Enterprise AI revenue is looking a lot less durable than startup pitch decks suggest

Enterprise buyers are still spending heavily on AI, but they’re not settling down. That’s the takeaway from fresh research out of Madrona, and it should make anyone staring at startup ARR figures a little uneasy.

The firm surveyed 150 enterprise IT professionals and found that 74% plan to expand AI budgets over the next 12 months. The rest plan to hold steady. That looks like clean growth until you hit the other half of the data: fewer than half of AI pilots make it into full production, and 77% of enterprises reevaluate AI vendors every six months or on a rolling basis.

That last number matters most. AI spending is real. The revenue behind it is brittle. A startup can post fast ARR growth, then lose the same account six months later when a steering committee asks whether the tool is still worth the bill.

Why AI sales don’t stick like SaaS used to

Traditional enterprise software got sticky because it became part of the plumbing. Email, identity, storage, ERP, HR. Once those systems were wired into daily work, ripping them out was painful. Multi-year contracts and integration costs did a lot of the locking in.

AI software has weaker defenses.

A lot of these products sit on top of an external model, a prompt layer, some orchestration code, and maybe retrieval. Useful stack. Also easy to swap. If output quality drops, or a competitor offers the same workflow with a better model and lower cost, buyers can move faster than they could with old-school SaaS.

The organizational side is just as important. Plenty of AI tools are still pilots inside one team, one function, one budget. They’re not buried deep in core systems yet. A customer support tool can be judged quarter by quarter. A legal review assistant can disappear fast if it saves less time than promised or makes too many mistakes.

That makes revenue recognition messy in a very practical way. ARR can look fine on a slide, but if the contract is tied to a pilot, a usage cap, or a six-month review cycle, that revenue is closer to rented time than durable enterprise software.

Why pilots keep stalling

The fact that fewer than half of AI pilots reach production isn’t surprising if you’ve spent time around enterprise systems. They die for familiar reasons, plus a few new ones.

The old reasons are integration, security, procurement, and politics. Production software has to connect to identity systems, logging, data governance, approval workflows, and sometimes compliance controls that were written before anyone said “copilot” without sounding silly. If it can’t fit, the pilot stays in a sandbox.

The new reasons are model behavior and operational risk. Enterprises don’t need a demo that works on clean test cases. They need systems that are predictable, auditable, and cheap enough to run at scale. An AI assistant that looks good in a controlled test can get expensive fast when it hits real traffic, real documents, and real edge cases.

Latency matters too. Three or four seconds might be tolerable for a support team. Twelve seconds, plus human review, usually isn’t. Once a product reaches production, the buyer starts counting every token, every API call, and every minute spent checking output.

That’s where a lot of vendors get into trouble. Their sales pitch is about intelligence. Their customer success problem is economics.

Pricing by tokens is under pressure

Andreessen Horowitz recently surveyed 50 technical AI buyers and found that more than half prefer pricing tied to work produced or outcomes, not token usage. That makes sense. Tokens are a billing unit, not a business result.

Per-token pricing made sense when AI products were closer to infrastructure. If you were buying model access, you paid for compute. Simple enough. But enterprises don’t buy compute for its own sake. They buy documents classified, calls summarized, tickets resolved, leads qualified. They want the bill to line up with work they can actually measure.

That’s why pricing around recognizable work is getting traction. Charge per report processed, per ticket closed, per claim reviewed, per contract summarized. Buyers can connect that to value. It also makes comparisons easier, which is good for customers and probably annoying for startups.

There’s a catch for vendors. Outcome-based pricing gets messy quickly. If the product saves 40% of the time but the customer’s workflow is broken, who gets credit? If the model handles easy cases and punts hard ones to humans, how do you count that? If the system sits inside a process with fuzzy attribution, revenue turns into an argument about measurement.

So the shift is logical, but not tidy. It adds trust requirements on top of technical ones. Vendors need telemetry, attribution, and enough instrumentation to show they’re doing the work they claim.

What founders should be paying attention to

A lot of AI startups are still talking as if enterprise adoption will eventually look like classic SaaS. Sell a pilot, win production, land a multi-year deal, expand over time. That path exists. It’s just not the default anymore.

The Madrona data points to something more volatile. Enterprises are trying more vendors, cycling faster, and revisiting choices with a frequency that would have looked odd in the old SaaS world. That lets startups grow quickly at the top of the funnel while building weaker revenue underneath.

For investors, that changes how ARR should be read. A 30% or 50% growth rate means less if the contracts are short, usage is erratic, and renewals depend on a quarterly value review. Gross revenue retention matters more than headline ARR. So do implementation depth, cost-to-serve, and whether the product sits inside a workflow or just hangs around it.

For founders, the lesson is sharper. If your AI product can be swapped out after one budget review, you don’t have a platform. You have a test case.

The companies that make it through this phase will probably do a few things better than their competitors:

• Integrate into core systems, not just side workflows
• Measure outcomes in customer terms, not model terms
• Keep inference costs under control as usage scales
• Make human review part of the product, not an afterthought
• Build auditability into the system from day one

That last one matters more than it used to. In enterprise AI, being able to explain what happened, trace a result back to inputs, and show logs or citations can be the difference between a pilot and a production contract. It’s not glamorous. It closes deals.

AI spending is real. Loyalty isn’t

None of this means enterprise AI is stalling. It clearly isn’t. Budgets are rising. IDC says companies are on pace to spend $4.25 trillion on technology in 2026, with AI driving most of it. Enterprises are still buying, still testing, still expanding where the math works.

But the buying pattern is unstable. The market rewards experimentation faster than it rewards permanence. That’s good news for vendors in the short term and a headache for anyone pretending this is just another SaaS cycle with better branding.

For technical teams, the takeaway is simple. If you’re building AI products for enterprise buyers, assume every win is provisional. Procurement will ask harder questions after deployment than it did during the pilot. Pricing will get challenged. Usage won’t automatically turn into retention.

That’s not a reason to avoid the market. It’s a reason to build like the customer can leave next quarter. Because right now, a lot of them can.