Your bank’s fraud model spotted a stolen card number in 40 milliseconds this morning. Your phone’s keyboard predicted the next word before you finished typing it. Neither system knows what money is, or what a word is. Both did their job flawlessly. That’s narrow AI in a nutshell: extraordinary competence inside a box so small it barely resembles intelligence at all.
AGI is the opposite bet. Not better, not faster, not trained on more data. Different in kind. And the confusion between the two shapes everything from how companies buy software to how regulators write law.
Narrow AI: the specialist that already runs your day
Narrow AI is any system built for one task family, and only that task family. It has no interest in your other problems. A recommendation engine that decides what you watch tonight cannot help you write a resignation letter. A speech-to-text model that handles 99 languages still cannot tell you whether the sentence it transcribed is a lie.
These systems are everywhere because they are genuinely excellent at their one thing. Visa’s network scores tens of thousands of transaction messages per second, most of them in milliseconds. Google Translate moves text across more than 130 languages. AlphaFold produced predicted structures for roughly 200 million proteins, a body of work that would have taken experimental labs centuries. Each one is a specialist. A very, very good specialist.
Business adoption has followed the same pattern. If you want a sense of how deep narrow AI has already sunk into normal operations, these charts on the current state of AI deployment are more useful than any think-piece, mostly because they count things instead of speculating about them.
Narrow does not mean small
This is where people trip up. GPT-class language models, image generators, coding assistants: all narrow by the technical definition. They are trained to do a particular thing (predict tokens, denoise images, complete code), even though that one thing turns out to be useful across dozens of contexts. Capability is not the same as generality. A system can be broad in application and still be a specialist at heart.
AGI: general by design, not by accident
Artificial general intelligence describes a system that can learn and perform essentially any intellectual task a human can, in an environment it has never seen, without being retrained for each new problem. The emphasis sits on learn. A person who has never played Go can pick up the rules over a weekend and lose respectably. A Go engine trained without those rules cannot play at all.
The definition is genuinely contested. Some researchers anchor it to human-level performance on a broad battery of cognitive tests. Others care about economic substitution: could this system hold down a job it wasn’t built for? If you want to see how wide the disagreement runs, the many competing meanings of artificial general intelligence are worth reading before you argue about timelines with anyone.
Generality, not scale, is the dividing line
Stack up every benchmark and one pattern holds. The moment a task requires transferring a skill to an unfamiliar setting, systems that looked brilliant fall apart. Humans score around 85% on ARC-AGI, a set of visual reasoning puzzles designed to resist memorisation. For years, the best AI systems sat in low single digits. Newer reasoning models have closed much of that gap, which is exactly why researchers keep building harder tests. The target keeps retreating because the thing being measured is real.
A side-by-side look
- Scope: narrow AI handles one task family. AGI would handle any cognitive task you hand it, including ones nobody anticipated.
- Learning: narrow systems need large task-specific datasets and retraining. AGI would pick up a new skill from a handful of examples, the way a new hire does.
- Transfer: a chess engine cannot debug your JavaScript. An AGI could learn chess, then apply the pattern-recognition habits it picked up to something unrelated.
- Autonomy: narrow tools sit and wait for a prompt. AGI would set its own sub-goals, notice when a plan is failing, and keep working across days.
- Failure modes: narrow AI usually fails inside its own lane, and often loudly. AGI failures would be unexpected by definition, which is the uncomfortable part.
- Evaluation: you can benchmark narrow AI with a held-out test set. Nobody has a clean, agreed test for “general.”
The confusing middle: general-purpose is not general
Modern chatbots feel general. The same model drafts an email, explains a tax rule, writes a Python script and translates a contract. That range comes from a single training objective applied to an enormous corpus, not from an ability to reason about the world. Ask one to keep a project moving for three weeks and the illusion dissolves: no durable memory, no stable goals, no grounding in anything beyond text patterns.
That distinction matters commercially. Right now you can buy a product that behaves impressively broadly while remaining entirely narrow underneath. One is a product, the other is a research bet, and mistaking the first for the second leads to bad contracts and worse expectations. The same applies inside engineering teams: a coding assistant wired into team chat is a superb narrow tool, but it is a tool. The shift toward collaborative coding inside Slack is a productivity story, not a step on the road to generality.
Why the gap between them is not academic
Three practical consequences follow from knowing which side of the line a system sits on.
Procurement gets clearer. If a vendor sells you “AI,” ask what task it was trained on and what happens when your data drifts. Narrow systems degrade predictably, and you can plan for that with monitoring and retraining schedules. A system claimed to be general gives you no such handle.
Regulation is already splitting along this line. The EU AI Act carves out separate obligations for general-purpose models versus narrowly scoped, high-risk applications. Different rules, different paperwork, because the risks genuinely differ.
Safety costs scale with generality. The further up the curve a model sits, the more a mistake is worth preventing. OpenAI slowing its own frontier work on Astra is a useful illustration: capability advances create pressure to pause and check, and that pressure does not exist for a spam filter.
How you would actually know AGI had arrived
Forget the Turing test. It has been passed, gamed and argued over for a decade. Better signals exist. Steve Wozniak’s coffee test asks a machine to enter an ordinary kitchen and brew a cup, which requires object recognition, navigation, planning and hands. ARC-AGI asks for novel reasoning under strict constraints. The most practical test of all is economic: drop the system into a job it was never trained for, hand it a two-page brief, and check whether it performs at median-human level eight weeks later.
Nothing on the market comes close to that last one. What we have are extraordinarily capable narrow systems stacked into pipelines that look general from the outside. An agent that books flights, writes the itinerary and files the expense report is three or four specialists in a trench coat, coordinated by scaffolding code.
So the useful question is not “is this AGI?” but “how much of this year’s work can I hand to a specialist, and where does the handoff break?” Pick a process, measure the baseline, deploy the narrow model, watch the failure cases. That discipline is what separates teams getting real value from AI right now from teams waiting for a machine that thinks like a person. And if the wait turns out to be long, the narrow deployment still paid for itself.