The Token Burn Episode: What Happens When Your Software Bill Has No Ceiling

Your AI bill just stopped behaving like a software bill. For twenty years, IT leaders got very good at counting seats: buy a hundred, pay for a hundred. Then AI swapped the seat for a meter, and the number stopped holding still.

This episode follows the burn from three vantage points: a financial analyst rationing a $250-a-month token budget he tore through in two days; the tech executive who watched enterprise AI bills climb 7x, 10x, 20x; and the IT leader at a 300-person company who refused to solve it with a usage dashboard. Along the way: Meta's leaked internal token leaderboard, Uber blowing its entire annual AI budget by April, and the uncomfortable question of who profits when everyone's told to use more.

In this episode:

• Why token-based pricing breaks the budgeting playbook IT has relied on for two decades
• What happens to the people using the tool when the meter starts running — and why rationing has a hidden cost
• Why measuring usage is the wrong scoreboard, and who benefits when you keep score anyway
• The mid-market move that beats policing: measure centrally, push the judgment to managers, and get clear on what you're optimizing for

Featuring Brian Elliott, CEO of Work Forward; Daryl Dore, Senior Director of IT & Information Security at Higher Logic; and Benjamin, a financial analyst who spoke with us on condition of anonymity.

Support our sponsor:

This episode is brought to you by Sophos MDR. Running Microsoft security tools and drowning in alerts? Sophos MDR's 24/7 experts investigate and stop the real threats. 

>>> Learn more at: https://www.sophos.com/en-us/solutions/use-cases/microsoft

#ITLeadership  #AICostManagement  #SaaSManagement  #FinOps  #EnterpriseAI  #TokenBurn  #ITAM

Show Notes & Resources

Referenced in this episode

• Meta's internal AI token leaderboard (Fortune) — 85,000 employees ranked by token consumption; shut down days after it leaked.
• Uber burns its 2026 AI budget in four months (Forbes; TechCrunch) — adoption jumps 32% to 84% in a month; spend later capped.
• Jensen Huang on token consumption as a productivity signal (Tom's Hardware).
• Gartner: worldwide AI spending forecast to grow 47% in 2026 (Gartner).
• Zylo 2026 SaaS Management Index — the scale of wasted SaaS spend (Zylo).
• Brian Elliott's newsletter, Work Forward.

Guest: Daryl Dore — Higher Logic.

This episode's sponsor: Sophos MDR, in partnership with Softchoice — 24/7 managed detection and response for Microsoft environments. https://www.sophos.com/en-us/solutions/use-cases/microsoft 

The Catalyst by Softchoice is the podcast dedicated to exploring the intersection of humans and technology. 

Transcript

Producer Ryan 0:02

This episode of the Catalyst is brought to you by Sophos MDR in partnership with Softchoice. If your organization runs Microsoft security tools and your team is still drowning in alerts, we can help. Sophos MDR is the 24-7 expert team that steps in, investigates the real threats, and stops attacks before they cause damage. Visit SoFos.com slash Microsoft or click the link in the description to learn more.

Katey 0:33

Benjamin is a financial analyst. He works inside a large company, one that's been pushing hard to get AI into everybody's hands. We're not using his last name or naming his employer because the project he's describing is still running. A few months ago, he got picked for a pilot. Five, maybe ten people in his part of the org given access to Claude and a budget.

Benjamin 1:02

So everyone in the proof of concept was given, I think at first it was $150, and then it was requested to extend that another $100. So everyone now I think has $250 worth of tokens a month per user.

Katey 1:21

Not a seat, not a license, a tank of gas. He started using it on his Excel models, rewriting formulas, cleaning things up, the unglamorous work that eats up an analyst's week.

Benjamin 1:35

Within a day or two of using it as an Excel add-in, I was curious to see what my token usage was. And so I checked back into my token usage and I realized I was at $60 out of the 250 used. And so I scaled back my usage a little bit following that, and I started using it for like high priority Excel model changes, just trying not to burn through my 250 in the month.

Katey 2:00

Two days, a quarter of the month gone. So he did what any of us would do. He started rationing. Now, while Benjamin was budgeting his spreadsheets, something very different was happening about 3,000 miles west. Earlier this year, an engineer at Meta built an internal dashboard. It ranked the company's 85,000 employees by how many AI tokens they had consumed. The top 250 got badges. In a single 30-day window, Meta staff burned through more than 60 trillion tokens. The number one user averaged 281 billion by himself. Mark Zuckerberg didn't crack the top 250. The dashboard was called Claude Economics. Meta shut it down two days after it leaked to the press. There's a word for what that dashboard was encouraging.

Brian 3:07

Token maxing is the act of developers and engineers primarily in Silicon Valley trying to figure out who can consume the most of them, like voluminous numbers of tokens, millions, billions, trillions, that they're going after to try to create more complex software and more complex tools. You literally have guys inside of some of these companies that set up leaderboards. Who can be the most aggressive user of tokens who can consume the most?

Katey 3:32

Two worlds, one analyst rationing $60 of spreadsheet help, one engineer running a small fortune through a model to win a badge. And they both end up on the same invoice. From SoftChoice, a worldwide technology company, this is the Catalyst. I'm Katie Tekasing. Today's episode: the software bill just stopped behaving like a software bill. For 20 years, IT leaders got very good at counting seats. Now the meter runs on behavior, and nobody quite knows sure how to budget for that. We're calling it the token burn episode. Act one, the meter is running. Here's what makes AI costs different from every software cost that came before it. A SaaS license is a fixed, boring, predictable number. You buy a hundred seats, you pay for a hundred seats. If 40 of those people never log in, then we will come back to that. You still pay for a hundred seats. It's wasteful, but it's knowable. AI isn't a seat, it's a meter. And what spins the meter isn't how many people you hired, it's what they do all day. Brian Elliott advises executives through exactly this kind of transition.

Brian 5:07

I'm Brian Elliott, I'm the CEO of Workforward. I spend a lot of time with executives wrestling with workplace transformation and how to make work better for teams and organizations. And these days, that's all about AI. The size of the bill is going up dramatically. The primary builders of the tools have started switching over from seat-based pricing to token-based pricing. And some companies have seen the cost go up 7x, 10x, 20x in terms of what the size of the bill is. You're literally now seeing some people that are paying as much for their AI tokens as they're paying almost for their engineering employees.

Katey 5:40

Seven times. The entire year in four months. Adoption of AI coding tools inside Uber went from 32% of engineers in February to 84% by March. Individual engineers were running $500 to $2,000 a month in costs. Uber had encouraged this. They'd put usage on internal leaderboards. But go back to Benjamin, because his version of this is quieter. And it's the one that will show up in your building. Once he started rationing, the tool actually changed. Not technically. Same model, same add-in. But what changed was his own relationship to it. He's holding tokens in reserve for an emergency that hasn't even happened yet. Which means that most days he's doing the work the slow way on purpose.

Benjamin 7:12

I made the joke to my girlfriend after I burned through all those tokens. I was like, after I stopped using it, I was like, wow, this is so much slower than c than anthropic. You know, I wish I still had more tokens. I was like almost bec uh like lost my self-reliance a little bit to a certain extent once I realized how powerful it was.

Katey 7:29

Lost his self-reliance. That's an unusual thing to hear from a man talking about a budget line, but it's pointing at something very real. The tool didn't just make him faster, it rewired how he worked. And then the meter took it back.

Benjamin 7:50

Like extensively. And then, like again, the bait and switch of like I started using it, and then all of a sudden I didn't have tokens to use anymore, or like I didn't feel comfortable using more tokens right away. And so I lost the ability to use it, and and it did impact my work efficiency for sure.

Katey 8:07

And remember, Benjamin is the careful one. On an internal call about AI usage, somebody else's numbers came up.

Benjamin 8:16

They brought up another anecdote of one of the people in our IT organization within that sits within finance, talking about how they had burned through their 250 tokens, like $250 worth of tokens, and a session or two of using it for, I'm sure more in-depth things that I was using it for, but they burned through it immediately.

Katey 8:35

A session or two, an entire month's budget, and nobody did anything wrong. That's what's so disorienting about this. No misuse, no rogue spending, no shadow IT. Somebody used a powerful tool to do a hard job, and then the bill arrived. Brian says that's exactly what happens when you push the cost decision all the way down to the person doing the work.

Brian 9:04

Partly what you're then asking users to do is to think on each individual query or activity or task or job or agent that they're building, what's the right underlying model to use from your service provider, right? That's just putting way too much weight on the individual. If you keep putting in front of users, you have to make a choice about which model to use in order to not blow out your budget on something. People are going to stop using the tool. It's just too expensive cognitively to have to do that work every time.

Katey 9:31

Which is precisely what Benjamin did.

Benjamin 9:34

Claude is definitely a more powerful model, but yeah, the token barn is real.

Producer Ryan 9:54

Not because teams don't care, but because there aren't enough hours in the day to chase every signal, separate the real threats from the noise, and keep everything else running. And that's when something does slip through overnight or over a weekend. The damage is done before anyone gets to it. That's the gap that SoFos MDR was built to close. SoFOS MDR is a 24-7 team of security experts that monitor your Microsoft environment, investigate suspicious activity, and, this is the important part, actually takes action. Not just an alert, not an escalation that lands back on your desk at 2 in the morning. SoFOS analysts can disable compromised accounts, kill active sessions, and isolate devices directly inside your Microsoft tenant so threats get stopped before they spread. And because SoFos works across Microsoft Business Premium, E3 and E5 without replacing the fender, you're not ripping anything out. You're making what you already have actually protect your business. SoftChoice now offers a dedicated SoFos MDR deployment service so you can get fully operational quickly with expert-led implementation and no disruption in your environment. To request a demo or security assessment, visit SoFos.com slash Microsoft or click the link in the description.

Katey 11:07

Act 2. So the bill is real, it's volatile, and it's moving faster than anyone's spreadsheet. Here's where it gets strange. The industry's response has been to measure it. Track who's using AI, how much, how often, put it on a dashboard. In a lot of companies, put it in a performance review. And Brian will tell you, we have run this experiment before.

Brian 11:38

We've gone through this game a bunch of different times, though, right? We've seen companies set up these things that are activity-based measures. So I don't have a great way of measuring outcomes for you. So instead, I'm going to look at your output. I'm going to measure on return to office how many days a week you sit in your seat, right? I'm worried about how responsive you are, so I'm going to monitor uh keystrokes and make sure that you're logging at least eight, nine, ten hours a day on the computer. Well, if you're logging usage, that's a pretty easy one to gamify. And the problem is play stupid games, win stupid prizes.

Katey 12:09

He tells a story about Wells Fargo, a company that decided to monitor whether its remote employees were actually at their computers.

Brian 12:18

What folks did was they went and spent $25 on Amazon to buy a mouse jiggler. A mouse jiggler, you plug it into your computer, and it basically makes sure that something's getting clicked every once in a while.

Katey 12:27

Wells Fargo fired people over it. But Brian's point isn't about the employees.

Brian 12:33

You shouldn't be surprised when people then game the incentive instead of saying, hey, are they actually generating outcomes? Are they actually getting the job done? Now swap the $25 mouse jiggler for an AI agent has caused people to gamify it to the point now where Amazon and Meta have both pulled back from that because they know that engineers are actually pretty good at finding ways to consume it. One engineer told me, you know, if I really wanted to max out on this, all I'd do is sit down and create a Nintendo game from the 2000s every morning at 10 a.m. Uh using my work computer.

Katey 13:07

An engineer, in other words, can run up an enormous number without producing anything at all. The metric is trivially easy to satisfy and almost impossible to argue with, which raises an obvious question. If usage is such a bad proxy for value, who exactly has been telling everybody to use more? At NVIDIA's developer conference in March, CEO Jensen Huang laid out a thought experiment. Take a software engineer you're paying $500,000 a year. At the end of the year, ask what they spent on tokens. If the answer is less than $250,000, Huang said he'd be deeply alarmed. Brian was speaking at a conference in Las Vegas not long after. Three speakers ahead of him quoted it.

Brian 14:02

And all of them were talking about this as if it was like a fait accompli. That's what we should be doing. Look, Jensen Huang has a commercial interest in doing this. This is the guys who run the pencil factory selling you pencils. So, you know, token maxing sounds really good if you're one of the primary model builders or you're a you know software development company, but it's not the greatest thing and it's not super smart when it comes back to pretty much everybody else in the ecosystem.

Katey 14:27

To be fair to Huang, he isn't exactly wrong that engineers with good tools produce more. But what he's selling is a ratio, not a strategy. And a ratio makes a very poor budget. According to Gartner, worldwide AI spending will reach $2.59 trillion this year, up 47%. And in the same forecast, Gartner's own analyst names the problem. CIOs are struggling to prove value from their AI investments and demonstrate tangible business outcomes. Enormous spend, unproven return. And the scoreboard the industry has handed you measures that spend.

Brian 15:44

So if you want to give employees a perk, if you want to give them a benefit, if you want them to be engaged in your company, giving them training, giving them support, letting them experiment and learn is a much better perk for almost everybody than thinking about like, okay, you get, you know, uh uh $500 a month token budget.

Katey 16:04

Act three. What are you optimizing for? Everything so far has happened at companies with 85,000 employees and billion-dollar research budget. That isn't most of us. So let's go somewhere smaller.

Daryl 16:20

So, Daryl Doer, I'm the senior director of IT and Information Security at Higher Logic. So I covered all of our corporate IT for a fully remote workforce, our corporate systems, as well as security across both corporate as well as our products.

Katey 16:35

HigherLogic is about 300 people, fully remote, mostly US, and some in Canada and Australia. Daryl runs corporate IT and security across all of it. And when he arrived four years ago, he walked into something a lot of IT leaders will recognize immediately.

Daryl 16:56

Yeah, when I first joined Higher Logic, I came in as our first corporate systems director. And so a big part of that role was to bring more strategy and alignment across our corporate systems environment. And when I came in, there was really just a challenge of knowing what are the systems we're using, who's managing them, how much do they cost, and what does the licensing look like for a lot of those. So when it meant budgeting, we got really good information from our finance team aligned to how finance sees it, but that's very different than how an IT leader looks at it in order to manage and optimize their licenses.

Katey 17:29

Good numbers from finance, just not useful ones. So he put in a SaaS management platform and pointed it at the financial system, the expense system, and identity.

Daryl 17:41

When we first put a SaaS management tool, it started pulling information from our financial system, our expense system, and our identity system. And it was quite surprising to see the number of apps that we weren't even aware of that we were using, that we were paying for, that we just didn't have an understanding of from a corporate IT perspective.

Katey 18:01

This is the part every IT leader has lived. It has a name, SaaS Sprawl. And the numbers around it are grim. Xylo's 2026 SaaS Management Index puts the average organization's wasted spend on unused licenses at close to $20 million a year, with roughly half of all licenses underused or unused in a given month. And Daryl will tell you, it was never negligence, it was velocity.

Daryl 18:35

So at that time, we definitely had some SAS sprawling. I think it's just a part of being a company that moves really quickly, that had had a couple significant acquisitions over the previous period, and a lot of staff who really came from other places where they used a lot of specialized tools to help them get their job done. And when they joined our company, they were really interested in using those same tools.

Katey 18:57

Fast growing company. Acquisitions. Good people bringing the tools they loved from their last job. Nobody was wrong, but everybody was spending. And then AI did the same thing to him again.

Daryl 19:11

AI is interesting because it definitely caused some initial sprawl when we were really fully embracing AI inside organization. The common AI providers didn't have a lot of robust tooling or capability to solve all of our business needs inside them. So there was an initial lift where we ended up buying a lot of purpose-built AI tools to help solve some immediate problems.

Katey 19:33

Here's the thing though. Daryl has already been through this movie. He has the platform, he has the visibility, and he has the discipline. He is, by his own account, not losing sleep over the token bill. What worries him is what's coming next.

Daryl 19:51

I think duplicate functionality is a big thing that I do focus on, and especially how that drives cost. So initially, we saw a lot of great AI capabilities introduced into tools that was free, and now I'm seeing a lot of vendors charging for the consumption of that or pushing for higher licenses. And I think there's a lot of vendors that are still on that path and that's coming. My biggest worry is we're enabling AI in every single tool that we have, which creates more fragmentation of our business processes and then starts to drive a lot of costs that we have to manage in a whole bunch of different tools.

Katey 20:23

It isn't the price of any single token that keeps him up. It's the prospect of running 20 separate AI meters inside 20 separate tools, each with its own bill, its own contract, and its own governance problem. Which brings us to what he actually did about it. And it's about the least IT leader thing an IT leader could do. He gave the decision away.

Daryl 20:50

I think we've been pretty careful in terms of how we manage and monitor tokens. But by the same vein, like there is a lot of value in fully enabling staff to do what they need to do with AI. So we've really tried to strike a balance where we're managing and monitoring, but a big part of that is really just charging it back and providing visibility to the managers so that they can manage appropriateness of AI use inside their own teams. Because I just don't think it works for an IT leader to sit here and trying to decide what is and isn't appropriate in terms of token usage across our business.

Katey 21:22

Listen to what he's separating there. IT does the measuring, the manager does the judging. Because the manager is the only person in the building who knows whether an $80 afternoon on a financial model was actually worth $80. Brian gets to the same place from the opposite direction. He works with the companies writing the biggest checks. And his advice starts with the arithmetic everybody skips.

Brian 22:00

Classic example here are places where it's already made a material difference. Let's take um software engineering and development as an example. The tokens themselves can be expensive, but it actually gets uh your development dollar to go a lot further. So what used to be an eight-person team might now be a three or four-person team that's able to do just as much as they could before. The starting point is not give every employee an infinite budget of tokens and let them have at it. The starting point is have your hypotheses about where AI can make a difference because you and then run some pilots in those areas, see if it's actually proving out, put more money into it until you start finding out that the money's not actually returning uh what you thought it was, right? Don't max it out, slowly increment it up, then go experiment in the next couple of areas.

Katey 22:42

And he's specific about how many bets. Two or three, not a thousand, which sounds like ordinary prioritization advice, right up until he explains why it isn't.

Brian 22:57

You can't feed a squirrel enough to turn it into an elephant. If all of your pilots are things that actually don't have major impact on the company, if they're all small, tiny incremental improvements, maybe that'll move things together in the aggregate, but you're a heck of a lot better off picking a couple of areas to invest that are actually really important to the business results that you're trying to generate.

Katey 23:17

And then there's the cost nobody puts in the budget at all. Brian told our producers about a chief people officer he'd spoken with, a tech company that dropped AI into its engineering org and watched the code start flying.

Brian 23:32

I actually talked with a uh chief people officer the other day of a tech company who described the acceleration of development inside their software organization as putting a Ferrari in the jungle. They dropped AI into it. It's accelerated code development like mad. But the product organization now can't feed the engineers fast enough. The marketing and sales team can't keep up with what's coming out of product and engineering.

Katey 23:54

A Ferrari in a jungle. The engine is magnificent, and there's nowhere to drive it because the bottleneck was never the code.

Brian 24:06

The major impediment isn't technology itself. Most of what we're trying to do with this stuff is a generation or two behind in the AI itself. It's the humans, it's workflow, it's process inside of companies, it's redesigning how you get work done.

Katey 24:25

One more thing about Benjamin. Somewhere in the middle of his rationing, he did the thing his entire profession is built to do. He ran the numbers on himself.

Benjamin 24:37

When I was doing the math on it, and I said, you know, this probably saved me X and X amount of hours. This is roughly what I get paid per hour, and I did the math on it. It worked out. I was able to, you know, transition my hours into doing something else. But again, did I add $60 to our bottom line within those three hours that I worked? Possibly, possibly not.

Katey 24:59

Possibly, possibly not. That is the entire enterprise AI ROI debate delivered by a financial analyst about a single afternoon. Near the end of his interview, our producers asked Daryl if there was anything else. He said he'd actually thought about it beforehand.

Daryl 25:17

The real advice to any IT leader is to be really careful about what you're optimizing for. I think a lot of us can be too guilty of optimizing for the cost, optimizing for the IT management, or even optimizing for the security. What we're optimizing for is an employee enablement. And sometimes our decisions might save a little bit of money, but really cost the organization because human productivity is the most important thing in a business like ours. Really just think about for your organization, for what you're doing, what are you optimizing for? And really focus on that. And a lot of times it's not the traditional drivers of IT. And a lot of times you might end up spending more money or be inefficient because it's the right thing to support your employees.

Katey 25:58

The meter is real. The bill is volatile in a way your finance team has never had to model. And there is enormous pressure right now from vendors, from the trade press, from the guy who owns the pencil factory to solve that by measuring how everybody uses. But usage was never about the thing being bought. Uber measured usage and got a leaderboard and an empty budget by April. Meta measured usage and got a token legend. Benjamin's company measured usage and got an analyst who is rationing a spreadsheet. The IT leaders who come out of this well will not be the ones who tracked tokens best. They'll be the ones who were clear before the invoice landed about what they were actually trying to buy. The catalyst was reported and produced by Tobin Dalrymple and the team at Pilgrim Content. Editing by Ryan Clark with support from Philippe Dimas, Joseph Bayer, and the marketing team at SoftChoice. Special thanks to Brian Elliott, to Daryl Doer, and to Benjamin, who spoke with us on the condition of anonymity. The catalyst is brought to you by SoftChoice, a worldwide technology company. If the topics in today's episode struck a nerve, visit SoftChoice.com to discover how we can help you manage IT waste and unlock ROI in your AI strategy. We help thousands of organizations, just like yours, do that every year.

Producer Ryan 27:40

Thanks again to SoFos MDR and SoftChoice for supporting today's episode. If your Microsoft security tools are generating more alerts than your team can handle, SoFos MDR is worth a serious look. Visit SoFos.com/slash Microsoft or click the link in the description.