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

 

[Otterly]

Stop giving them tokens, you stupid fucks. Problem solved. Now pay me my $10k consultation fee.

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?

 

[Meriasek]

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?

Let it format Word documents and transform them to PDF and back. Have it rewrite texts endlessly. Let it write your emails, and never accept the first draft it gives you.

 

[AssignedEva]

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?

Images will be the biggest one. If you have any rationale for making it generate images, or edit existing images, that'll burn a fair few. For example "I am using 12 bits of clipart/stock images on my slide deck. Please make them all look like the same style".
Otherwise anything long context. If you have meeting transcriptions, put an hour long transcript into it and ask it to summarise themes, analyse sentiment from key stakeholders and generate a list of action points with suggested risks/pinch points, that'll use loads. The more it has to read, "think" and write, the bigger the burn.
The other one will be revisions. Instead of taking its output and tweaking it, keep going "no that's not what I wanted, make it more formal" (or whatever).

 

[22 Rimfire]

Something I find really interesting to think about with all of this is that, if widespread adoption of LLMs is genuinely a 10x multiplier to individuals when properly applied, you would expect to see some exceptional parties making exceptional use of it's capabilities. But that kinda doesn't seem to be happening.

I believe the first part of that. Chucking LLMs at white collar workers properly gives them a kind of personal assistant at (relatively) negligible cost. But the downstream of that should be some really weird turbo output from 1%ers in the corpo world, and we're not seeing that yet. Or maybe we are and I haven't noticed.

 

[northstar747]

Or maybe we are and I haven't noticed.

I would assume the super users are smart enough to not automate themselves out of a job

 

[John McAfee]

We are still being required to use gemini at work. They don't give us google cloud access to make it work. It's a useless chat bot. Ive feed it rubrics and criteria for the quarterly reviews and yearly reviews. Had chatgpt make me a prompt. Each day I put what I accomplished in my role into a basic text document. I then feed my bullshit into the bullshit machine and it spits out bullshit. I paste it into the portal. Management comes to me and says this is excellent and exactly how they want these filled out. They have me teach a class on prompt building. Now my org pumps bullshit into it, to feed bullshit to our bullshit management to appease more bullshitters.

 

[22 Rimfire]

I would assume the super users are smart enough to not automate themselves out of a job

That's not quite what I'm talking about. If you're a genuinely high-performing coder you should be able to effectively harness LLMs to augment your current output. That was already high, so the result should be genuinely impressive.

Possibly what's happening is this is able to improve the output of the middle of the pack but isn't able to meaningfully assist the top end of the spectrum. It's just closing the gap.
OR it's possible the top of the pack is the most resistant to adopting this, so we're seeing a flat roll-out due to natural resistance.

That's why I find it interesting to think about. I wonder which it is. Maybe something else entirely.

I know for a FACT it's augmenting middle-of-the-road coders, because today I'm meaningfully contributing to several pretty major software projects that, prior to my adoption of AI coding tools, I completely lacked the spare attention to devote towards improving them.

 

[Schpoots]

I think what companies will do is have someone that acts like a gatekeeper for AI requests. You send your prompt to this person and if it meets some requirements they pass it onto claude and send the results back to you.

 

[Breadbassket]

That's not quite what I'm talking about. If you're a genuinely high-performing coder you should be able to effectively harness LLMs to augment your current output. That was already high, so the result should be genuinely impressive.

Possibly what's happening is this is able to improve the output of the middle of the pack but isn't able to meaningfully assist the top end of the spectrum. It's just closing the gap.
OR it's possible the top of the pack is the most resistant to adopting this, so we're seeing a flat roll-out due to natural resistance.

That's why I find it interesting to think about. I wonder which it is. Maybe something else entirely.

I've seen articles here and on there on what the "ideal" is for coders relying on AI which amounts to attempting to manage multiple AI agents at once :


They very much seem like the minority though.

 

[crowabunga]

>Deep Seek V4 exists, retards don't use it
>Retards waste tokens on PDF conversions
>Their AI is too retarded to just pull one from github or whatever

omegalol

PyMuPdf my beloved~

AI is the cutting edge office productivity equivalent of "printing an email from your desk inkjet to hand carry to the other side of the office where the Ricoh sits to scan to the recipient's inbox" but the immediate inefficiency is abstracted away far enough that fewer people get a gut reaction of horror.

In personal experience, I recently made a quick-n-dirty printable calendar in Word using cells in a table for days in a month, and thought "why not ask Copilot to populate the cells with dates?" It took just as long as typing it by hand, and if I really needed to do this frequently I'd make a template with a VBA script in it.

 

[northstar747]

I know for a FACT it's augmenting middle-of-the-road coders, because today I'm meaningfully contributing to several pretty major software projects that, prior to my adoption of AI coding tools, I completely lacked the spare attention to devote towards improving them.

I see your point.

i suck at coding, a few years ago I need to take a python class and I was a peice of shit that used chatgpt to do everything.

 

[DiabeticSydePhil]

'We want our employees to use AI more'
'We want potential employees asking how many tokens they'll get'
'We spend more on AI than competitors'
'Am employee not using tokens is a red flag'

'Oh no we're spending so much money whan happened'

Niggers did you think AI was free? Did you assume the AI company was your friend and would let you drain water and cut down trees so you could convert pdfs without charging you?

The AI bubble will burst because all but thr largest companies (that are invested in AI anyways) can't afford to waste money on this shit but felt obliged to. I hope AI dies ajd it bankrupts a bunch of jeet and chinese companies and a ton of silicon valley fags kill themselves.

 

[22 Rimfire]

I see your point.

i suck at coding, a few years ago I need to take a python class and I was a peice of shit that used chatgpt to do everything.

Where I think I'm at with this is perhaps it's just useful to a midwit like me, and then diminishing returns from there upwards. That seems to track with observable reality at least. Maybe this genuinely isn't useful for the top x % of coders.

I suspect that's wrong though. I have a sneaking suspicion there's some active resistance at those upper levels to adoption of what is fundamentally, just a fancy screwdriver. Because the screwdriver is kind of insulting, and takes away from processes that are highly satisfying to engage in.
"Hey what if the actual act of coding is not that difficult and we could farm this off to some AI or whatever?" - oof
That shit stings, man. Especially so for the folks at the top of that game who are churning out a lot of lines per day and base their ego at least partially on the act.

Again, fascinating to think about. I'm glad I don't have too much skin in the game.

 

[TJT]

This was always going to happen.

I left a job earlier this year and I was telling leadership that this tokenmaxing is going to destroy your budget once the free taste is up. They are scrambling because AI costs are now >$200k a month and they have to be tyrannical about decreasing it. After hounding everyone to use AI constantly the past two or so years.

 

[22 Rimfire]

This was always going to happen.

I left a job earlier this year and I was telling leadership that this tokenmaxing is going to destroy your budget once the free taste is up. They are scrambling because AI costs are now >$200k a month and they have to be tyrannical about decreasing it. After hounding everyone to use AI constantly the past two or so years.

It probably works if you are able to magically align to it's implications.

As in, if this enables your employees to double their individual productivity, which perhaps it does if they take it seriously, then if you're willing to drop (quite literally) half your staff immediately maybe you convert that 200k in AI spend into 800k in reduced loaded costs.

But is anyone actually doing that? This suddenly depends on the entire class of middle management being willing to dump the human slurry beneath them at rapid pace. Plus you know that whole idea about people willingly obsoleting their own roles. Somehow I'm not feeling it.

 

[30+GameOvers]

I can't wait until the AI bubble pops. I am saving up a bottle of champagne just for that, and I hope everyone who abuses it goes broke.

 

[Soyjak Enjoyer]

There's some cartoon episode where the dumb character randomly gets a ton of money and becomes a socialite and has a house party where he's just giving money out to people insisting they take more and then is suddenly shocked when he runs out of money and the viewer is supposed to laugh at how dumb he is. That is how I view these companies

(Google's AI overview correctly identified the show and episode as Spongebob 'Porous Pockets', which is a decent pun of a title. I suppose this is evidence AI can at least do one thing more productive than converting PDF to powerpoint)

 

[Retrodogkiller]

Ai was advertised as making the mundane tasks more automated so I don't know why companies are surprised when their employees have claud or chatgpt do the bullshit everyone hates doing. I've done AI assisted coding and its basically just a more immediate Google search. Instead of me reading through multiple stackoverflow pages, the AI does it instead. I still have to fix the code to fit the rest of my program but its definitely faster than before. AI has its uses but lol companies deserve to suffer for implementing it without thought or proper governance. Suffer retards.

 

[No-Eyed Square Guy]

(Google's AI overview correctly identified the show and episode as Spongebob 'Porous Pockets', which is a decent pun of a title. I suppose this is evidence AI can at least do one thing more productive than converting PDF to powerpoint)

One of the several reasons Google destroyed its search ability was so users would credit "AI" with being able to find things that searches can't (anymore).

The actual innovation in "AI" search is that in situations where traditional searches can only be made to fail, it can outright lie.

Ask Google about anything you wouldn't trust Wikipedia about, or anything that present-day Reddit won't let its users say—or any subject you know in detail.

You'll approach perfect error: the ideal and intended state of "AI."

 

[Tanner Glass]

Where I think I'm at with this is perhaps it's just useful to a midwit like me, and then diminishing returns from there upwards. That seems to track with observable reality at least. Maybe this genuinely isn't useful for the top x % of coders.

It isn't.

The difference between a run of the mill coder and a top x% coder is consistency. If you're writing important code - you don't just write it and have it spit out whatever, you need to know it, have others be able to build into it, explain it to stakeholders, document it, and build change control processes around it. Being lazy with any of these steps will have a forecastable negative impact that's referred to as a Technical Debt - which is the loss that's going to occur when something messes up/you can't do something.

You can't effectively do that if you didn't write it and there's no forced consistency in AI. Version to version, bot to bot, it changes as it scrapes new and different data. It also doesn't remember or recall things it's done previously and is able to lie instead of answering questions. It also doesn't learn and can't incorporate things it has done previously or elsewhere. You can easily find yourself in a spot where it's easier/simpler/more consistent to start a project over then to work through the AI code.