But it’s knowing when not to use the biggest model.
If you use the same AI model for every task, you probably aren’t saving time. You’re just outsourcing your judgement.
And I think this will become one of the most practical skills in the AI era.
Not prompting.
Not memorising every new tool.
Not chasing every model launch.
Judgement.
Knowing what you are trying to achieve.
Knowing how much accuracy the task actually requires.
Knowing when speed matters more than intelligence.
Knowing when a cheaper model is good enough.
And knowing when AI should not be trusted without human review.
That, to me, is what AI literacy is starting to look like in the real world.
The AI Race Is Making Us Ask the Wrong Question
Every week, there seems to be a new model.
Smarter.
Faster.
Cheaper.
Better at reasoning.
Better at coding.
Better at research.
Better at whatever the previous model was supposedly bad at.
So naturally, people ask:
“Which AI is the best?”
But I think that is often the wrong question.
The better question is:
“Which AI is good enough for this particular job?”
Because “best” without context is almost useless.
Would you use a Formula 1 car to buy groceries?
Technically, it might be impressive.
Practically, it would be ridiculous.
The same thing happens with AI.
People use their most powerful model to summarise a short email.
Then they complain about cost.
They use a fast, lightweight model for a complex business decision.
Then they complain about hallucinations.
They expect AI to make a strategic judgement when they haven’t given it enough context.
Then they blame the tool.
Sometimes the problem isn’t the AI.
Sometimes the problem is the person choosing how to use it.
And I think that distinction will matter more and more.
AI Literacy Is Becoming Less Academic and More Practical
When people hear the term AI literacy, they may imagine something technical.
Understanding neural networks.
Knowing how models are trained.
Learning machine learning terminology.
That knowledge can certainly be useful.
But I don’t think practical AI literacy requires everyone to become an AI engineer.
For most people, I see it more like this:
Can you look at a task and make a sensible decision about how AI should help you do it?
That is a practical skill.
Imagine you have five jobs to complete today.
-
- Summarise a 30-page document.
- Brainstorm 20 campaign ideas.
- Analyse why your sales conversion has dropped.
- Rewrite 500 product descriptions.
- Review an important contract before sending it to your legal team.
Should all five jobs use the same AI model?
I don’t think so.
They don’t carry the same level of complexity.
They don’t have the same consequences if the answer is wrong.
They don’t need the same speed.
And they certainly don’t need the same budget.
This isn’t just my personal theory.
Even the companies building AI models increasingly frame model selection around trade-offs.
OpenAI’s current model guidance, for example, distinguishes between frontier capability for difficult work, a balance of intelligence and cost, and efficient models for cost-sensitive or high-volume workloads.
Its guidance is based on the idea that the model should fit the workload rather than assuming the strongest model is automatically the right choice.
That sounds obvious.
But in practice?
Most people still choose AI the same way they choose a restaurant.
They pick the one they happen to know.
The “Best AI” Is Usually an Expensive Way to Think About the Problem
Here’s the assumption I would challenge:
The smarter the model, the better the outcome.
Not necessarily.
A smarter model may be unnecessary.
It may cost more.
It may take longer.
And if the task is simple, the difference in output may be meaningless.
Stanford’s 2025 AI Index reported just how dramatically this economics has changed.
The cost of querying an AI model performing at the level of GPT-3.5 on the MMLU benchmark fell from $20 per million tokens in November 2022 to $0.07 by October 2024, a reduction of more than 280 times in roughly 18 months.
At the same time, smaller models are becoming far more capable.
The AI Index noted that in 2022, the smallest model exceeding 60% on MMLU had 540 billion parameters, while by 2024 a 3.8-billion-parameter model had crossed the same threshold.
To me, this makes one thing clear.
The question is no longer simply whether AI is powerful enough.
The question is whether you are good enough at deciding how much power the task requires.
That is a very different skill.
A Simple Test I Keep Coming Back To
Whenever I look at an AI task, I mentally ask four questions.
How difficult is the job?
Is this simple repetition?
Or does it require deep reasoning?
For example, extracting names from a spreadsheet is very different from analysing why a business strategy failed.
One needs consistency.
The other may require judgement, context, competing hypotheses and deeper reasoning.
Why pay for a brain surgeon to put a plaster on your finger?
Use the capability you actually need.
How expensive is a mistake?
This is probably the question people underestimate most.
If AI gives me five mediocre headline ideas, nothing catastrophic happens.
I choose another one.
If AI makes a mistake in a high-stakes financial, legal, medical or strategic decision, the consequences are very different.
So model selection shouldn’t only be about output quality.
It should also be about risk.
A simple task with low consequences can tolerate a simpler, faster and cheaper system.
A complex task with serious consequences deserves more capability and more human review.
Notice that I said more human review.
Not zero review.
This is where people sometimes misunderstand AI literacy.
Knowing how to choose a good model doesn’t mean blindly trusting the output.
It means knowing where the responsibility still belongs.
With you.
How often will I do this?
This is where cost becomes much more interesting.
If I use a premium model once to help me think through an important business problem, I may not care much about the cost.
But what if I need AI to process 100,000 customer requests?
Now a tiny difference in cost can become significant.
What if the workflow needs to respond in seconds?
Now latency matters too.
This is why “the best model” is not a fixed answer.
The best model for one important decision may be completely wrong for a high-volume repetitive workflow.
OpenAI’s practical guidance on building AI agents makes this point clearly, different models have different trade-offs in task complexity, latency and cost, and not every task requires the most capable model.
The recommended approach is to establish a quality baseline, then test smaller models to see whether they still meet the required standard.
I think that is useful advice beyond AI agents.
It is a useful way to think about AI in general.
Start with the outcome you need.
Then choose the cheapest and fastest approach that can reliably get you there.
Not the other way around.
What exactly am I asking AI to do?
This might be the most important question of all.
Because many AI frustrations are not actually model problems.
They are thinking problems.
Someone says:
“AI gave me a bad answer.”
Okay.
What did you ask it?
“I asked it to create a marketing strategy.”
For what company?
What market?
What customer?
What price point?
What objective?
What constraints?
What previous strategy?
What resources?
What definition of success?
Silence.
Then the AI becomes the villain.
This is why I don’t think AI literacy is simply about knowing which button to press.
It is about understanding the job well enough to direct the tool.
If you don’t understand the problem, AI can help you produce an answer.
But how will you know whether it is a good one?
The Real Skill Is Becoming a Better Editor of Machines
I think the future will create a strange advantage.
The people who benefit most from AI may not always be the people who know the most about AI.
They may be the people with the best judgement.
The best editor can recognise weak writing.
The experienced salesperson can recognise a generic sales pitch.
The marketer can spot when an “insight” is just an obvious observation dressed up in impressive language.
The business owner can see when a recommendation ignores commercial reality.
AI can generate.
But someone still has to decide:
Is this useful?
Is this true?
Is this relevant?
Is this good enough?
What is missing?
That is why I believe human expertise becomes even more valuable when AI becomes more accessible.
Not because AI cannot do impressive things.
But because more output creates a bigger need for judgement.
McKinsey’s research on the changing workplace makes a similar argument, as AI takes on more routine digital tasks, people will increasingly need to focus on asking better questions, interpreting results, guiding machines and exercising judgement.
That, to me, is the real opportunity.
We are moving from a world where value came from being able to produce something manually…
…to a world where value increasingly comes from knowing what should be produced, how it should be evaluated, and when the machine is wrong.
Stop Collecting AI Tools Like Pokémon
Another thing I find amusing is how AI adoption has become a collection hobby.
Someone uses one AI for writing.
Another for presentations.
Another for research.
Another for meetings.
Another for images.
Another for automation.
Another because someone on LinkedIn said it would “change everything.”
Before long, they have 17 subscriptions and no better workflow.
I understand the temptation.
New tools are exciting.
But more tools don’t automatically mean more productivity.
Sometimes they just mean more tabs.
McKinsey reported in 2025 that while nearly all companies were investing in AI, only 1% of leaders considered their organisations mature in AI deployment—meaning AI was fully integrated into workflows and driving substantial business outcomes. The same research found strong familiarity with generative AI among employees and leaders, but familiarity clearly isn’t the same thing as operational maturity.
That statistic should make us pause.
Almost everyone is experimenting.
Very few have fully figured out how to turn it into repeatable business value.
Why?
Probably because buying access to AI is easier than redesigning how you work.
A new tool takes five minutes to sign up for.
Building judgement takes much longer.
My Rule: Don’t Ask “Which Tool?” First
When I approach a task, I try not to begin with:
“Which AI should I use?”
I start with:
“What is the actual job?”
Then:
“What would a good result look like?”
Then:
“What happens if it gets this wrong?”
Then:
“Do I need this quickly, cheaply, or at the highest possible quality?”
Only after that does the tool question become useful.
Because AI tools are increasingly interchangeable in some situations.
The workflow is not.
The judgement is not.
Two people can use the exact same model.
One saves three hours.
The other wastes three hours correcting nonsense.
Same AI.
Different outcome.
The difference is rarely explained by a lack of access.
It is explained by how well the person understands the work.
There Will Never Be One Model for Everything
I think we need to get comfortable with this.
There probably isn’t going to be one perfect AI model for every situation.
And even if one model becomes remarkably good at almost everything, that still doesn’t make it the most efficient choice for everything.
A complex strategic analysis may justify deeper reasoning.
A repetitive classification task may not.
A high-volume customer workflow may prioritise speed and cost.
A one-off critical decision may prioritise capability.
A creative exercise may benefit from generating many cheap alternatives.
A sensitive or high-risk task may require tighter controls and substantial human involvement regardless of model strength.
Different jobs.
Different trade-offs.
Different answers.
And that is why I think the most important AI skill isn’t loyalty to a particular model.
It’s **model judgement**.
The ability to look at a job and say:
This needs more intelligence.
This doesn’t.
This needs speed.
This needs scale.
This needs human review.
This isn’t an AI problem at all.
That last one may be the most valuable judgement of all.
The Tool Will Keep Changing. Your Judgement Is the Portable Skill.
Today’s favourite model will eventually be replaced.
Today’s “must-have” AI tool may disappear.
Prices will change.
Capabilities will improve.
New companies will enter.
Old leaders will fall behind.
If your AI skill is simply knowing the name of today’s most popular tool, you may need to relearn everything every few months.
But if your skill is understanding tasks, evaluating trade-offs and judging output quality?
That skill travels.
From one model to another.
From one company to another.
From one AI generation to the next.
This is why I think AI literacy is becoming a practical business skill rather than a technical speciality.
The future won’t only reward people who can use AI.
It will reward people who can use AI appropriately.
And there is a huge difference.
Anyone can ask AI for an answer. The real skill is knowing how much intelligence the question deserves—and whether the answer deserves your trust.
That’s the part I think we should spend more time learning.
Not just how to prompt.
But how to judge.
Not just how to use the tool.
But when to use which tool.
Because as AI gets cheaper, faster and more widely available, access will matter less.
Judgement will matter more.
What do you think?
When you use AI today, do you consciously choose the model based on the job—or do you simply use whichever tool is already open?
And if you are trying to figure out where AI actually fits into your marketing, customer journey or business workflow, sometimes the biggest opportunity isn’t adding another tool.
It’s getting clearer about the problem you’re trying to solve first.