Silicon Valley spent much of the past year celebrating AI usage at almost any cost. Companies pushed employees to ask more, generate more and automate more, often with little scrutiny over whether the extra spend produced real business value. That era is now colliding with a harder question: what, exactly, is AI worth?
That question sits at the center of NEA partner Tiffany Luck’s current view of the market. A veteran investor who once helped convince businesses that e-commerce would reshape commerce, Luck now sees AI moving through a similar adoption curve — but with a sharper focus on measurable outcomes. Speaking on TechCrunch’s Equity podcast, she argued that the industry is shifting away from open-ended experimentation and toward a more disciplined search for return on investment.
The change is visible across the enterprise. Reports of fast-rising AI bills, reduced license counts and internal efficiency pushes suggest that many companies are no longer content to simply “use more AI.” They want proof that the tools save time, improve quality or unlock new revenue. Startups, in turn, are adapting by selling infrastructure that helps buyers track usage, compare model performance and deploy AI more selectively.
Luck’s broader thesis is that the AI market is not just about the model layer — the foundation models at the top of the stack — but about value creation at every level, from applications and workflow tools to deployment services and measurement platforms. In her view, that widening opportunity is one reason AI remains one of the most active areas in venture capital even as the market becomes more demanding.
The end of “tokenmaxxing” and the start of the ROI reckoning
Early in the AI boom, some companies treated model usage like a badge of honor. Teams were encouraged to maximize token consumption, experiment broadly and discover where AI could fit into daily work. But the economics of that approach have become harder to ignore. As budgets tighten and enterprises seek accountability, the focus is shifting from volume to value.
In practical terms, that means firms are asking more detailed questions. Which workflows actually benefit from generative AI? Where does a premium model justify its cost? When is a smaller or cheaper model good enough? And how can finance teams tell whether AI spending is producing measurable productivity gains?
That shift is not unique to any one company. It reflects a broader phase change in how enterprises adopt technology. New tools often enter the workplace through experimentation, only to be followed by scrutiny once the novelty wears off and budgets face pressure. AI is now at that inflection point.
Luck’s view, as discussed on the podcast, is that the market has moved from enthusiastic consumption of AI capacity toward a harder accounting of what those systems deliver in return.
This “ROI reckoning” is forcing both buyers and sellers to mature. Vendors can no longer rely on vague promises about transformation. They need to show how AI changes specific metrics, from customer response times to employee output to support deflection rates. Buyers, meanwhile, are learning that model choice, workflow design and internal adoption patterns all affect the economics.
Why AI spending is getting more complicated
One reason the math is getting trickier is that AI products often spread horizontally across an organization. A single team may use a chatbot for support, an assistant for coding, another tool for summarization and a separate platform for internal search. Each use case may seem modest on its own, but together they can create substantial costs.
Some firms have responded by narrowing access. Others have renegotiated contracts, shifted workloads to lower-cost models or limited usage in departments that were not delivering enough business impact. These changes are not a sign that AI is failing; they are evidence that the first wave of exuberance is giving way to operational discipline.
Luck’s framing suggests that this moment should be understood less as a backlash than as a normalization. In most technology cycles, spending becomes more targeted after the early rush. The same pattern may now be unfolding in AI, where the companies that survive the next phase are likely to be those that can prove repeatable value.
Forward-deployed engineers become a Trojan horse for adoption
Another theme Luck highlighted is the growing importance of forward-deployed engineers, a role that has become especially popular among AI startups selling to enterprise customers. These engineers work directly with clients to integrate products into real workflows, customize implementations and troubleshoot deployment issues.
The reason they matter is simple: AI adoption rarely happens just because a company buys a license. Real deployment requires systems to fit existing processes, data environments and organizational habits. Forward-deployed teams help bridge that gap, often becoming the first point of contact between a company’s abstract interest in AI and its actual use in production.
Luck described these engineers as a kind of “Trojan horse” for AI adoption, because they often open the door to much broader implementation. Once a vendor proves value inside one team or one workflow, the product can spread to adjacent departments and become part of the company’s operating rhythm.
This model also explains why services-heavy startups have become more common in AI. Unlike software that can be handed over and left to run on its own, AI products often need substantial early support to prove themselves. That dynamic has created opportunity for vendors willing to invest in implementation as much as in product development.
From sales motion to product strategy
What looks like customer support is often part of a broader strategy. Companies that place engineers close to users can learn faster about what breaks, what matters and what customers will actually pay for. In that sense, forward deployment is not just a sales tactic. It is a product feedback loop.
For startups, this can mean faster iteration and stronger retention. For buyers, it can reduce the risk of paying for tools that never move beyond pilot mode. The challenge, of course, is that this hands-on model is expensive and hard to scale. It favors companies that can balance high-touch deployment with long-term software margins.
Enterprises are mixing models instead of picking a winner
Luck also pointed to another important change in enterprise AI strategy: many companies are no longer betting on a single provider. Instead, they are mixing and matching models depending on the task.
That behavior reflects both technical and financial realities. Different models can excel at different jobs, and the best choice for customer support may not be the best choice for coding, search or content generation. At the same time, price-performance differences can be significant enough that a company might reserve its most powerful model for premium tasks while relying on cheaper options elsewhere.
This multi-model approach reduces dependence on any one vendor and gives buyers more leverage. It also creates new opportunities for startups building orchestration layers, evaluation tools and governance systems that help companies route requests intelligently across providers.
For the model makers, this is a sign that competition is becoming more nuanced. Winning the entire enterprise may matter less than becoming indispensable in the right slice of the workflow. That in turn helps explain why the AI stack continues to expand into adjacent layers instead of consolidating around a few dominant foundation-model players.
AI value is spreading across the stack
Luck’s outlook is notably broader than the idea that the biggest companies in AI will automatically be the model providers. She believes value is being created in many parts of the stack, including infrastructure, deployment, safety, workflow automation, evaluation and customer-facing applications.
That perspective matters for investors and founders alike. If value accrues only to the model layer, then the market is likely to concentrate quickly. But if each layer of the ecosystem can produce durable companies, the opportunity set widens dramatically.
That is one reason AI has attracted such sustained venture attention. Even as headlines focus on giant model releases or valuations, the practical business opportunities increasingly lie in the tools that help companies put those models to work. Startups are building everything from agent platforms and model routers to observability dashboards and cost controls.
In that sense, the market is beginning to resemble earlier software cycles. Infrastructure companies support application companies; services help deployment; analytics tools help buyers understand what they are getting; and end-user applications turn raw capability into something people will actually pay for.
Why the application layer still matters
For consumers and businesses, the most visible value is often created far from the model itself. A well-designed interface, a tailored workflow or a specialized assistant can matter more than the size of the underlying model. That is especially true when users care about speed, reliability and task-specific outcomes more than technical benchmarks.
Luck’s comments suggest that investors are increasingly willing to back those layers. In AI, the old assumption that the largest model equals the strongest company is giving way to a more layered view of the market.
Personal agents remain a compelling, but unfinished, promise
One of the more intriguing topics Luck discussed was the future of personal agents. The idea has become one of the most talked-about visions in consumer AI: software that can act on a user’s behalf, remember preferences, manage tasks and coordinate across apps with minimal supervision.
The appeal is obvious. A truly useful personal agent could reduce friction in everyday life, from scheduling and shopping to travel, email and administrative work. It could also become a new front door to digital services, changing how people interact with software in general.
But the promise is still ahead of the product. Building a personal agent that is both helpful and trustworthy requires more than a fluent chatbot. It requires memory, permissions, context, integration, reliability and safeguards. That makes the category one of the most ambitious in consumer AI — and one of the hardest to get right.
Luck’s interest in “magic moments” in consumer products points to why investors are still watching this space closely. Consumer technologies often break through when they deliver a clear emotional or practical payoff. In AI, that moment may arrive when an agent does something users can feel immediately: saves time, reduces stress or anticipates a need before it becomes a burden.
Consumer AI and the search for delight
Unlike enterprise software, where ROI can be counted in hours saved or workflows automated, consumer AI has to prove itself in more subjective ways. Delight, usefulness and habit formation matter as much as raw efficiency. That is why the best consumer products in AI may not look like enterprise tools at all.
Instead, the winners may be the products that quietly reduce cognitive load or make everyday tasks feel effortless. Luck’s background in e-commerce may also shape her thinking here: many consumer platforms succeed when they collapse friction at the exact moment of decision.
AI IPOs and the market’s next test
The discussion also touched on this year’s AI IPOs, which have become an important signal for the sector’s maturity. Public-market debuts put pressure on high-growth companies to justify their valuation with revenue quality, margin structure and sustainable demand rather than sheer excitement alone.
For the AI industry, IPO performance is more than a Wall Street event. It is a test of whether the market believes AI companies are building durable businesses or simply riding a speculative wave. Investors will be watching for evidence of customer retention, repeatable sales motions and a path to profitability that can survive beyond the first wave of enthusiasm.
That scrutiny is arriving just as the AI category is fragmenting into multiple submarkets. Some companies sell infrastructure, others offer workplace productivity tools, others focus on consumer agents, and still others aim to help enterprises govern and measure AI usage. Public investors may be more selective than private ones, which could sharpen the pressure on startups to define exactly what kind of business they are building.
| Theme | What is changing | Why it matters |
|---|---|---|
| AI spending | From broad experimentation to tighter ROI tracking | Buyers want proof of value before expanding budgets |
| Deployment | Forward-deployed engineers are increasingly common | Hands-on support speeds up enterprise adoption |
| Model strategy | Companies are mixing multiple models instead of standardizing on one | Creates leverage, flexibility and cost optimization |
| Market structure | Value is forming across the stack, not only at the model layer | Expands opportunity for startups beyond frontier labs |
| Consumer AI | Personal agents remain promising but underdeveloped | Could become a major product category if trust and utility improve |
Why investors are still bullish despite the pressure on budgets
The tightening of AI budgets might sound bearish, but investors like Luck do not appear to see it that way. If anything, the discipline may be healthy for the market. When buyers become more selective, the strongest products win faster and weak products disappear sooner.
That dynamic can benefit serious founders. A company that can demonstrate measurable savings, revenue lift or workflow transformation will have an easier time standing out. In this environment, product quality and business impact matter more than narrative alone.
There is also a structural reason for continued optimism. AI is not a single product category but a broad technological shift that can touch nearly every industry. That gives venture investors a wide range of entry points, from enterprise software to consumer applications to tooling and infrastructure.
Luck’s career trajectory reinforces that view. Just as e-commerce evolved from an idea into a foundational layer of the economy, AI may eventually become a default capability embedded across business and consumer software. The question is not whether companies will use AI, but how smartly they will pay for it.
The bigger lesson for startups
The main takeaway from Luck’s perspective is that the AI market has entered a more sophisticated stage. Early adopters may still be eager to experiment, but the rest of the market is demanding evidence. That shifts the burden onto startups to sell outcomes, not just access.
Founders now need to answer a set of more practical questions:
- Does the product save time or money in a measurable way?
- Can customers deploy it without massive internal change management?
- Is the model choice flexible enough to control costs?
- Can the company prove value in production, not just demos?
- Does the product solve a real workflow problem, or merely add another layer of novelty?
Those questions are forcing the market to mature quickly. But they are also creating room for new winners. Startups that can help organizations govern AI use, evaluate model performance, deploy agents safely or integrate AI into core workflows are likely to remain in demand.
In the consumer market, the challenge is different but equally demanding. The winning products will need to feel personal, intuitive and useful enough to become part of daily life. If personal agents eventually cross that threshold, they could become one of the defining interfaces of the AI era.
What comes next
The AI story is no longer just about breakthroughs in model capability. It is increasingly about adoption, cost control, distribution and trust. That is a more mature story, and arguably a more important one.
Luck’s remarks reflect a market that is moving from fascination to accountability. The companies that thrive next may not be the loudest or the most heavily funded, but the ones that can show a clear line from AI usage to business value.
For investors, that means a more selective deployment of capital. For startups, it means sharper product design and clearer economics. For enterprises, it means learning how to use multiple AI tools without letting spend run away. And for consumers, it means waiting for the personal agent promise to become more than a demo.
If the first act of the AI boom was about possibility, the next may be about proof.
Key takeaways from Tiffany Luck’s outlook
- AI adoption is shifting from broad experimentation to hard ROI measurement.
- Enterprises are using multiple models rather than standardizing on a single provider.
- Forward-deployed engineers are emerging as a major enterprise sales and adoption tool.
- Value in AI is being created across the entire stack, not only in foundation models.
- Personal agents remain a major consumer opportunity, but the category is still early.
For Silicon Valley, that means the AI era is not cooling off — it is getting more serious.
