Madrona’s Enterprise AI Survey Turns Startup ARR Into a Moving Target

For most of the SaaS era, annual recurring revenue was a promise as much as a metric. A signed enterprise contract meant a vendor could count on that revenue again next year, and probably the year after, because switching business software was slow and expensive. New research from Madrona Venture Group suggests that promise does not hold for artificial intelligence. In a survey of 150 senior enterprise decision-makers, the venture firm found that companies are not just buying AI cautiously: they are re-evaluating what they already bought on a rolling basis, treating even signed AI contracts less like commitments and more like open auditions.

The findings were published August 13, 2026 in a report titled “Harnessing Enterprise Value: The ROI of AI,” written by Madrona’s Rolanda Fu, Matt McIlwain, and Vivek Ramaswami. They land at an awkward moment for AI startups that have spent the past two years boasting about how fast their revenue grows. As TechCrunch’s Julie Bort reported on September 3, 2026, enterprise contracts are exactly what let some AI startups claim they went from zero to $10 million in annual recurring revenue within three months. Madrona’s data suggests that speed cuts both ways, even as competing model providers battle for the same enterprise wallet share.

The Budget Is There. The Certainty Is Not.

Enterprise appetite for AI spending has not slowed down. According to Madrona’s survey, 74 percent of enterprise decision-makers plan to expand their AI budgets over the next 12 months, and the rest plan to hold spending steady rather than cut it. Nearly half of enterprises now carry a dedicated, net-new AI line item in their budgets, a sign that AI spending has moved from experimental innovation money into its own permanent category.

What Madrona calls “the number that should reorganize how founders think about their go-to-market” is less encouraging: 83 percent of those same enterprises converted fewer than half of their AI pilots into production over the last 12 months. More than a third converted fewer than one in four pilots, and just 1 percent moved more than three quarters of their pilots into production. Madrona frames the gap between committed budget, realized deployment, and enterprise value as the defining dynamic of enterprise AI in 2026.

Pilots Rarely Die in the Demo

It is tempting to assume a stalled pilot is a technology problem: the model was not good enough, or the output was not accurate enough. Madrona’s data says otherwise. “Didn’t work as promised” ranks sixth among the reasons enterprises give for why a pilot failed to convert to production. What actually kills pilots, according to the survey, is everything that surrounds the product: integration complexity ranks first as a cause of failure, followed by security, privacy, and compliance requirements, then ROI scrutiny.

The same survey found that strong end-user adoption, not a formal ROI case, is what tips a pilot into a production deployment. Sixty-five percent of buyers rank end-user adoption and feedback among their top three factors for scaling a pilot, putting it slightly ahead of executive sponsorship and seamless integration. A clear, quantifiable ROI case ranks fourth. Madrona’s reading of that order is that by the time a formal ROI analysis happens, the real decision has usually already been made informally, once a practitioner falls for the tool and an executive agrees to back it.

Fast In, Fast Out

Sales cycles for enterprise AI have compressed sharply compared to the traditional SaaS playbook. Madrona’s survey found that 52 percent of AI deals close in under six months from the first meeting to a signed contract, a pace that would have been unusual for enterprise software a decade ago. The back end of the relationship is where AI breaks from precedent. Seventy-seven percent of enterprises re-evaluate their AI vendors at least every six months, and 29 percent do it on a rolling, continuous basis rather than at a fixed interval.

“Annual and multi-year contracts used to provide a moat of inertia,” Madrona’s report states. “In enterprise AI, switching costs are lower and the re-evaluation cadence is relentless.” The firm’s advice to founders is blunt: landing the deal is the beginning of the relationship, not the end of it, and vendors need continuous, visible value built into the product itself to survive renewal season, because a signed contract no longer buys the years of inattention that SaaS deals used to.

Part of what lowers that switching cost is where AI tools enter the enterprise in the first place. Madrona found that discovery is overwhelmingly bottom-up: 41 percent of enterprises primarily find AI tools through internal technical-team research rather than a formal sales process. A tool that lands through a practitioner’s own experimentation, instead of a boardroom pitch and a procurement cycle, has fewer institutional habits protecting it once a better option shows up.

Why Buyers Want Outcomes, Not Tokens

The pricing side of the same puzzle is the subject of a separate, complementary piece of research from Andreessen Horowitz, published August 27, 2026 by a16z partners Tugce Erten and Sarah Wang. Their argument starts with a specific complaint: many AI applications still price their product the way the underlying model provider prices raw inference, in tokens, and that choice quietly works against the vendor.

“When an application prices that work in tokens, it imports the model provider’s cost structure into the relationship with the customer and anchors the product’s value to a unit whose cost keeps falling,” Erten and Wang write. Token-based pricing, in their framing, trains a customer to mentally compare a purpose-built application to raw compute, a comparison the application will usually lose as model providers keep cutting prices to stay competitive. It also gives away the value of the data, workflow, and orchestration the application layers on top of the model, and it exposes customers to a bill they cannot easily forecast: predicting context length, retrieval volume, retries, or reasoning time is a different skill than budgeting for a known number of completed conversations or resolved tickets.

The firm’s survey of 50 technical AI buyers backs the argument up: 27 said they preferred credits tied to a recognizable unit of work, such as a completed account brief or an implemented code change, while only 14 preferred straightforward token pricing. “Pricing around the recognizable work does the opposite,” Erten and Wang write of the alternative to token metering. “It makes value legible, spend forecastable, and product improvements economically valuable to both sides.”

Madrona’s own survey data lines up with that finding. It found that outcome-based pricing is the model enterprise buyers say they most prefer, yet the one they least commonly encounter in the market. Usage-based pricing runs the opposite direction: nearly half of buyers say it is what they primarily encounter, even though fewer of them actually prefer it. Read together, the two reports describe a market where vendors are billing for the wrong unit, in a way that leaves their own revenue harder to defend once renewal time arrives, the same fragmented, token-denominated spending that infrastructure providers have started building dedicated tooling to track in the first place.

What Fragile ARR Means for AI Startups

None of this is happening in a vacuum. Enterprise AI budgets are expanding against a backdrop where a large share of AI initiatives still are not proving their worth. In 2025, MIT’s NANDA initiative reported that roughly 95 percent of enterprise generative AI pilots were failing to produce a measurable business return, a widely cited finding that has shaped how cautiously enterprises now treat new AI purchases. Madrona’s newer data suggests the caution does not disappear once a pilot clears that bar and reaches production. Getting a deployment approved no longer buys a vendor the multi-year runway that a comparable SaaS win once did.

For AI startups reporting fast, headline-grabbing ARR growth, the practical implication is that a signed logo is a weaker predictor of durable revenue than it used to be. A contract with a customer that discovered the product through its own engineers, rather than a formal procurement process, behaves less like recurring revenue and more like a subscription that can be canceled with a few months’ notice. Madrona’s own advice to founders reflects that reality: build continuous, visible value demonstration into the product itself, not just into the sales relationship, and price around the unit of work a buyer already understands rather than the token count a model provider happens to bill in.

For enterprise buyers, the same dynamics look more like leverage. A rolling re-evaluation cadence, bottom-up tool discovery, and a preference for outcome-based pricing all shift negotiating power toward the customer, at the cost of more ongoing vendor management overhead than the old model of signing a contract and revisiting it in three years. Whether that trade is worth it may depend on how quickly AI vendors adapt their products and pricing to a market that no longer rewards signing the deal and moving on.