Mention Artificial General Intelligence (AGI) and the first argument is always when we’ll get it: is it next year or twenty years from now? Some say all it will take is one year, while others argue 20+ years.
The reality is we don't know when this may occur, or how — and it might not matter.
In practice, today’s productivity gaps aren’t intelligence gaps; they’re context gaps. Your model can draft perfect code, but only if it knows your private APIs. It can schedule follow-ups, but only if it sees your calendar. Until we solve the problem of information aggregation, chasing human-level cognition might be the wrong mission.
I cannot help but wonder if AGI is actually worth pursuing.
Intelligence ≠ Context
AGI that matches or surpasses humans across all cognitive tasks sounds transformative, but smarter models alone still can’t ship products or provide services. Why?
- Data blindness. Roughly 90 % of enterprise data is unstructured: emails, PDFs, call transcripts, and various other sources that the model never sees.
- Gartner puts it bluntly: 80 – 90 % of all new enterprise data is unstructured and triples in size faster than structured data.
I believe we may be chasing the wrong issue. The case for AI is to increase the productivity of society, most of which involves automating complex workflows and systems inside of companies, both big and small. The interesting case to make, however, is although AGI will be "smarter" than current LLM models we have now, it still will not be able to complete tasks.
The ability to complete a task actually doesn't lie in "intelligence", it relies on information. Even with the most complex and advanced model, it still won't tell me what meetings happened in my company, what approach to take when solving technical issues related to internal codebases, and who to reach out to for help.
Toward Task-Aware AI
So is AGI really useful if it can't move the progress of society forward?
I think we need to take a step back. Maybe we are solving for the wrong problems; trying to fit a square peg in a round hole. What we actually want is not AGI, we want focused, adaptable systems. Benjamin Mann, co-founder of Anthropic mentions that he refrains from mentioning the term "AGI" at Anthropic. He instead brings up this idea of "transformative AI".
In his eyes, the inflection point of AI is not when we achieve some form of super intelligence, but rather when we can replace >50% of jobs utilizing AI. Now, this isn't to say that improvements to our current models are meaningless, in fact I argue it is even more important. We know that GPT is capable of emulating human behavior, and maybe if it's just marginally better that is enough.
We already have inventions such as tool-calling, libraries like LangGraph, and even OpenAI's Agent mode. These models and tools are capable, but we haven't explored the space of AI enough to understand if what we want is AGI, or if it is something else.
AGI could open up the exploration of new workflows and capabilities, which would be amazing for society, but even with the most advanced reasoning model, we still need to define flows for reliability. Companies, and even individuals, will likely use these AI models by providing an input and expecting a specific output. If AGI can reason, that is great, but it really just needs to be able to complete a task.
Especially with libraries such as LangGraph, we can already add steps such as evaluation and decision trees. It may require more manual instruction in the form of code, but it achieves what we want.
The Real Bottleneck: Private Data
The problem with AGI is that even with the ability to reason, AGI will not (for example) be able to create copy for a company if it doesn't have:
- Access to the business mission
- Access to their value props
- Access to a marketing style guide
- Access to advertising analytics to see what works and what doesn't
AGI doesn't solve any of this. Companies will not give away proprietary data and leave it up for the taking, especially in an age where software could very likely become a commodity. In fact, this means companies will be even more adamant about locking down their IP than they are currently.
Even further, 54% of data that organizations view and analyze is generated internally, while 25% is generated externally, and 21% comes from a combination of the two. This means that 75% of the knowledge required for a task is off-limits for AGI, unless we supply that data through existing technologies such as RAG.
What is really slowing us down is our data. 75 % of AI initiatives fail to scale, and messy, siloed data is the top blocker. How can we expect to answer questions if we don't provide ways for models to access necessary context?
Barking Up The Wrong Tree
Sometimes in the world of AI it seems like we are solving the wrong problems. Acting without questioning. Just because it sounds like a good idea in theory doesn't mean it is in practice. AGI will not help companies automate flows anymore than current models without the right resources, and companies are not comfortable with opening up the knowledge of internal systems and processes to a general intelligence. This means, in terms of AI usage for corporations, AGI may not even be worth pursuit.
If you are building systems today, it's important to question not the task that needs to be done, but what information and thought is required to complete such a task.