Probably far less than the people building AI.

But enough to understand why a probabilistic system shouldn’t necessarily be doing deterministic work.

Enough to recognise when professional judgement needs to shape the system rather than merely approve its output.

Enough to ask what evidence the machine used, what it might have missed and when it should stop and involve a human.

And enough to encounter a new AI capability and ask:

Which human capabilities are being replicated here — and what changes because they are?

That’s not AI expertise.

It’s enough understanding to exercise judgement.

And perhaps that’s what being AI fluent really means.