Introduction
2023 is often called the "Year One of ChatGPT," where ChatGPT became a topic of discussion not only in the tech industry but across all sectors, and this momentum shows no signs of slowing down in 2024 and beyond 2025.
During this unprecedented ChatGPT boom, companies began initiatives to use ChatGPT internally, making headlines daily.
You've probably seen many news articles like "Company X deploys ChatGPT to approximately 10,000 employees, aiming to improve productivity across all staff."
Not wanting to fall behind early adopters, many companies rushed forward thinking "We need to do this quickly too!" and similarly proceeded with initiatives to distribute ChatGPT internally.
However, data from companies that led early adoption is now emerging, and when we look at actual usage, despite all the attention ChatGPT has received since 2023 and the media frenzy it has caused, overall usage remains at around 10% or less.
This "around 10%" result represents a significant gap from initial expectations, with many negative comments like "We thought more people would use it, but it's being used much less than expected..."
In this article, I'd like to examine "Why isn't ChatGPT being used much internally even though it received so much attention?" - exploring the root causes of this gap and the challenges we should truly be addressing.
Note: While "LLM" would be the technically correct term to use, I'll continue using "ChatGPT" for easier understanding by a broader audience.
Why Isn't ChatGPT Being Used Much Internally Even After Distribution?
So why is ChatGPT being used less internally than initially expected?
In many cases, this expectation gap is attributed to two main factors:
- ChatGPT isn't as useful as expected (ChatGPT is at fault)
- There are problems with employee literacy (Employees are at fault)
In other words, many cases blame either ChatGPT's performance issues (it's still developing and unusable due to hallucinations, etc.) or employees' inability to use it properly (education is needed).
While these factors do play a role, to get straight to the point: ChatGPT will continue to improve but is already sufficiently capable, and the problem isn't with employee literacy either.
Particularly in cases where employee issues are blamed, we see executives loudly proclaiming "Why aren't they using it more! They should be using it more in their daily work!" and companies rushing to implement ChatGPT training for employees.
However, from the employees' perspective, they're honestly just confused by being told to "use it more."
In many surveys, when asked "Why don't you use ChatGPT?", the most common response is "I don't know where to use it."
However, this should be understood not as "I can't use it because I don't know how" but rather "There aren't many places where I can use it."
I often said "It's fine to distribute ChatGPT internally for now, but most people probably won't use it much," and the 10% figure actually reflects reality. The real problem isn't with ChatGPT or employee literacy, but with the expectations themselves.
In other words, the problem lies in the low resolution of providers' expectations that "If we distribute ChatGPT, it will be used more internally."
So why isn't ChatGPT used much internally just by distributing it? This becomes clear when we consider the mapping between what ChatGPT can do and internal business operations.
What ChatGPT Can Do
In most cases, companies have simply distributed it (made it available to employees), so let's consider vanilla ChatGPT as our premise.
What vanilla ChatGPT can do broadly falls into two categories:
- Tasks that can be solved with general knowledge
- Tasks that can be solved by searching for the latest information on the web
First, while GPT has a data cutoff at a specific point in time, all tasks that can be solved with general knowledge up to that point fall under category 1.
This includes commonly mentioned summarization, idea generation, and draft creation. Since it also has scientific and historical knowledge, all tasks based on publicly known information can be solved. Programming code generation and environment setup also fall into this category with impressive accuracy, as they're based on public information.
However, since it only has data up to a specific point in time, some tasks require access to the latest information, which can be solved by category 2.
For example, when you need the latest news, paper information collection, constantly updated technical information, or market trend research, these can be solved through web search in category 2.
As some may have already noticed, the tasks vanilla ChatGPT can solve are essentially those based on public information. It has mastered humanity's collective knowledge up to a certain point and can access the latest internet information, so it can basically solve all tasks within that scope.
So why doesn't internal usage progress as expected with just these features?
Mapping ChatGPT Capabilities to Business Operations
Let's break down the internal business operations that would be ChatGPT's application targets.
Building on our context so far, internal business operations can be broadly divided into three categories:
- Tasks that can be solved with general knowledge
- Tasks that can be solved by searching for the latest information on the web
- Tasks that require internal company information
While tasks in categories 1 and 2 can be solved with vanilla ChatGPT, tasks in category 3 that require internal information cannot be solved by vanilla ChatGPT.
The significant variation in ChatGPT usage within companies isn't due to literacy issues but simply occupational differences - the more tasks in categories 1 and 2 in your work, the greater your potential for ChatGPT utilization.
Areas with particularly high utilization potential include planning/marketing, researchers, and IT engineers.
In planning/marketing, ChatGPT can be utilized for various tasks within categories 1 and 2, including idea generation, brainstorming with virtual personas, copy creation, and trend research.
Researchers benefit significantly from productivity gains through category 2's latest paper and data research/summarization, and it's highly valuable for output scenarios like writing papers and reports. Translation and proofreading are particularly valuable.
Personally, I think IT engineers see the most dramatic productivity gains - having it write base code, researching technical areas you're less familiar with, and various environment setups are extremely helpful.
Questions like "I know how to do this in AWS, but how do I do it in Azure?" or "How do I write this process in this language?" - tasks that previously required half a day to several days of searching through various sources can now be solved in minutes. This past year, I personally felt tremendous productivity gains in this area.
What's common among these ChatGPT-friendly domains is that they involve "tasks solvable with public information."
However, in most companies, these job categories don't represent the majority. For the majority in other job categories, most of their work falls into category 3.
This is because most internal business operations are based on internal rules and documentation, making tasks that can be solved without internal information extremely limited.
This tendency becomes more pronounced in larger organizations. As the number of people involved increases, the organization cannot function orderly without certain rules.
In other words, tasks solvable within categories 1 and 2 represent only about 10% overall, while the majority are category 3 tasks requiring internal information, which is why vanilla ChatGPT usage remains around 10%.
Let's personify this to make it clearer.
What vanilla ChatGPT represents is like hiring an incredibly talented person, giving them internet access, but completely isolating them from all internal information.
This person knows everything about general knowledge, can use any language, and can write programming code. They're probably unmatched in the company when it comes to web searches. You can ask this person anything through chat.
However, this person knows nothing about internal information and cannot access it. They don't know the internal organization structure, who works in the company, standard business processes, approval processes, what systems are used, salary regulations, work regulations, or which companies are clients - they cannot know any internal information.
So while this person is incredibly talented, when you ask "What do I need to do first to proceed with this task?" "Where can I find this information in our internal documents?" "Which positions from which departments should I invite to this meeting?" "What approach should I take in the next meeting with this client?", they can only respond with "I'm sorry, I don't know."
It's the same as hiring a Harvard graduate with an MBA and impeccable business credentials - they won't perform brilliantly from day one. Because without the context of internal information, no matter how talented, they first need to catch up.
For people whose work mostly requires the context of internal information, distributing vanilla ChatGPT naturally means there's little they can use it for.
I use ChatGPT daily myself, and while I can't work without it for technical paper research and programming tasks, I hardly ever use it for summarization, idea generation, or document drafting.
As long as ChatGPT cannot access internal information, this situation won't change regardless of how much ChatGPT's performance improves. While educating users on ChatGPT and teaching prompt engineering is certainly good, this is like learning how to better communicate with that talented person isolated from internal information - it's not addressing the core issue but rather trying to squeeze more from an already wrung-out cloth.
What We Should Really Work On to Bridge the Gap
So what should we really work on to bridge the gap? It's not complaining about ChatGPT's performance or lamenting users' IT literacy. As should be obvious from our discussion, it's giving ChatGPT the context of internal information.
Using our earlier analogy, while there's limited work you'd want to delegate to an incredibly talented person who has zero access to internal information, what if this person was well-versed in internal information?
When you ask "How do I use this internal system again?" "I can't find the internal regulations file..." "This department seems new, what's their mission?" "What approach should I take in the next negotiation with this client?", unlike the isolated state, they would now give you immediate answers.
Additionally, they would provide information you didn't even know existed, like "Regarding that department's policy, this document was issued internally recently. Here's the link." or "Last week, Manager ●● had a meeting with △△ Company, and here's the meeting minutes."
To put it simply, this means "Let's properly implement RAG," but by enabling ChatGPT to respond and handle tasks based on internal information, it can now cover category 3 areas that vanilla ChatGPT couldn't handle, dramatically increasing the number of users.
An incredibly talented person who probably knows more about the company than anyone else will support you 24/7/365, never complaining even if you ask questions somewhat roughly, always ready to help you.
In other words, the strategy to bridge the gap in internal usage becomes: "How much can we improve the level of RAG that provides internal information context?"
More Specific Keywords
For those who find this still too abstract, in more specific terms: the first thing to work on is giving ChatGPT internal information and working towards zero instances of "asking people" or "searching for information" within the company.
The value-add of work lies in transforming input information into valuable output, but there's no value in spending time understanding already-known internal information, so this time should be minimized to zero.
Moreover, as labor shortages are increasingly discussed, interpersonal communication time should be used for idea-generating discussions and decision-making, so time spent just being taught internal information that could be looked up should also be minimized to zero.
In a society where job changes are becoming normal and mobility continues to increase, this will become essential as internal infrastructure.
For people accustomed to an environment where ChatGPT is well-versed in internal information and basic internal matters can be understood by asking ChatGPT, having to "ask people for everything" would be incredibly cumbersome, making it difficult to recruit talented digital natives.
First, enable ChatGPT to have internal information as context, then work towards zero instances of employees "asking" and "searching." Then, in the next phase, move into the "delegating tasks" phase. Since it would obviously be wasteful to only have ChatGPT handle Q&A when it's well-versed in internal information, gradually delegate actual tasks as well.
In any case, "ChatGPT being able to access internal information" becomes the prerequisite.
Summary
How was this article?
The essence of the gap between expectations and reality - that ChatGPT isn't being used as much as expected - isn't due to ChatGPT's performance deficiencies or employees' literacy issues, but because ChatGPT knows nothing about internal information and therefore has limited applications.
Does this mean distributing vanilla ChatGPT internally is meaningless? Absolutely not. As mentioned, people in areas with high utilization potential find it extremely valuable, so this itself is a must. Rather, they couldn't work without it.
However, there's significant occupational bias, with overall usage at around 10%, which isn't a gap from expectations but rather meets expectations in a sense.
Of course, ChatGPT's performance will continue to improve, and employee education should proceed in parallel, but the highest priority isn't there - it's establishing an environment where ChatGPT can function with internal information included.
While 2023 was about "distributing ChatGPT internally and trying it out," as some companies have already begun working on, shouldn't 2024 and beyond be about seriously working on "creating internal experts who support all employees" in the realm of business efficiency and automation?
Naturally, this isn't a zero-or-one issue, so work needs to begin early and continuously level up. And since internal information and its understanding are key, this is an area that's difficult to completely outsource to external vendors, including decisions about information necessity and evaluation.
I hope this helps in properly understanding what "ChatGPT usage remaining at around 10%" means and connecting it to the next strategic initiatives. We help organizations go beyond tool rollouts and make AI adoption actually stick.
