# Agentic AI Widens Enterprise Performance Gap

**Podcast:** The AI Daily Brief (Formerly The AI Breakdown): Artificial Intelligence News and Analysis
**Published:** 2026-08-25

## Transcript

There's always been a gap between an average AI user and the most advanced AI users, but my goodness has that gap grown.
In recently released research, OpenAI showed that the gap between the most advanced users and the average AI user had grown from 2.6x back in January to 8.3x by the end of June.
In other words, the most advanced users of AI were using 8 times as much AI as were their average counterparts.
The reason, of course, is agents.
At the beginning of the year, agentic use cases became viable and significantly upgraded the difficulty, complexity, and importance of the work that AI could take on.
The top users have jumped in headfirst, figuring out how to significantly increase the value they get from their AI usage.
The average users, on the other hand, just haven't.
But as the power users use agents to take on increasingly valuable work, that gap is just poised to grow.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends, quick announcements before we dive in.
First of all, thank you to today's sponsors, KPMG, Blitzy, Harbor, and Hyperagent.
To get an ad-free version of the show, go to patreon.com slash ai daily brief, or you can subscribe on Apple Podcasts.
And to learn more about sponsoring the show, send us a note at sponsors at ai daily brief dot AI.
You can also find information on the ai daily brief dot AI site.
While you're there, you can also find a link to our free webinar and hands-on lab, Agentic Loops for Knowledge Workers.
That is happening on Wednesday.
And even if you can't make it, if you register, we will send you the recording after.
And of course, if you were looking for a little bit more hands-on support, our next executive catch-up and executive agent leadership program is starting in a couple of weeks.
And you can find a link to that program from the top of aidailybrief.ai.
According to roadmap documents viewed by the information, Meta is putting the finishing touches on their consumer agent ahead of release in the coming weeks.
Now, this is something we've been hearing about for a while, but we're getting more details as the product becomes imminent.
Internally, the product is known as Hatch, and sounds like it could be sort of in the GrokBot family of delivering a more streamlined version of an open-claw-style agent experience.
The company is reportedly looking at using Hatch as part of a new AI agent subscription, which could justify a $200 a month price tag for high usage accounts.
Meta also plans to launch a new platform on WhatsApp to allow better integration for third-party agents.
The platform will reportedly allow multiple agents to coordinate with each other using WhatsApp messages, again, which is a mirroring some of the functionality, like I said, of GrokBot.
For what it's worth, this doesn't strike me at all as Meta cribbing off of GrokBot.
I think these are just interaction patterns that we're likely to see more of.
The rollout could begin as soon as this week, as a preview to a smaller group of customers.
Finally, in Meta model news, a larger model known as Watermelon is being prepared for an October launch.
Back in July, Meta's AI CEO Alexander Wang told staff that Watermelon had already caught up with GPT-5.5 on internal benchmarks.
Now, obviously, the frontier has moved forward substantially with the release of GPT-5.6, and will likely move again by the time October arrives.
So we'll see whether Watermelon can actually keep pace or continues to fall in the column of Meta getting closer to the frontier without actually reaching it.
Speaking of GrokBot, one of the barriers for a lot of you guys testing it has been its extreme premium pricing.
In fact, it wasn't just high pricing, it was kind of confusing pricing.
Initially, users weren't sure if they had to subscribe to both Cursor Ultra and Super Grok Heavy, which would be a total of $500 a month to access the service.
But then many, myself included, were able to get it just through Super Grok.
which is itself not a cheap subscription, but it was very clear that this was an intentionally rate-limiting launch to make sure that things didn't go down, with the hopeful anticipation that prices would be reduced later.
As of this week, GrokBot is included in the $60 a month Cursor Pro subscription, as well as the $100 a month SuperGrok subscription.
Speaking of dropping prices, OpenAI is also dropping prices for GPT-5-6-Soul.
Accessing Sol over the API will now cost $4 per million input tokens and $20 per million output tokens, down from $5 and $30 respectively.
Costs for Luna and Terra were already cut late last month.
Many are speculating that this is OpenAI trying to put pressure on Anthropik ahead of their IPO.
My guess is that for whatever ancillary benefit that might have, OpenAI is more likely to just be realizing that they've got a new set of challenges based on what we talk about every week here on this show.
that especially business customers are not just going to use the most expensive state-of-the-art model for everything anymore and have to think in more sophisticated ways about their complete model stack.
To the extent that OpenAI has the compute to deliver their frontier models cheaper, it seems like they've decided it makes sense to do so.
Now, following up on news from yesterday, Business Insider reported that Hugging Face was courting an acquisition at a $13 billion valuation, and according to sources speaking with the information, part of what might justify their high asking price, is that the company is now generating more than $150 million in annualized revenue, which is up 50% from two months ago.
Now, that number may seem low relative to, for example, the coding agent startups, but Hugging Face is a company that has specifically not been focused on generating revenue.
97% of users access the platform entirely for free, including downloading the latest model weights.
The primary revenue drivers for Hugging Face are premium and enterprise-grade accounts, serving inference in partnerships with Hyperscale Clouds.
Essentially, up until now, the profit-seeking segments of the platform have existed to subsidize the free hosting and distribution of open models.
In June, CEO Clem DeLange said that the number of premium accounts had doubled in the first half of the year, and based on that, the platform was approaching profitability.
For acquirers, this is very likely not a strict revenue-multiple type of conversation.
Summing up the logic, Jess Fields writes, Hugging Face should be worth as much as Cursor is, way more than $13 billion, maybe three to four times that.
Considering that Hugging Face is the backbone of the open weights challenging frontier gatekeeping, it occupies a uniquely powerful position in the entire economy.
Now, one of the companies that people are speculating on might be an interesting fit for Hugging Face is, of course, NVIDIA.
If for no other reason than they seem to be in the conversation for every acquisition right now.
In fact, with a flurry of reporting around NVIDIA's dealmaking in recent days, some are wondering just what Jensen is building.
Over the past week, we've heard that NVIDIA signed a deal to license technology and acquire talent from poolside, buy a stake in data labeling company Merkur, and potentially invest in perplexity at a potentially perplexing $30 billion valuation.
That's to say nothing of other equity investments in neoclouds, land and power deals, and data center backstops.
Some have started to conceptualize NVIDIA as the central bank of compute, standing behind the AI economy, as well as setting the price of the key resource.
Martin Peers of the Information compared Jensen's approach to that of John Malone, who built up a giant cable TV empire in the 1980s.
Another rough comparison point might be Google's approach with their holding company Alphabet in the mid-2010s, when Google restructured the company and created their other bets division to house Moonshot investments in Waymo as well as Google Ventures.
The basic idea was to reinvest their massive earnings from internet advertising into the broader tech ecosystem, and of course subsequently a lot of those bets have now paid off.
NVIDIA's approach is obviously different.
Rather than fanning out across next-generation tech, NVIDIA is sticking to the AI ecosystem.
Still, they are building up a formidable portfolio of other bets.
During their last earnings call in March, they reported $42.3 billion invested in private companies.
And the number will certainly be higher when they report later this week.
And before we scream up and down shouting circular dealmaking, one thing that's important to recognize is that part of this is a reflection that NVIDIA is more or less tapped out when it comes to reinvesting in their own business.
Nvidia does not operate their own fabs.
So at this stage, growing chip revenue is limited by constraints across their network of suppliers.
In other words, constraints they can't control.
By turning their investments outwards, Nvidia supports the entire AI economy, and that in turn ensures that their revenues can stay strong for years to come.
At least that appears to be the goal, as they get even more ambitious with their other bets.
Still, chips remains the big game, and over on the other side of the world...
Taiwanese prosecutors have charged nine people for chip smuggling, including one NVIDIA manager.
On Monday, the Taiwanese announced nine indictments in relation to a scheme to smuggle cutting-edge Blackwell 300 systems into China.
In addition to the person who was identified as a manager in NVIDIA's distribution business, two others worked for NVIDIA partner Supermicro.
Earlier this year, a Supermicro co-founder was charged in relation to another smuggling incident.
Both NVIDIA and Supermicro have indicated the problems are contained to a few rogue employees and that they're working with authorities.
To give you a sense of the scale of this incident, the group allegedly ordered 130 supermicro servers containing Blackwell 300s, with prosecutors claiming that 74 were delivered to buyers in China, while another shipment of 56 were stopped by Taiwanese officials.
In total, you're talking about less than 10,000 chips, which is not nothing but certainly not enough to build a frontier training cluster.
Lastly today, an interesting peek under the hood around how much AI one hyperscaler's employees are using.
Business Insider got hold of an internal spreadsheet where Microsoft employees self-report key metrics including salary, bonuses, and AI usage.
Now, this is not a story about token maxing, as BI found no correlation between token burn and financial compensation or promotions at Microsoft.
Both across and within different departments, AI usage was extremely jagged.
In the Azure department, the range of monthly AI spend was $1 to $7,500.
In the cloud and AI division, the top end went all the way up to $15,000.
And in Microsoft Customer and Partner Solutions, there was at least one person who ran up a bill of $28,000.
The median AI usage across all the different departments was a little more clustered together.
Seven of the eight had median AI usage of right around $150 to $500, with CoreAI being the outlier where their median AI usage was $975.
Now, importantly, this is all voluntary self-reporting.
The spreadsheet is maintained to allow staff to voluntarily share compensation figures.
in an attempt to promote pay transparency.
Only a tiny sliver of Microsoft's 223,000 employees overall contribute to this chart, just 600 US employees in fact, only 350 of whom reported their AI usage.
Still, it gives you a sense of the type of magnitude you're seeing among perhaps the more prominent AI users inside a company like Microsoft, which I think is actually the perfect segue into our main episode, which we will begin now.
A new study from KPMG in the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes.
Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs.
These top performers, called AI amplifiers, weren't defined by what they knew alone, but by how they worked with AI.
Learn more about what separates AI amplifiers from everyone else at kpmg.com slash us slash AI amplifiers.
Blitzy deeply understands your code base before it writes code.
Here's the first place that pays off.
Security in the age of AI.
Vulnerabilities don't live in isolation.
They live buried inside millions of lines of interconnected code where patching one thing quietly breaks three others.
That's why surface-level scans fail.
Blitzy starts from its knowledge graph of your entire application, identifies and surfaces CVEs across the full estate, Proactively recommends patches and can execute the PR.
Each fix is grounded in how your systems connect and validate so nothing new breaks.
And the knowledge graph dynamically updates, keeping you ahead of an ever-accelerating threat landscape.
One Blitzy customer resolved 21 active CVEs across six core microservices in four days.
Zero compile errors, every validation scanned clean, months of planned work fixed in less than a week.
Security remediation grounded in real architectural context at the speed of compute.
Harden your codebase at Blitzy.com.
That's B-L-I-T-Z-Y dot com.
If you listen to this show, you likely have a thesis.
Maybe it's enterprise adoption.
Maybe it's compute.
Maybe it's a specific lab.
Harbor Capital's AI Lab ecosystem ETFs let you express it via five actively managed ETFs, each seeking exposure to the ecosystem around one major lab.
Anthropic, OpenAI, DeepMind, Meta, or SpaceX AI.
Your view of the AI race in ETF form.
Harbor Capital Advisors' AI Lab Ecosystem ETF Suite gives investors a way to invest in the AI ecosystem they believe is best positioned for success.
Search Harbor AI Lab Ecosystem's ETFs wherever you invest or follow at Harbor Capital on X to learn more.
Visit harborcapital.com for a prospectus containing investment objectives, risks, fees, expenses, and other important information.
Read and consider it carefully before investing.
Risks include principal loss and artificial intelligence-related risks.
Harbor ETFs are distributed by Foresight Fund Services, LLC.
Harbor is not affiliated with AI Daily Brief, and the funds are not affiliated with, sponsored by, or endorsed by any AI lab.
This is a paid advertisement and not personalized investment advice.
Investing involves risk, including possible loss of principle.
This episode of the AI Daily Brief is brought to you by HyperAgent, where you run fleets of agents your team can manage together.
New users get $1,000 in inference.
Forget local agents and chat workflows waiting on your laptop to be prompted.
HyperAgent deploys always-on agents in the cloud, doing real work across the tools your team already uses.
Marketing's agent turns competitor moves into landing pages.
Sales's agent enriches leads, drafts emails, and updates the CRM.
Ops agent chases the paperwork and tracks the budget.
Every agent has access to shared context and follows your rules about scope and approvals.
It's time you add agents that feel like teammates.
Hire yours at HyperAgent, built by the team at Airtable.
Claim your $1,000 in inference at hyperagent.com slash AI Daily Brief.
Welcome back to the AI Daily Brief.
One of the best things that's happening right now, when it comes to AI narratives at least, is that we're starting to see a shift away from the accepted without question kind of premise that AI is obviously going to be job destroying.
Now, regular listeners will know my position on this is pretty clear.
I am not Pollyannish at all about the potential scale of challenge when it comes to the transition between two totally different work paradigms.
You are inevitably going to see certain types of roles that this new category of technology will obviate the need for.
That will cause real personal disruption.
that society would do well to be ready to support the people affected.
The idea, however, that there was going to be some radical and rapid jobs apocalypse was never accurate.
On this point, Sam Altman has increasingly gone out of his way to not just have changed his position, but to explain that he believes that he was wrong and to try to explain why he thinks he was wrong.
In a recent podcast interview, he said, I thought when we got to GPT-4, which was back in 2023, that very quickly after that, there was going to be much more disruption.
Software business is up for grabs right away than it turned out to be.
I think I was wrong about a few things, but one in terms of the speed.
The economy just has so much inertia.
People keep doing the same things, buying from the same company, wanting to use their tools the same way.
I think this is actually a positive in many ways, and it's going to make this big transition go smoother and slower.
I'm grateful for it, but it means we've all been too ambitious on timelines.
Even with this incredible technology, society and the economy will adapt more slowly.
In other words, as I like to think about it, who needs a pause AI movement when you've got corporations?
What's interesting, though, is that Sam actually went farther and recognized that while, yes, institutional inertia is perhaps the biggest driver of the slowdown in the rollout, that that inertia happens on an individual level, even with really advanced users as well.
In that same interview, he said, The thing that feels most psychologically inconsistent about myself is that I have for 20 years been using computers the same way.
I now have a magic thing called codex.
So do you, so does everybody.
That means I should completely be using my computer in a different way.
I should not be clicking around, pasting from one messaging app to another.
I should not be scrolling mindlessly through my emails trying to figure out which one is the least painful to open.
I should not be keeping a to-do list and doing these rote computer tasks the same way I have for so long.
And yet there's something in my mind that is encoded, that doing this kind of stuff is what it means to work and be productive.
If you asked me, I would never say I like doing it that way.
In fact, I'd say the opposite, and I'd mean it.
But by revealed preference, I have a better way to do it now, and I still do it the old way.
It makes no sense other than I must secretly like it or feel good about it.
I don't think it's even some secret satisfaction, though.
I just think we all get stuck in the patterns of how we've always done things.
We build intellectual mind-muscle memory.
And the thought of undoing that to go build a new type of mind muscle that is much less comfortable often appears exhausting when we know exactly how long it will take us to do it the old way and we could just get that done right now and move on to whatever it is we actually want to be doing.
Which is why it's so important to spend time looking at how people who have broken out of their mind muscle memory are doing things.
And interestingly, a couple of weeks ago, OpenAI published some research about exactly that.
Now, I don't know why they weren't screaming about this article from the rafters, but if I miss something that's putting numbers around how frontier firms are doing things differently than others, you know that thing was not promoted nearly widely enough.
In any case, I only noticed the data when A16Z reposted it as part of their charts of the week.
The chart that grabbed mine and many others' attention was this one showing Codex User Growth by Enterprise Job Title.
It was indexed back to the beginning of February of this year, and in that time, While every role has gone up, it is the non-technical roles that have grown the fastest.
Now, of course, part of this is because they had a lower starting point.
But to give you some examples, while use among engineering and technical practitioners is up 5x in that time, use in finance and accounting is up 20x.
In marketing and communications, it's 26x.
In people and recruiting and separately sales and account management, it's up 41x.
And in legal, codex usage is 108x from where it was back in February.
Joking, but not really joking about the lawyer side of this.
Spellbook Scott Stevenson quoted Jack Newton saying, LLMs are for lawyers what spreadsheets were for accountants.
But that was hardly the only data that was available here.
The big through-line signal in this piece, which was called enterprise signals, what frontier firms are doing differently, was two parts.
First, more work is being delegated to agents.
And second, because of that, there is a compounding effect where the firms that are farthest along are getting farther away from the firms that are behind.
In other words, agentic use compounds their lead.
And importantly, this is not a model question.
It's about, as OpenAI puts it, how they put those models to work.
Moving from assistants to delegation, giving agents the context and tools to complete complex tasks, and accelerating agentic AI beyond software development.
So let's talk about a few of the most interesting numbers.
A year ago at this time, in August of 25, when GPT-5 was announced, The balance between ChatGPT usage and agentic usage, as measured by the output tokens produced by the enterprise interacting with the ChatGPT ecosystem in any way, was basically 100% ChatGPT tokens and not agentic tokens.
Between October and February, the first glimpses of the actual agentic era started to be seen.
Codex becomes generally available in October, and a low single-digit percentage of enterprise output tokens are now in that agentic category.
In December, GPT 5.2 comes out.
and we see a bit of a bump that extends up into the beginning of February.
In February, when the Codex app launches for macOS, the percentage balance between ChatGPT and agentic usage, again as measured by enterprise output tokens, was 87% ChatGPT to 13% agentic.
But then from there, the agentic use cases just take off.
We get to March, Codex is released for Windows and GPT 5.4 comes out.
And we're now at 73% ChatGPT, 27% agentic.
Towards the end of April, just a month later, GPT-55 comes out and we get the flippening, where all of a sudden agentic output tokens are representing 53%.
And by June, when this data set ends, we're down to 36%, while agentic tokens were up to 64%.
Now keep in mind, this does not mean that all of a sudden 64% of the times that someone sits down to use an OpenAI product at work, they're doing something with agents.
Instead, what this means is that if you use the amount of output tokens that the aggregate set of prompts lead to, in other words, if you use output tokens as a proxy for the amount or volume of work being done, the preponderance of it, almost two-thirds by the time these statistics were captured, is now being agentic work.
So point one was that agentic use is up, and point two was that the gap between the leading AI users, i.e.
the ones using agents the most and the best, and the general enterprise users was getting wider.
OpenAI defines frontier firms as those in the top 10% of usage in a month, as measured by output tokens per active user, with the average firm to be between the 45th and 55th percentile.
And providing the best numerical advice, you can see that going back to about April of 2025, throughout the year of 2025, the gap in output tokens per active user between the typical firm and the frontier firm was only about 2x.
as in the average user at a frontier firm used about twice as many output tokens as an average user at an average firm.
The gap started to widen around October, and in January stood at 2.6x.
The gap now has absolutely exploded, with the distance between the typical firm and the frontier firm now at 8.3x.
Overall, while the average firm is using about twice as many tokens as they did a year and a half ago, frontier firms are using 17 times as many tokens.
as they were a year and a half ago.
Now, part of this is, of course, that they're just using agents more, but part of it is that they're also better at using agents.
To measure this, OpenAI looked at the number of weekly active users who use either plugins or skills.
Plugins are, of course, capability sets that can connect to other applications or data, whereas skills are reusable instructions that can help with common workflows.
At typical firms, about 9% of weekly active users are using plugins, and only about 3% are using skills.
At frontier firms, which is again the top 10% of enterprises, 19% are using skills and 21% are using plugins.
Which is not to say that those frontier firms have topped out in terms of their usage.
By way of comparison, at OpenAI right now, 93% of their employees are using skills and 95% are using plugins.
But still, coming back to the gap between employees at frontier and typical firms, you're talking about two and a third times as many people using plugins and more than six times as many people using skills.
And in terms of what work they're deploying this towards, as we saw from that chart that kicked off this show, the fastest growth in these agentic use cases is coming from knowledge workers that are outside software and engineering functions.
OpenAI diagnoses it like this.
Software, they said, moved first for a reason.
Codebases give agents clear context, tests make outputs easier to verify, and progress in coding helps accelerate AI research and development.
In contrast, they say, Progress in general knowledge work has been slower because many tasks provide limited context, can be difficult to specify, and lack clear criteria for verifying the result.
But continued scaling, advances in reinforcement learning, and targeted efforts to improve performance on evaluations such as GDPVal are bringing more real-world tasks, tools, and work environments within the reach of frontier models.
As a result, agentic AI has increasingly found product market fit with general knowledge workers since the beginning of the year.
Look, it is great that OpenAI is working hard to have their models and harnesses work better for knowledge work.
But stuff that OpenAI has done is not the reason that agentic use has grown among these non-software engineering knowledge workers.
The reason that agentic use has grown among general knowledge workers is that we've started to figure out the patterns that actually allow agents to thrive in our own contexts.
So what are those patterns?
From this, I'm combining the research from OpenAI about the types of use cases both in chat and agentic across departments, plus our own experience at both AIDB and at Superintelligent, to provide a little bit of a picture of what that more advanced agentic use actually looks like in practice.
You can see in OpenAI's chart that there's a massive shift in the type of work when you move between chat and agentic.
And to be clear, this doesn't mean that the type of work being done with chat is not valuable.
Support for writing and communications, knowledge retrieval and search, those things do bring a lot of value.
But agentic takes a lot of those individual workflows and instead moves them into systems-level work.
that can impact more than just the individual.
You can almost think about a use case ladder.
At the base is generation.
Think drafting an email, a report, an Excel formula, things like that.
On the second rung is synthesis, being able to take disparate data sources and produce something more complete that is informed by them.
Moving farther up the ladder, we have execution, where the AI is actually being tasked with interacting with existing systems and doing things within them, which of course is closely tied to the next level of maintenance.
where agents are tasked not just with executing something specific, but maintaining a system over time.
So for example, if you're looking in legal, the context that people are drawing on is of course things like contracts, policy documents, precedent, as well as any relevant counterparty history.
The types of agentic work patterns you're going to see are around things like comparing terms, flagging deviations, drafting red lines, recording decisions, monitoring commitments.
Humans will continue to negotiate the material terms, to set risk tolerance, to approve exceptions in final language.
So basically you have a division of labor where the agent is handling coverage and coordination, while people still own risk judgment and accountability.
If you go back and look at the difference in the categories of work in the legal field that are being done via chat or agentic, writing makes up a full 57% of legal work in chat, followed closely by knowledge retrieval at 20.5%.
System operation barely registers at 0.2%.
Now you move over into agentic and you've got writing down to 16.2%, knowledge retrieval down to 8.3%, and a whole bunch of new categories coming online in a huge way.
Classification and extraction jumps to 4.6%.
System operations jumps to 17.7%.
Workflow automation jumps to 7.7%.
And coding, actually building applications, even though these aren't the software engineers, jumps to 32.9%.
And you see this pattern in basically every other department as well.
Writing and knowledge retrieval with a side of education and guidance remain the preponderance of chat tokens, while Agentec gets into deeper systems integration work.
Now, one thing that I think is going to supercharge this to the next level is the emergence of multiplayer and team AI as opposed to just individual AI.
I think that even though you are seeing these frontier users move more into these higher order tiers of execution and maintenance of systems, most of these agents are still operating within individual silos.
And my belief is that where a lot of the next generation of gains are going to come from is actually at the intersection of different teams.
Ultimately, none of this is all that surprising.
It has been clear for a while that 2026 was the year that agents became real.
But boy, the numbers do not lie.
And while even the frontier firms are still just barely beginning to figure it out, the fact that they seem to be racing ahead and putting more and more distance between themselves and the average firms should be a wake-up call for those who aren't deploying agentic uses at scale yet.
This is probably where I should insert a shill for our training programs over at Super Intelligent, but if you are a regular listener, you will already know that those are there and available for you.
In any case, this is great stuff from OpenAI.
Please keep doing this and please promote it more heavily next time.
And of course, for you guys, appreciate you listening as always.
Until next time, peace.