Imagine that artificial intelligence doesn’t get dramatically better from here.
No sudden breakthrough.
No artificial general intelligence.
No machines capable of doing everything a human can do.
Just the technology we already have, gradually finding its way into businesses.
That would still be significant.
In less than two years, generative AI has progressed from a text interface that surprised people with its ability to write and answer questions to systems capable of working across text, images, audio, video and increasingly enormous amounts of information.
In May, OpenAI demonstrated GPT-4o responding to speech in something approaching the rhythm of a human conversation while reasoning across audio, vision and text. Google has made Gemini 1.5 Pro capable of processing up to one million tokens of context — enough to work with multiple large documents, substantial codebases, audio or video in a single interaction. Anthropic’s Claude 3.5 Sonnet, released last month, improved again across areas including reasoning, coding and visual understanding.
There is already plenty for businesses to work out.
But there is another possibility worth considering.
What if this is just the beginning?
Not because anybody can reliably predict what AI will be capable of in five years.
They can’t.
But because after the rate of change we’ve just witnessed, assuming today’s capabilities are close to the endpoint seems like a surprisingly confident prediction of its own.
AI doesn’t need to become AGI#
A lot of discussion about the future of artificial intelligence eventually arrives at AGI: artificial general intelligence.
Will machines reach human-level intelligence?
When?
What happens afterwards?
They’re fascinating questions.
I’m not convinced they’re the questions most businesses need to answer.
AI doesn’t need to become more intelligent than every human being to have profound economic consequences.
It doesn’t need consciousness.
It doesn’t need to autonomously run a company.
It doesn’t even need to eliminate humans from most workflows.
It just needs to become good enough at enough economically valuable tasks.
Imagine systems that become somewhat better at research.
Somewhat better at writing software.
Somewhat better at analysing information.
Somewhat better at understanding customers.
Somewhat better at creating content.
Somewhat better at translating languages.
Somewhat better at helping people make decisions.
Now combine those improvements with systems becoming faster, cheaper, easier to use and capable of working with more of a company’s information.
No individual improvement needs to transform the economy.
The combination might.
AI doesn’t need to replace human intelligence to change the economy. It only needs to make useful intelligence significantly less scarce.
And scarcity is where this becomes a business question.
Businesses ration intelligence#
Businesses ration scarce resources all the time.
Capital is scarce.
Time is scarce.
Expertise is scarce.
Management attention is scarce.
Software-development capacity is scarce.
Analysis is scarce.
Good customer service is scarce.
Much of what a business would like to do doesn’t happen because the value of doing it doesn’t justify the human effort required.
A company could analyse every customer conversation manually.
It usually doesn’t.
It could produce individually researched communications for every prospect.
It usually doesn’t.
It could ask a developer to build software for every irritating internal process.
It usually doesn’t.
It could have an analyst investigate every small decision.
It usually doesn’t.
It could translate everything it produces into twenty languages.
It usually doesn’t.
The constraint isn’t necessarily that these things have no value.
Often, they simply cost too much.
Human cognitive effort is expensive.
We therefore reserve it for problems important enough to justify it.
This is why I think one of the most interesting questions about AI isn’t simply which jobs it might automate.
It’s what happens to all the work we currently choose not to do at all.
Falling costs change behaviour#
We’ve seen this pattern with other technologies.
When something becomes dramatically cheaper, we don’t necessarily consume the same amount and pocket the saving.
We find new uses for it.
Computing is an obvious example.
If the cost of computing had remained where it was several decades ago, we wouldn’t simply own more expensive smartphones.
Huge categories of software and digital services wouldn’t exist because their economics wouldn’t make sense.
The same principle applies inside a business.
Suppose analysing a particular dataset requires ten hours of an experienced employee’s time.
Perhaps the insight is worth £200 to the company but costs £500 to produce.
So nobody does it.
Reduce the cost of performing that analysis to £20 and the calculation changes.
Nothing has been “automated” in the dramatic sense.
No job necessarily disappears.
An activity that previously wasn’t economically rational simply becomes worth doing.
That distinction could become extremely important with AI.
The economic opportunity may not only be:
How can we perform the work we already do with fewer resources?
It may increasingly become:
What useful things don’t we currently do because intelligence is too expensive to apply to them?
That’s a much larger search space.
Capability is only one curve#
There is a tendency to measure AI progress by asking whether the latest model is smarter than the previous one.
Businesses should probably watch several curves.
Capability is one.
Cost is another.
Speed is another.
The amount of information a model can work with is another.
Accessibility is another.
Reliability is another.
In May, OpenAI said GPT-4o matched GPT-4 Turbo’s performance on English text and code while being faster and 50% cheaper through its API.
Google introduced Gemini 1.5 Flash specifically as a lighter model designed for speed and efficiency at scale, while Gemini 1.5 Pro’s long context window expanded how much information a model could process at once.
Anthropic’s Claude 3.5 Sonnet offers another interesting example. Anthropic reported that it outperformed its previous flagship Claude 3 Opus across a range of evaluations while operating at the speed and price of its mid-tier Sonnet model.
The precise benchmark leader will change.
That’s almost beside the point.
What’s interesting is the direction.
We are seeing competition not simply to produce more capable intelligence, but to make increasingly capable intelligence faster, cheaper and easier to apply.
For businesses, those improvements multiply.
Something that wasn’t useful enough can become useful.
Something that was too slow can become practical.
Something that was too expensive can become economical.
Something that couldn’t understand enough context can become viable once it can work across more of the relevant information.
The boundary of the business opportunity moves even if no single breakthrough occurs.
Abundance creates different businesses#
This is where following the trajectory becomes more interesting than predicting individual AI products.
Imagine useful machine intelligence becoming substantially more abundant.
What changes?
Perhaps a small company can afford analytical capabilities previously available only to businesses with teams of analysts.
Perhaps every customer interaction can be examined rather than a sample.
Perhaps software gets built for internal problems that would never previously have justified development.
Perhaps every salesperson can research every account.
Perhaps every employee can interrogate thousands of pages of company information.
Perhaps products can be personalised for individual customers where previously businesses could only afford broad segments.
Perhaps ten ideas can be tested where previously there was time to test one.
None of these requires a science-fiction version of AI.
Some are already possible in limited forms.
The important question is what happens as the quality improves and the cost of applying it falls.
Businesses designed around scarce intelligence naturally develop processes for rationing it.
Managers decide which questions deserve analysis.
Development teams prioritise which software is worth building.
Marketing teams decide which customers deserve personalisation.
Customer-service organisations decide how much human attention different enquiries can economically receive.
If the economics of intelligence change, some of those assumptions change with them.
And eventually, businesses designed around the old constraint may start to look very different from businesses designed around the new one.
There are good reasons to be sceptical#
None of this means the current trajectory continues indefinitely.
It might not.
AI systems remain unreliable in important ways. They can invent information, misunderstand instructions and fail unpredictably.
The Stanford AI Index published earlier this year found that AI still trails humans on more complex tasks including competition-level mathematics, visual commonsense reasoning and planning.
Training frontier models is also becoming extraordinarily expensive. Stanford estimates the compute used to train GPT-4 cost around $78 million, while Gemini Ultra’s training compute cost approximately $191 million.
There may be technical limits ahead.
There may be economic ones.
Regulation could slow deployment. Energy and infrastructure constraints could matter. Businesses may discover that integrating AI reliably into real workflows is considerably harder than the technology demonstrations imply.
And human judgement may prove stubbornly difficult to reproduce in areas where context, trust, accountability and experience matter.
All of that is possible.
But scepticism cuts both ways.
There is uncertainty in assuming AI progresses rapidly from here.
There is also uncertainty in assuming it doesn’t.
And after the changes of the last eighteen months, I don’t think businesses should treat the second assumption as inherently more sensible simply because it feels more conservative.
Strategy happens before certainty#
The uncomfortable thing about strategy is that you don’t get to wait for the answer.
Businesses make decisions while the future is unresolved.
They hire people.
Build products.
Choose technology.
Design processes.
Allocate capital.
Enter markets.
Develop capabilities.
Their competitors do the same.
By the time the consequences of a technological shift become completely obvious, many of the decisions that determine who benefits from it have already been made.
That doesn’t mean betting the company on the most aggressive prediction about artificial intelligence.
It means asking a more useful question.
What would we do differently today if useful intelligence were going to become significantly more capable, significantly cheaper and significantly more abundant over the next few years?
Perhaps the answer is very little.
Perhaps it changes everything.
But it’s becoming a question worth asking.
Because the most consequential possibility isn’t necessarily that AI suddenly becomes more intelligent than us.
It may be much simpler than that.
Useful intelligence could just keep getting better.
And cheaper.
And easier to access.
The risky assumption may not be that AI keeps improving. It may be that this is as good as it gets.