Let's cut the preamble.
Tokenmaxxing was a fucking mess.
I mean, really. CEOs were handing out AI access like it was free coffee at a startup hackathon. "Just use it," they'd say. "Push it to the edge. See what happens." And people did. Uber burned through its entire annual AI budget in three months. Meta killed its internal leaderboard because employees were gaming it. Companies started cutting Claude licenses like they were trimming a Christmas tree.
This wasn't innovation. It was an orgy of compute.
And now? The bill came due.
Tiffany Luck, the NEA partner who helped bet on e-commerce before it was cool, is sitting in the middle of this wreckage. She's not surprised. She's seen this movie before. The same thing happened with cloud adoption in 2012. Everyone rushed in. Everyone assumed it was magic. Then the CFOs showed up with spreadsheets and asked, "Where's the return?"
AI isn't different. It's just faster.
The truth? Enterprises are still figuring out their AI ROI. And if they don't fix this fast, the next wave of AI investment won't come from VCs. It'll come from auditors.
What's wild is how predictable this all was. We've been here before—every major tech wave follows the same arc. Hype. Excess. Reckoning. Then, eventually, real value.
But the reckoning always hurts more than anyone expected. Because by the time you realize you're burning cash, you've already committed to the narrative. You've told your board it's strategic. You've hired people around it. You've built culture around it.
Luck knows this better than most. She didn't just watch the cloud cycle—she lived it. And now she's watching AI repeat the same pattern, just at lightspeed.
The difference? There's no going back. You can't un-burn that budget. You can't undo the licenses you've already cut. All you can do is figure out what actually works and double down on that.
That's where we are right now. Past the hype. Into the mess. And somehow, that's actually a good place to be.
Because when the smoke clears, only the real value remains. Everything else? Just noise.
Why "Magic Moments" Aren't Enough
Luck talks about "magic moments" like they're the holy grail. And sure—they're real. I've seen them. A customer service rep who used an AI agent to cut their ticket resolution time from 12 minutes to 90 seconds. A marketer who generated 50 campaign variants in ten minutes and found one that outperformed the last quarter's top performer.
But here's the thing: magic moments don't scale. They're accidents.
Personal AI agents? They're not going to click because they're "smart." They're going to click because they're indispensable. And that requires something enterprises haven't figured out yet: intentional design.
Think about it. You don't use Siri because it's magical. You use it because you need to send a text while driving. You don't use Google Maps because it's brilliant—you use it because you don't want to get lost.
The same goes for AI agents. If your agent doesn't solve a specific, repeated, painful problem, it's just another tab in your browser you'll close in three days.
Luck says forward-deployed engineers are the Trojan horse. And she's right. Those are the people who aren't waiting for permission. They're building agents for their own teams—automating status updates, summarizing meetings, pulling data from five different systems into one Slack message. They're not trying to impress the CTO. They're trying to get home on time.
That's the real playbook. Not the grand vision. Not the keynote demo. The quiet, unsexy automation that saves someone 20 minutes a day.
I keep thinking about how most enterprise software fails. Not because it's bad technology. Because nobody actually uses it. You build something fancy, launch it with fanfare, and then... crickets.
AI agents are at that exact same inflection point. The tech works. The demos are impressive. But if you're not solving a problem that keeps someone up at night, they'll find something else to do with their time.
The forward-deployed engineers get this. They're not building for the boardroom. They're building for their own pain. And that's why it works.
That's the lesson here: start small. Start specific. Start with the problem that actually hurts.
Not the vision. The pain.
The AI Stack Is the Real Gold Mine
Here's what nobody's talking about: the value isn't in the models.
Everyone's obsessed with who's got the biggest LLM. OpenAI. Anthropic. Mistral. But the real winners? The startups building the plumbing.
There's a company called PointFive that's raising money to help enterprises track AI spend across dozens of vendors. Another one, Baseten, is building a layer that lets engineers deploy and monitor agents without needing a PhD in distributed systems. There are teams building guardrails, cost dashboards, usage analytics, even AI audit trails.
This is the infrastructure layer that'll make or break enterprise AI. For a deeper look at why trillions in AI infrastructure spending haven't translated into returns, see our analysis of the AI investment gap.
We're not in the model race anymore. We're in the management race.
Luck gets this. She's not betting on the next GPT-5. She's betting on the companies that will help enterprises stop burning cash on AI like it's a startup's last $50k credit card.
And here's the kicker: the companies that build this infrastructure? They're not going to be the flashy ones with the slick websites. They're going to be the quiet ones with the boring names—like the people who built the routers and firewalls that made the internet work.
Think about it. When you think about the internet's infrastructure, what comes to mind? Cisco? Juniper? Maybe some obscure company that makes network switches.
Nobody's talking about them at parties. But they built the foundation everything else runs on.
That's where we are with AI right now. The models get all the attention. The demos get all the clicks. But the real value? It's in the boring stuff.
The cost tracking. The deployment tools. The guardrails. The analytics.
These are the companies that'll survive the reckoning. Because when enterprises finally start asking, "Where's my ROI?" they'll need tools to answer that question.
And Luck knows this. She's not just watching the model race. She's watching the infrastructure build-out. And she's betting on the right side.
Because here's the thing: models will commoditize. They always do. But infrastructure? That sticks around.
That's where the real money is.
The Real Test: Can AI Agents Replace Your Boss?
Here's the question nobody wants to ask: if your AI agent can do your job better than your boss, will you keep it?
That's the threshold.
Not "Can it write an email?" Not "Can it summarize a meeting?" But: "Can it make a decision I'd be afraid to make myself?"
That's what personal AI agents need to become—not assistants, but delegates.
Think about it. Your assistant doesn't just take notes. They schedule your calendar. They push back on meetings. They filter your emails. They know when to say "no" for you.
An AI agent that does that? That's not a tool. That's a co-pilot. And that's what Luck's talking about when she says "magic moment." It's not the demo. It's the moment you stop thinking about the agent at all.
You don't ask Siri to send a text. You say, "Hey Siri, tell Mark I'm running late." And you don't think about the API call or the latency or the token cost.
That's the future. Not the hype. Not the ROI spreadsheet. That moment of total, effortless trust.
And we're not even close.
I keep thinking about how most people interact with technology today. It's all friction. You open an app. You navigate menus. You fill out forms. You wait for responses.
The best technology disappears. It becomes invisible. You don't think about it because it just... works.
That's what AI agents need to become. Not another tool in your stack. Not another tab in your browser. Something that just... works.
Something you trust enough to delegate real decisions to.
That's the bar. And we're nowhere near it yet.
But Luck thinks we'll get there. Not because the tech is ready. Because the need is real.
And need always finds a way.
The Only Path Forward: Discipline, Not Hype
The truth? We're in the middle of an AI winter. Not because the tech failed. Because we got lazy.
We thought innovation meant using more AI. It doesn't. It means using AI better.
The companies that win will be the ones that treat AI like capital, not candy.
Set budgets. Track spend. Measure outcomes. Kill the projects that don't deliver. Build guardrails. Empower engineers—not just the execs.
Tiffany Luck didn't become one of the most respected VCs in AI by betting on hype. She bet on discipline.
And that's the lesson here.
Personal AI agents won't click because they're smarter.
They'll click because we finally stopped treating them like toys.
And started treating them like teammates.
I keep thinking about how most companies approach new technology. They throw money at it. They hire people around it. They build culture around it.
And then they wonder why nothing works.
Because innovation isn't about spending more. It's about spending better.
It's about knowing what works and doubling down on that. It's about killing the projects that don't deliver. It's about building guardrails, not just launching features.
That's what Luck is saying. That's what the market is telling us.
The hype cycle is over. The reckoning is here. And the only way through it? Discipline.
Not vision. Not ambition. Just... discipline.
Set budgets. Track spend. Measure outcomes. Kill the losers. Double down on the winners.
That's how you get through an AI winter.
And that's how you come out the other side stronger.
Because when the smoke clears, only the real value remains. Everything else? Just noise.