Business The Tokenpocalypse Is Here: Companies Are Scrambling To Stop Spending So Much on AI - Leaked audio from Accenture says a big source of AI token ‘chewing’ is people just converting PDFs to presentation slides.

Photo by Sebastian Herrmann on Unsplash, collage by 404 Media with logos from Microsoft and Anthropic.

Consulting giant Accenture is trying to figure out how to stop non-technical workers from blowing through companies’ AI token budget on trivial tasks like converting PDFs to presentation slides, according to leaked audio obtained by 404 Media. Across the industry Accenture is seeing “soaring token spend,” according to the audio.

The news highlights a major shift in the tech industry and other companies that use AI: the wave of uninhibited AI growth is over. Some AI providers like GitHub are now charging customers per token rather than a flat subscription fee, leading some companies to burn through their tokens. Uber recently capped employees’ use of AI tools like Claude Code and Cursor; that came after Uber told employees to use AI as much as possible and Uber’s CTO said the company had blown its entire AI budget in four months. And Accenture itself reportedly started requiring senior staff to start using AI or risk losing out on promotions.

It also undercuts the narrative that superpowered engineers generating mountains of code are behind the AI boom. In many cases it is non-technical staff burning through tokens for non-specialized tasks.

“We’re seeing from some of the data internally at least that it’s actually not our engineers that are driving the token consumption. It’s a lot of the non-engineers that are doing some of those behaviors [...] you were talking about,” Justice Kwak, Accenture’s agentic AI strategy lead, said in a recent internal meeting, according to the audio obtained by 404 Media.

At one point in the meeting, Kwak and Eduardo Salamanca de Diego, senior manager of product management at the company’s Center for Advanced AI, start presenting about what is described as “token ops.”

Kwak says he knows people aren’t using slides these days, but he has some. As he appears to be preparing to present, Stuart Henderson, Accenture’s client group lead, interrupts. He jokes he hopes Kwak didn’t just convert a PDF into images and then into markdown files. “I’m learning that’s one of the big token chewers,” Henderson says. “Turning PDFs into markdown: is that right?”

That’s when Kwak says that’s what Accenture’s own data shows.

“What we’re seeing right now is just rapid escalation in AI token spend,” he says “As companies start to scale AI, moving from like simple chatbots into use cases that feature agentic workflows and automation and then enterprise-wide deployment of some of these tools like Copilot, Claude Code, and Codex, we’re hitting this inflection point where AI is becoming material to the cost structure; spend is becoming very unpredictable; and leadership, especially at the CFO, COO, and CIO level, are still asking the question of whether they’re getting value from what we’re spending on in the context of AI.”

“It’s really not a niche problem. It is a problem that every enterprise will face if they are bullish on AI, if they haven’t already,” he adds. The amount of token spending is increasing “exponentially, as more and more people are starting to use AI.”

Kwak says after Accenture tried to get enterprises to adopt AI as quickly as possible, AI has reached scale in most areas in both Accenture and its clients. But with that scale is a new opportunity for Accenture regarding its clients: “to really think about token economics.” The bill of the overall AI spend is visible, Kwok explains, but attributing that AI spend at the token level to the value outcomes on the projects where AI is being used is not visible.

Finally, the “controls are just arriving too late.” Those are things that might stop someone spending a bunch of money on tokens, like budgeting or different tiers.

Following the Financial Times’ reporting of Accenture’s policy to force AI adoption or risk missing promotions, an Accenture spokesperson told CNBC, “Our strategy is to be the reinvention partner of choice for our clients and to be the most client-focused, AI-enabled, great place to work. That requires the adoption of the latest tools and technologies to serve our clients most effectively.”

Kwak says Accenture plans to formally launch a product called “Token IQ” soon. Accenture did not respond to a request for comment.

As 404 Media has reported, some startups have bragged about how much they’ve spent on AI instead of human workers. Walmart also capped its staff’s use of AI tools following high demand.

Article Link

Archive

 

[Gandhi's Left Hook]

Society has uncovered new and exciting ways to become even more brainless.

Students graduating High School without learning how to read, what they can read is handicapped by whole language, they go to college and can't do fifth grade math and now they can delegate every task on a computer to an AI.

Learning is obsolete. Be proud.

They actually don't have a high enough grasp of language to prompt AI properly.

 

[Disdainer of the Plebs]

So this is the bubble pops. With people using AI for shit they could do otherwise.

This is actually happening and it’s going to cause worldwide shockwaves. Companies all bought in on the hype, but now the major providers need to become profitable for their IPOs so the “free/cheap access” portion of the Silicon Valley business cycle is ending. OpenAI and Anthropic are now moving from a free quota towards pay-per-token, and the price of that for a large enterprise is quite shocking.

Here’s the economics of it:
A solid senior level software developer typically requires somewhere between $500 and $1000 in tokens per month to really realize value, and that’s if they’re using it very responsibly and not tokenmaxxing. Tokenmaxxers typically spend more like $3-5k, but typically produce worse code in higher volume. That adds up to an extra 12-60k per year per engineer. If you have 100 engineers, that’s 6 million bucks on the high end.

To put that $6 million in perspective, for $6million you can host a pretty sizable company’s entire IT operation in AWS dev to prod. Tokenmaxxing is not economically viable and the sticker shock companies are seeing will have unavoidable consequences.

lol no it’s in my goals to use AI more. I literally have to. Now, what’s the most computationally intensive task I can do as a non coding middle ranking wagie?

Some gamesmanship is necessary here, so find out what KPIs you’re judged on.

If it’s a combination of requests sent and tokens spent and total money spent, prompt the AI to be as long winded, detailed and verbose as possible when it responds, use the most premium model you can, crank reasoning levels up to max and just replace web search with it. Make it review and revise any written communication you write. Give it any trivial search task you can imagine and you’ll be burning money like a pro in no time.

An advanced technique is to tell it to spin up multiple subagents and have them discuss your task amongst themselves. This burns a fuckload of tokens, and is the only way you can get the fucking thing to produce code that can even dream of getting near production in a responsible shop. Also, if you have to write extensive documentation, make AI do that. Burns tokens by the bushel fucking basket but is worth every penny, because your documentation actually exists and is in legible English.

 

[Dante Alighieri]

lol no it’s in my goals to use AI more. I literally have to. Now, what’s the most computationally intensive task I can do as a non coding middle ranking wagie?

https://github.com/Lomnus-ai/TokenBurner

There are tons of projects on GitHub that amount to "please burn as many tokens as possible".

 

[Point and Laugh]

I use AI for my more tedious admin tasks, give it a few examples of when I did it manually and it can copy my writing style mostly. Is that not what AI was explicity designed to do?

It’s literally been “sold” on automating tedious bullshit therefore allowing people to focus on productive work. It’s hilarious that’s now a problem.

Also, I hope Accenture implodes under the weight of its jeet slaves using AI tokens to change their voices and create poorly made posts about aryan supremacy.

 

[Diana Moon Glampers]

It also undercuts the narrative that superpowered engineers generating mountains of code are behind the AI boom. In many cases it is non-technical staff burning through tokens for non-specialized tasks.

Bullshit. The tokens wasted by the non-technical people are a rounding error compared to the nutty agentic loops run in the engineering department and product department. Absolute nonsense.

They just know that all the big AI valuations cannot hold if it turns out the engineering use case costs more than it's worth. You can tell businesses to stop finance and marketing and sales from using it if you want to, but telling the engineers to curtail use would be the pop in the AI bubble.

 

[Betonhaus]

They literally made it so that employees needed to run this in order to get promoted instead of fired then have the audacity to be shocked when they spend tokens like water. https://github.com/dtnewman/burn-baby-burn

Like, did they not take a second to understand the business model for AI companies and recognize that this was inevitable? For fucks sake understanding the business model of the very things you pivot the core functions of your company to is like chapter 1 of The Art Of Business

 

[Hitman One]

Quick, we need more data centres. Just put the fucking water pipe into the ocean, the fish will help the machine think.

 

[Snekposter]

Wait, can I cause my company pain and misery by using AI more? Nobody told me that.
How can I burn maximal tokens with plausible deniability?

Why deny it? Just get AI to spin up a message in peak manager speak about how beneficial your AI use is to your workflow and the company. In fact, you should have it produce multiple versions for you so you can pick the most optimal one, and then have it write a follow-up afterwards about how it helped you avoid a tedious revision process.

 

[Not A Bunny]

Why would you invest in AI and then be mad that people are using them to make slide decks?

Slide deck creation is one of the most brainless and least worthwhile task a person can do in a day.

It's all some Government organizations do. Between me and a number of my friends, I've confirmed that Government is a whole metric ton of brain-rot that is incapable of surviving without everyone being in every meeting, and every document being pored over in team meetings because you have no autonomy anymore.

Anyone with half a brain to see all of this going down is suffering from learned helplessness.

I am determined like hell to get out of this fucking hellscape.

EDIT: Oh right, I had a point to make, and that's this: since everyone has no autonomy anyway, what's the difference between ceding it to the group, versus ceding it to AI?

 

[TJT]

Here’s the economics of it:
A solid senior level software developer typically requires somewhere between $500 and $1000 in tokens per month to really realize value, and that’s if they’re using it very responsibly and not tokenmaxxing. Tokenmaxxers typically spend more like $3-5k, but typically produce worse code in higher volume. That adds up to an extra 12-60k per year per engineer. If you have 100 engineers, that’s 6 million bucks on the high end.

WTF are you smoking? Can I get some?

I am a staff engineer, org wide we have a $20/month limit to use Cursor. Multi billion dollar company.

 

[Betonhaus]

So what will break first? Companies all reporting to shareholders that their entire profit has been eaten up by AI or AI companies reporting to investors that their customers have been cancelling in droves and they will never make back all the money they spent?

 

[Disdainer of the Plebs]

WTF are you smoking? Can I get some?

I am a staff engineer, org wide we have a $20/month limit to use Cursor. Multi billion dollar company.

Fuck, at that point it’s not even worth it. You won’t even get one query out of Opus for that. Or are you still on the legacy plan where you get the 500 “request points”? I currently run Claude code, and we pay by token for that that. The pilot came with unrestricted use, and the numbers I put in my financial analysis are roughly what we saw. They imposed limits this last week because they realized we’re gonna burn a couple million. I’ve blown through about $500 in tokens over the last month. I run opus to hunt down production incident root cause though, and I spent about $300 of the 500 on that. Worth every penny to end a managed incident though.

I used to be on the $20 cursor plan with $50 on demand usage but I ran out of capacity in 2 weeks and that was when I was being super conservative and putting everything into composer, haiku and gpt mini. Those models can’t handle a lot of the work we throw at them. Most power users at my company just signed up for all the tools we can and just swapped tools when we ran out of requests.

 

[TJT]

Fuck, at that point it’s not even worth it. You won’t even get one query out of Opus for that. Or are you still on the legacy plan where you get the 500 “request points”? My company runs Claude code, and we pay by token. They only imposed limits this last week because they realized we’re gonna burn a couple million but I’ve blown through about $500 in tokens over the last month. I run opus to hunt down production issues though, and I spent about $300 of the 500 on that. Worth every penny to end a managed incident though.

I used to be on the $20 cursor plan with $50 on demand usage but I ran out of capacity in 2 weeks and that was when I was being super conservative and putting everything into composer, haiku and gpt mini. Those models can’t handle a lot of the work we throw at them. Most power users at my company just signed up for all the tools we can and just swapped tools when we ran out of requests.

I am not some genius but I have been heads down programming for over 20 years. I work in data infrastructure right now and not app development. Strictly backend engineering.

I generally ration myself well enough, since the IT guys here are tards I supplument Cursor with my own subscription sometimes when I happen to run out. I am quite efficient at this job and I know exactly what I am trying to do, why I am doing it this particular way, and how to do so. A new feature for me in this context has generally cost me $1-$3 in Cursor use.

IMO anyone generating thousands of lines of code with AI is usually full of shit and doesn't understand what they are doing. Nor do they work in anything resembling normal development as one task you might have to develop X feature will almost never be thousands of lines of code.

If I asked I could probably get more Cursor spend. But I don't need it that much at the moment.

 

[Fartificial Insmelligence]

Fuck, at that point it’s not even worth it. You won’t even get one query out of Opus for that. Or are you still on the legacy plan where you get the 500 “request points”? I currently run Claude code, and we pay by token for that that. The pilot came with unrestricted use, and the numbers I put in my financial analysis are roughly what we saw. They imposed limits this last week because they realized we’re gonna burn a couple million. I’ve blown through about $500 in tokens over the last month. I run opus to hunt down production incident root cause though, and I spent about $300 of the 500 on that. Worth every penny to end a managed incident though.

I used to be on the $20 cursor plan with $50 on demand usage but I ran out of capacity in 2 weeks and that was when I was being super conservative and putting everything into composer, haiku and gpt mini. Those models can’t handle a lot of the work we throw at them. Most power users at my company just signed up for all the tools we can and just swapped tools when we ran out of requests.

Imagine going back in time and saying these words to some ancient Egyptian Pharaoh, they’d either think you were a God or possessed by demons…

 

[Disdainer of the Plebs]

Imagine going back in time and saying these words to some ancient Egyptian Pharaoh, they’d either think you were a God or possessed by demons…

Never mind 5000 years, just go back 5.

 

[Tanner Glass]

So what will break first? Companies all reporting to shareholders that their entire profit has been eaten up by AI or AI companies reporting to investors that their customers have been cancelling in droves and they will never make back all the money they spent?

They're going to stall this out until Anthropic/OpenAI/SpaceX is able to cash in on their IPOs and then leave investors holding the bag like they always do.

They're going to try and get out ASAP so I expect by the end of this year all of those IPOs will be completed and then the whole collapse will be allowed to begin.

 

[Disdainer of the Plebs]

I generally ration myself well enough, since the IT guys here are tards I supplument Cursor with my own subscription sometimes when I happen to run out. I am quite efficient at this job and I know exactly what I am trying to do, why I am doing it this particular way, and how to do so. A new feature for me in this context has generally cost me $1-$3 in Cursor use.

That’s all well and good, and I can do the same, but if you throw a much bigger problem than a feature at it, it gets more expensive very quickly.

For example: I have Claude hooked into my public cloud, log and monitoring systems to the pin t where I can say “ Claude we have a production incident, here’s the ticket, go find out what’s happening.

It will blow 10 million tokens, but 250 bucks is a cheap fix compared to wrangling 50+ people on a P1 incident call, not to mention the estimated cost to business for the outage is measured in millions by the powers that be.