Three years after ChatGPT lit the fuse on the generative AI boom, most big companies are still waiting for payback.
An MIT survey in August found that 95% of enterprises weren’t getting a meaningful return on their AI investments. That hasn’t stopped investors from making the same prediction for the third year in a row: next year is when enterprise AI finally goes from toy to toolkit.
TechCrunch spoke with 24 enterprise-focused VCs. Almost all of them say 2026 is the year AI becomes a real line item, not just an experiment. The difference this time: they’re a lot more specific about what has to change.
From hype to harder problems
The easy phase of the AI cycle is over.
“Enterprises are realizing that LLMs are not a silver bullet for most problems,” said Kirby Winfield, founding general partner at Ascend. His example: just because Starbucks can use Claude to write its own CRM software, that doesn’t mean it should.
Winfield expects attention to move to the unglamorous pieces that actually make AI work at scale: custom models, fine-tuning, evals, observability, orchestration and data sovereignty.
Molly Alter, partner at Northzone, thinks a lot of enterprise AI startups are about to discover they’re really services businesses. She expects “a subset of enterprise AI companies” to shift from product to AI consulting, replicating the forward-deployed engineer model and building bespoke use cases on top of their own platforms. In other words, many niche AI product companies will turn into generalist AI implementers.
Where VCs see real adoption in 2026
A few themes cut across the investor answers:
Voice as the primary interface. “Voice is a far more natural, efficient, and expressive way for people to communicate with each other and with machines,” said Marcie Vu, partner at Greycroft. She’s “eager to see how builders reimagine products, experiences, and interfaces with voice as the primary mode of interaction with intelligence.”
AI in the physical world. Alexa von Tobel, founder and managing partner at Inspired Capital, argues that 2026 will be the year AI reshapes infrastructure, manufacturing and climate monitoring. She frames it as a shift “from a reactive world to a predictive one,” where physical systems can sense problems before they become failures.
Labs moving up the stack. Lonne Jaffe, managing director at Insight Partners, is watching how frontier model labs behave. Many people assumed they’d just ship models and let others build the apps. Jaffe no longer buys that. He expects labs to ship more turnkey applications directly into production in finance, law, healthcare and education.
Quantum’s ‘momentum’ year. If quantum computing had a tagline for 2026, OpenOcean general partner Tom Henriksson would pick “momentum.” He sees trust in quantum advantage building as companies publish roadmaps. But he warns not to expect big software breakthroughs yet; hardware still has to catch up.
The boring stuff is big business
Behind the scenes, a lot of venture dollars are headed into infrastructure and efficiency.
Salesforce Ventures principal Emily Zhao says the firm is targeting two frontiers: “AI entering the physical world and the next evolution of model research.”
Microsoft’s venture arm M12 is leaning into what managing partner Michael Stewart calls future “token factory” technology — everything inside the walls of the data center, from cooling and compute to memory and networking.
NEA partner Aaron Jacobson puts it bluntly: “We are at the limit of humanity’s ability to generate enough energy to feed power-hungry GPUs.” He’s looking for software and hardware that can drive breakthroughs in performance per watt, whether through better GPU management, more efficient AI chips, next-gen networking like optical links, or rethinking thermal load in AI systems and data centers.
What an AI moat actually looks like
If you’re an AI startup, investors don’t believe you just because you have a clever prompt or a slightly better benchmark.
“A moat in AI is less about the model itself and more about economics and integration,” said Rob Biederman, managing partner at Asymmetric Capital Partners. He looks for startups that are deeply embedded in workflows, sit on proprietary or continuously improving data, and create real switching costs or cost advantages.
Wing Venture Capital partner Jake Flomenberg is skeptical of moats “built purely on model performance or prompting,” because those advantages “erode in months.” His killer question: if OpenAI or Anthropic ships a model tomorrow that’s 10x better, does this company still have a reason to exist?
For Northzone’s Alter, it’s easier to build a durable moat in vertical categories like manufacturing, construction, health or legal, where data looks similar across customers. That allows real data moats, where every new customer or interaction makes the product better. She also calls out “workflow moats” — defensibility that comes from deeply understanding how work moves from A to B in a given industry.
Snowflake Ventures director Harsha Kapre takes a similar line: strong AI moats come from how effectively a startup can turn an enterprise’s existing data into better decisions, workflows and customer experiences, without creating new data silos.
Will 2026 finally deliver enterprise value?
Most of the investors TechCrunch surveyed think the answer is yes — but not in a hockey-stick way.
Winfield expects enterprises to abandon “random experiments with dozens of solutions” in favor of fewer, deeper deployments.
Norwest Venture Partners’ Scott Beechuk sees 2025 as the infrastructure year and 2026 as the test: “If last year was about laying the infrastructure for AI, 2026 is when we begin to see whether the application layer can turn that investment into real value.”
Exceptional Capital founder Marell Evans agrees value is coming, but calls it “still incremental.” He thinks solving simulation-to-reality training will unlock new opportunities across industries.
Andreessen Horowitz general partner Jennifer Li pushes back on the idea that enterprises aren’t seeing any return. Her litmus test: “Ask any software engineer if they would ever want to go back to the dark ages before they had AI coding tools. Unlikely.” She argues value is already showing up and “will multiply across organizations next year.”
Black Operator Ventures partner Antonia Dean offers a warning from the boardroom: executives will happily say they’re increasing AI investments to justify cuts elsewhere. “AI will become the scapegoat for executives looking to cover for past mistakes,” she said.
Budgets will rise — and consolidate
On one point, there’s near-consensus: AI spend is going up, but not for everyone.
Sapphire managing director Rajeev Dham expects AI budgets to rise, but notes it’s “nuanced.” Some of that growth will come from shifting labor spend into AI tools, or from AI investments that pay for themselves “three to five times over” through top-line ROI.
Biederman predicts budgets will rise only “for a narrow set of AI products that clearly deliver results” and “decline sharply for everything else.” Gordon Ritter, founder and general partner at Emergence Capital, also expects concentrated spending, especially on tools that expand a company’s proprietary advantage rather than simply automate tasks.
Databricks Ventures VP Andrew Ferguson thinks 2026 is the year CIOs clamp down on AI vendor sprawl. Right now, many enterprises run multiple tools for a single use case because it’s cheap to experiment and hard to tell vendors apart, especially in go-to-market workflows. As proof points emerge, he expects companies to “cut out some of the experimentation budget, rationalize overlapping tools, and deploy those savings into the AI technologies that have delivered.”
Maverick Ventures managing director Ryan Isono sees a shift from pilots to production: experimental AI line items will start turning into budgeted spend, and enterprises that tried to build in-house will move to specialized startups once they hit production complexity.
The Series A bar for enterprise AI in 2026
If you’re raising a Series A next year, narrative alone won’t cut it.
“The best companies right now combine two things,” said Wing’s Flomenberg: a compelling ‘why now’ story tied to genAI, and “concrete proof of enterprise adoption.” One to two million dollars in annual recurring revenue is “the baseline,” but what really matters is whether customers see the product as mission-critical. His verdict: “Revenue without narrative is a feature; narrative without traction is vaporware. You need both.”
Insight Partners’ Jaffe wants to see founders going after markets where lower costs from AI actually expand the total addressable market rather than shrinking it away.
Work-Bench co-founder and general partner Jonathan Lehr looks for customers who use the product in day-to-day operations, will take reference calls, and can clearly articulate how it saves time, reduces cost or increases output — and survives security, legal and procurement scrutiny.
M12’s Stewart says pilot revenue no longer carries the same asterisk, given how many options customers are evaluating. But by around six months, he expects investors to see conversions become “the leading part of the story.”
Evans at Exceptional Capital boils it down to execution and traction. His north star: 12-month-plus contracts, delighted users, real technical sophistication and founders who can attract top-tier talent away from hyperscalers.
AI agents: from demo to ‘coworker’
AI agents are coming, but they won’t be fully running enterprises by the end of 2026.
Nnamdi Okike, managing partner and co-founder of 645 Ventures, thinks agents will still be in their initial adoption phase thanks to technical and compliance hurdles — plus a lack of standards for agent-to-agent communication.
Sapphire’s Dham, on the other hand, expects “one universal agent” to emerge inside companies as siloed agents for inbound and outbound sales, support, and product discovery merge into a single system with shared memory and context.
Dean argues the winners will be organizations that treat agents as collaborative augmentation, not a clean human-versus-bot split. She expects “more sophisticated collaboration between humans and agents on complex tasks, with the boundary between their roles continuously evolving.”
NEA’s Jacobson goes so far as to predict that “the majority of knowledge workers will have at least one agentic co-worker they know by name.” Hustle Fund co-founder and general partner Eric Bahn pushes that logic even further: he thinks AI agents could make up the larger share of the enterprise workforce, because proliferating agents is “essentially free and zero marginal cost. So why not grow through bots?”
What’s actually working right now
If you want to know where enterprise AI is already sticking, follow the growth and retention metrics.
Flomenberg says his fastest-growing companies all share a pattern: they spotted workflow or security gaps created by genAI and executed hard to fill them. In security, that means tools for data protection so LLMs can safely touch sensitive data, and agent governance to keep autonomous systems under control. In marketing, it’s entirely new categories like Answer Engine Optimization — getting discovered in AI responses, not just search.
Ferguson sees companies that land with a very focused wedge — a narrow persona or use case — and then expand once they become sticky.
Andreessen Horowitz’s Li says the strongest growth is in companies that help enterprises put AI into production: data extraction and structuring, developer productivity for AI systems, infrastructure for generative media, voice and audio for media, and applications like support centers and call centers.
On retention, Flomenberg points to three traits: being mission-critical (removal breaks production workflows), accumulating proprietary context that’s hard to recreate and solving problems that intensify as customers deploy more AI.
Henriksson notes that “serious enterprise software providers, especially those enhanced with AI” show the best retention, citing Operations1, which digitizes employee-led production processes end-to-end, as an example of software that embeds deeply in operations.
Stewart highlights startups in data tooling and vertical AI apps that pair strong products with forward-deployed teams for customer success and quality — a model he says “has been adopted by all leading startups in those markets.”
Lehr points to AuthZed for authorization and policy, and to Courier Health and GovWell as systems of record and orchestration layers in healthcare and government. Once those are embedded, ripping them out is extremely costly.
Put it all together, and the 2026 enterprise AI story looks less like another hype wave and more like a stress test. Budgets will rise, but vendors will be cut. The tools that survive will be the ones that sit closest to real workflows, real data and real infrastructure — and that can prove, not just promise, that they’re worth the spend.
