Wellness
Rewards
Recognitions
Learn more about Wellable's all-in-one
employee engagement platform
The AI Token Gold Rush Is Over (At Least for Tesla). Now What?
Estimated reading time: 6 minutes
In this week’s episode, Nick and Geoff reunite after Nick’s summer hiatus to break down three recent headlines. Tesla caps employee AI token spending at $200 a week, a dramatic reversal from the gamified leaderboards of just six months ago. Starbucks bets it can replace $400 million in enterprise software spend by rebuilding tools in-house with AI. Lastly, Gartner projects one in five companies will cut more than half their middle managers by year-end, a structural efficiency trend with consequences that are only starting to show up.
Short on time? Here are the key takeaways:
- Tesla now requires manager approval for AI token spend above $200 per week, and Uber blew through its entire 2026 token budget by March
- Caps might be signaling negative ROI, not just cost management, since companies with these resources would keep spending if the returns were there
- Starbucks is attempting to replace roughly $400 million in annual enterprise software contracts by rebuilding tools in-house with AI and internal engineers
- If Starbucks succeeds, it could mark a meaningful shift in the build-versus-buy calculus for enterprise software and accelerate the so-called SaaSpocalypse for software vendors
- Gartner projects one in five companies will eliminate more than half their middle managers by the end of the year, but a cautionary case study shows a VP of engineering with 47 direct reports approving decisions in Slack with no context or coaching until their best engineer quietly quit
- Only 6% of Gen Z aspires to senior leadership according to Deloitte research, a rational response to watching middle management disappear and remaining managers stretched to unsustainable spans of control
Episode Summary
The AI Token Spending Reversal
Six months ago, companies were building leaderboards to encourage employees to use as many AI tokens as possible, tying consumption to bonuses and performance metrics. The reversal, at least at some organizations, has been swift. Tesla now requires manager sign-off for any AI token spend above $200 per week. Uber burned through its entire annual AI budget in the first quarter and has since imposed a $1,500 monthly cap. Microsoft has shifted its own internal policies multiple times without finding stable ground.
Nick’s read is that the caps are not primarily a cost management story. Tesla, Uber, and Microsoft all have the resources to absorb higher AI spend if the returns justified it. The more likely explanation is that the returns are not there. In some cases, Nick argues, token spend has likely produced negative ROI, with engineers using tokens on personal projects and low-value tasks just to climb leaderboards, rather than on work that actually benefited the business. The gold rush mentality has been replaced by a reckoning: companies are now in the difficult and slower work of figuring out where AI actually delivers value and limiting spend to those use cases while they do.
Geoff connects this to a broader pattern: the next phase of enterprise AI adoption is probably not more spending but more specificity. Companies that can identify the exact workflows where AI produces measurable output improvement, and standardize tooling around those use cases, will be better positioned than the ones still experimenting across every function simultaneously.
The Starbucks Bet: Replacing $400 Million in Software with AI
Starbucks is attempting something that, if it works, has significant implications far beyond the coffee industry. The company is planning to rebuild much of its enterprise software stack, including tools from Oracle and other large vendors that it has used for decades, using its own engineers and AI. The total software spend they are targeting is approximately $400 million annually. The goal is to complete meaningful portions of the rebuild by the end of the year.
Geoff’s skepticism is well-founded. Starbucks already attempted an AI-powered inventory management system and quietly reverted to manual counting after the tool failed to deliver. The history of enterprise software replacement projects is not encouraging, and Starbucks is not a technology company. But Nick’s counterpoint is worth taking seriously: Starbucks has set a short enough timeline that we will know relatively quickly whether the bet is paying off. If they successfully cut even one or two major vendor relationships by year-end, it will send a signal to every company paying large enterprise software contracts that the calculus on build versus buy has meaningfully shifted.
The broader implication, which Nick raises directly, is that if AI makes enterprise software rebuilds feasible for companies like Starbucks, the threat to SaaS vendors becomes more concrete and more immediate.
Middle Management: The Cautionary Tale Inside the Gartner Data
Gartner projects that one in five companies will eliminate more than half their middle managers by the end of 2025. The efficiency argument for that move is real: fewer management layers accelerate decision-making, reduce headcount costs, and create cleaner organizational structures. But a case study buried in the data Nick and Geoff discuss tells a different story.
One technology company cut 70% of its engineering managers in 2024, saving $3.2 million. Six months later, the VP of engineering had 47 direct reports and was approving decisions in Slack between meetings with no context and no coaching. The company’s best engineer quit. In her exit interview, she said nobody knew what she was working on or why it mattered. The VP had been too stretched to notice she had disengaged until the resignation arrived.
Nick’s point is that the short-term math on cutting middle management often works. Recent feedback is still in circulation, people know what they’re doing, and the cost savings are real and immediate. The long-term damage is slower and harder to see until someone like that engineer walks out the door. The people who leave are disproportionately the high performers with options, the ones who can afford to act on their dissatisfaction.
The Gen Z data adds another layer. Only 6% of Gen Z aspires to a senior leadership role, according to Deloitte research. Nick’s interpretation is not that Gen Z lacks ambition but that they are making a rational calculation: they are watching middle management layers disappear, seeing the managers who remain stretched thin across impossible spans of control, and concluding that individual contributor roles offer more stability and less exposure to the risks of a trend that has nothing to do with their individual performance.
The final thread Geoff raises is manager accountability for AI-generated work. Research shows that managers reviewing AI-produced documents spend less time on that review than they would spend on human-generated work, because the perception is that AI output requires less scrutiny. Nick’s counterargument is that it requires more: hallucinations, confident wrong answers, and the absence of context that a human contributor would naturally include all make AI work product a higher-risk review challenge, not a lower one. As agent usage grows and more work is produced without human authorship, the governance gap between what managers are expected to own and what they are actually reviewing carefully is going to widen.
Frequently Asked Questions
Tesla requires manager approval for AI token spending above $200 per week, a significant reversal from earlier this year when the company encouraged high token consumption through gamified leaderboards. Nick’s interpretation is that the cap reflects poor ROI rather than simple cost management. Companies with Tesla’s resources would continue spending if the returns justified it. The cap suggests they are not seeing meaningful, measurable returns from unconstrained AI spend and are pulling back while they figure out where the value actually is.
Starbucks is planning to replace approximately $400 million in annual enterprise software spend, including contracts with Oracle and other major vendors, by rebuilding those tools in-house using AI and internal engineers. The company has committed to meaningful progress by year-end. If successful, it would be one of the most significant examples of a non-technology company using AI to exit large enterprise software relationships, and it would send a signal to every company paying comparable contracts that the build-versus-buy calculation has shifted.
The SaaSpocalypse refers to the scenario where AI enables companies to build custom software in-house at a cost and complexity level that makes enterprise SaaS contracts unnecessary. It has been a theoretical concern for software vendors for several years. The Starbucks story is one of the first high-profile real-world tests of whether a non-technology company can actually execute that transition. The outcome is not yet clear, but the timeline is short enough that the answer will be visible by early 2028.
Gartner projects that one in five companies will eliminate more than half their middle managers by the end of 2025. For HR leaders, the relevant question is not whether the number makes financial sense in the short term but what it does to development pipelines, employee retention, and accountability structures over time. The case study of a VP with 47 direct reports and a best engineer who quietly disengaged and then resigned is a concrete illustration of what the short-term math misses.
According to Deloitte research, only 6% of Gen Z workers aspire to senior leadership. Nick interprets this not as a lack of ambition but as a rational response to what they are observing: middle management layers being cut, remaining managers stretched thin, and an overall trend that penalizes people for being in certain organizational positions regardless of their individual performance. The individual contributor role looks more stable, more insulated from structural risk, and increasingly capable of being high-impact with AI tools available.
Yes, and current behavior suggests most are not doing so. Research shows managers spend less time reviewing AI-produced documents than human-produced ones, operating under the assumption that AI output is reliable enough to require lighter scrutiny. Nick argues the opposite is true: AI outputs are more susceptible to hallucinations, confident errors, and missing context than human work, and require at least as much review, if not more. As agent-produced work becomes more common, the governance gap between accountability and actual review behavior will become a significant organizational risk.
Full Episode Transcript
Nick: Welcome to the Wellable Weekly Podcast, where we talk about key topics and trends at the intersection of well-being, technology, and HR. I’m Nick, along with my good friend and co-host Geoff. Welcome back, Nick. I’ve been gone for a little bit.
Geoff: The dream team’s back together. Good to have you back. We had some fantastic guests, but nothing like getting the original two back in the studio.
Nick: Completely agree. We did have some great guests — definitely check those episodes out if you haven’t. Feels good to be back. Had a good little vacation, got a little sun. AI is still dominating the news, and the big story now is expenses. How much does it cost? Which model should you use? Which ones are cheaper? Open source versus closed source? Tesla came up today and it’s interesting because Tesla is kind of the quintessential non-AI AI company. They are not out there developing their own frontier model — Elon Musk has his own version with SpaceX — but the future of Tesla is really robots and autonomous driving, which is basically an AI tool in the vehicle. So they’re as AI-forward as you can be without being a frontier model company. And what was really interesting is that they announced a new policy: Tesla employees will now need to get sign-off from their manager to spend more than $200 a week in tokens from AI. The one exception is SpaceX’s XAI — if you want to use their beta applications, there’s no limit. But for people who want to use Claude, OpenAI, or any true frontier model in a meaningful way, they’ll likely need manager approval.
Geoff: It really illustrates the spectrum of AI adoption across different types of companies. At the high end, you’re putting guardrails on how much employees can spend on AI credits. At the other end, companies are just trying to get employees to experiment with tools at all. And what we saw earlier this year was almost a gold rush mindset — spend, spend, spend, use as many credits as you can. Token maxing, exactly. And companies like Tesla are now completely reversing that, saying we have to rein this in. It went too far. And even companies still figuring out where to start should watch this, because Tesla and Microsoft are still seesawing on their own policies and haven’t found stability.
Nick: What a turnaround from six months ago. Companies were gamifying token consumption, building leaderboards, engineers competing to get to the top — to the point that they were using tokens inefficiently on projects that didn’t need them, on personal projects, all of it. And then they just completely reversed and put a pretty material cap in place. Tesla is in the news partly because the cap is so low. Uber went through their full 2026 annual token budget by March and has since capped at $1,500 a month. These companies were encouraging everyone to use tokens for every little thing and now they’ve gone completely the other direction.
Geoff: The next phase for a lot of organizations is probably picking a lane and committing to it. Your average employee right now is probably trying to figure out: am I allowed to use Claude? Should I just use Copilot because it’s in the Microsoft stack? Can I use ChatGPT? When organizations can pick a lane, give good guidance, that’s probably when you’ll start to see more consistent and measured adoption.
Nick: The real story behind this is that Tesla has the funds, Uber has the funds, Microsoft has the funds. If there was a real ROI to this token usage, I don’t think we’d be having this conversation. Even if token expenses were quadruple, if Uber had burned through its budget in one quarter but had real results to show for it, they would find a way to allocate more dollars. The Silicon Valley mindset is that it’s okay to invest heavily in areas with strong returns. The fact that they’re capping it tells me they’re not seeing those returns. And my guess is that in some cases the dollars spent have a negative ROI, at least today. While they figure it out, they’re going to limit usage to better manage the spend.
Now, Tesla is one side of the story. There’s another big article about Starbucks. Which is interesting because Tesla is a technology company, and Starbucks — while they have a CTO and proprietary products like a mobile app — is a coffee company. Their annual software spend is about $400 million. A lot of that goes to Microsoft, Oracle, IBM, for point-of-sale systems, supply chain logistics, and so on. They believe they’re going to use AI to eliminate a huge portion of that spend. Starbucks is trying to do $2 billion of cuts across the company, and a big portion of those cuts are going to come from software. They believe that their own engineers, plus AI, plus leadership from a new CTO, are going to rebuild software they’ve been using for decades from these software giants.
Geoff: The software companies you mentioned are going to closely watch this. It could be a pivotal exercise in that SaaSpocalypse narrative — whether companies are eventually going to be able to reproduce the software they’ve been paying third parties for for decades. And Starbucks is not a technology company. They’re taking on pretty lofty initiatives. In the same article, I remember reading that they tried an AI-powered inventory solution and then abandoned it and reverted to manual counting. You wonder how much of this is really going to play out in practice. The idea of bringing everything in-house, cutting software spend, sounds great in theory, but in practice I’m not sure you’ll see the results that allow you to keep the same efficiency and output with tools cobbled together from scratch.
Nick: I’m going to lose my mind if the final product from Starbucks ends up being an abacus and some guy in the back counting boxes. That said, a lot of the AI narrative is always so future-oriented. Tesla talks about humanoid robots and fully autonomous driving — while some of that exists today, we’re still really far from literally no drivers at all. But Starbucks is claiming they’ll see some of these software rollouts by the end of the year. We’re talking less than six months from now. We’ll see relatively quickly if this comes to fruition. If they successfully remove Oracle from their stack, that’s going to make every company — including companies like Wellable — really think hard about whether you can vibe code your way out of enterprise software relationships.
Moving to the third story: Starbucks is trying to save money, and every company is trying to save money in turbulent economic times. They’re also trying to invest heavily in AI. That money needs to come from somewhere. And as it relates to headcount, it’s disproportionately hitting middle managers.
Geoff: The middle management story is one we’ve been covering for a while. Gartner is now projecting that one in five companies will eliminate more than half their middle managers by the end of this year. Flatter structure, faster decisions, better margins. The surface narrative is appealing. In practice, I think it’s a tale of two stories. Some organizations genuinely have bloat and layers of management that slow decision-making. But some are overdoing it. The case that stood out to me: one tech company cut 70% of its engineering managers in 2024, saved $3.2 million. The VP of engineering now has 47 direct reports, approves decisions in Slack between meetings with no context, no coaching. Six months in, their best engineer quit. In the exit interview: “Nobody here knows what I’m working on or why it matters.” The VP was too swamped to notice she had completely checked out until the resignation hit.
Nick: That happened pretty quickly. In the short term, you cut middle managers, your team still has recent feedback, they know what they’re working on, you’ve lowered costs, technically sped up decision-making. That all looks good initially. And maybe AI can supplement and fill the gap — that’s the belief. Jack Dorsey’s vision is no managers at all, with all 6,000 Block employees reporting to him. He knows he can’t give feedback to all 6,000, but he believes there can be an AI version of Jack Dorsey that gives feedback to every employee and aligns the organization around a common vision. Remains to be seen. But until that’s proven, you’re really making a short-term bet. And the people you’re losing are the high-potential employees who aren’t satisfied, aren’t getting developed, and have the options to act on it.
The stat that actually stood out to me more than the one-in-five projection is the Deloitte finding that only 6% of Gen Z aspires to be in senior leadership. I think they’re seeing the risk. They’re thinking: why not just be an individual contributor? More stable, less exposure to a structural trend that has nothing to do with my personal performance. I can’t imagine being 22 to 26 and thinking about career trajectory in an environment where only 6% of your peers are even trying to get to senior leadership.
Geoff: They see middle management being cut and the managers who remain stretched to impossible spans of control, without the time or capacity to develop their people. The individual contributor role looks pretty appealing. And the article that stood out to me was the BCG research on managers reviewing AI-generated work. When managers were vetting work conducted by agents rather than human employees, they spent less time reviewing it — the perception was that because it came from software, it doesn’t require the same diligence as a human-generated work product. But with hallucinations, AI confidently giving wrong answers, there are so many reasons why AI-produced deliverables require just as much scrutiny, if not more. That perception gap is a whole new area of development that most people don’t have experience with yet.
Nick: The agents don’t need feedback. If you think your job as a manager is to review work product and the buck stops with you, but also to give feedback — and 360 reviews make that visible — being required to review the work makes you do both. With agents, you’re only doing the first part, and you’re not even doing that well. I remember seeing that BCG article and thinking that’s just crazy. You have to review it just to do the second part of your job, which is give feedback. And skipping that review because “it came from AI” is a governance gap that’s only going to grow as agent usage scales.
Geoff: Can’t waste those tokens. Thanks as always to everyone who tunes in. You can get your Wellable Weekly fix on Apple Podcasts, Spotify, or wherever you get your podcasts. Be sure to subscribe to the Wellable Weekly newsletter for all the latest insights. Thank you.
-
Share
-
-
-
-
