The AI churn wave?
A data-driven investigation into retention for AI apps
👋 Hi, it’s Kyle Poyar and welcome to Growth Unhinged, my weekly newsletter exploring the hidden playbooks behind the fastest-growing startups.
ICYMI: 🎧 Mostly Growth podcast | The State of B2B GTM report
Annual recurring revenue (ARR) isn’t a GAAP metric. Yet it has become the building block of SaaS metrics. And it’s the basis of SaaS valuation multiples.
We think of ARR as (usually) high margin, predictable and growing. Which means SaaS companies are (usually) on track to become highly profitable at scale.
AI startups are increasingly throwing this off. What’s not necessarily recurring revenue but gets treated as such:
Experimental revenue from AI products (not in production)
Contracts with a three month opt-out where 70-80% opt-out
Usage-based or success-based pricing
Professional services from forward-deployed engineers
AI startups force us to rethink ARR. There’s a gross margin problem, as I’ve written about previously. What’s even more concerning is that there are real questions about the durability of AI revenue, which Cassie Young from Primary Venture Partners recently predicted would be an impending “gross retention apocalypse”.
Buyers will scrutinize their AI experiments. AI tourists will move on to the next hot product. Big platforms will swallow dozens of thin LLM-wrappers as models improve.
Durability is an existential question for AI apps. Yet there’s been relatively little objective data on the topic. To change that, I teamed up with ChartMogul where I’m an Analyst-in-Residence.
We scraped the websites of 3,500 software companies, using AI to categorize companies as B2B SaaS, B2C SaaS, or AI-native companies (which might be B2B or B2C). Then we compared gross revenue retention (GRR) and net revenue retention (NRR) rates across these groups. Today I’m sharing what we found, and what steps you can take to protect yourself from what could be an impending churn wave.
AI has a tourist problem
The AI scraping produced a dataset of roughly 2,700 B2B SaaS companies, 600 B2C SaaS companies, and 200 AI-native companies. AI-native companies included a mix of B2B and B2C businesses, as well as startups that sell to both consumers and businesses.
I first looked at annualized retention rates throughout 2025 among the businesses that had reached at least $250k ARR. Why $250k: historical retention data is less reliable for very early-stage startups with few customers coming up for renewal.
What the data shows:
B2B SaaS is relatively sticky. The median NRR was 82%. (The upper quartile was 97%.)
B2C is much less sticky with minimal upsell. The median NRR was only 49%.
AI-native companies had even worse retention than B2C. The median GRR was a mere 40% and median NRR was 48%.
On the optimistic side, AI retention has gotten much better since the beginning of the year. Median GRR jumped from 27% in January to 40% in September. I suspect many of the early AI tourists left; those who remain are more committed or shifting from experimentation to production.
Even still, AI-native companies on the whole don’t look much like the B2B SaaS companies we know and love (and like to value on the basis of ARR). It’s hard to say a business has recurring revenue when 60% of the install base leaves each year.
Easy to buy, easy to cancel?
The AI churn wave won’t hit every AI company the same way.
Lower-priced products have always attracted more churn. This gap is especially pronounced for AI products. The best protection against churn is shifting from consumers and self-serve purchases (<$50 per month deals) to larger B2B purchases.
What the data shows:
AI-native products that sell for >$250 per month see 70% GRR and 85% NRR. This is essentially the same as B2B SaaS.
AI-native products that sell for $50-$249 per month see 45% GRR and 61% NRR. This is 15 points worse than B2B SaaS. (Few B2C products charge beyond $100 per month.)
AI-native products that sell for <$50 per month see just 23% GRR and 32% NRR. This is 20 points worse than either B2B or B2C SaaS.
It’s the curse of the AI wrapper. The downside of being easy to buy is being easy to cancel.
I see a second factor at play here, too. AI self-serve users might be paying customers. They see themselves as trial-ers.
Low-priced SaaS products tend to be freemium (or reverse trial) products where users have plenty of time to try before they buy. With 80%+ gross margins, supporting free users doesn’t need to break the bank. Users upgrade when they’ve consciously decided to buy. And there’s minimal disruption: people know what they’re getting when they purchase SaaS.
The calculus changes with AI products. Each token has a cost. AI products push for conversion either immediately or upon any level of serious usage. Some even impose a daily usage paywall before users need to upgrade, something that would be unheard of in SaaS. People start paying because they’re intrigued by the possibility of what could be, but they haven’t planned for how the products fit into their business long-term.
We need to adjust how we measure retention (and ARR, for that matter) to cleanly delineate experimental spend from serious spend rather than lumping them together. I’d propose we treat serious spend as revenue from customers who’ve paid at least $250. This could be from an annual plan paid upfront, from a deal size above $250, or from a low-spending customer who sticks around.
Retention = durable ARR
AI-native companies have been shattering growth records with many scaling from zero to $100 million in only a year. This is certainly impressive.
But it becomes nearly impossible to sustain hypergrowth when you have a larger and larger install base that walks out the door each month. There’s a term for this: burning through your TAM.
I mapped out the 3,500 software companies in the dataset across two dimensions: net revenue retention (NRR) from September 2025 and year-on-year revenue growth.
What the data shows:
On the whole there’s a very strong correlation between NRR and long-term growth.
Early-stage startups can grow fast (200%+) even with NRR below 40%.
Yet low retention companies are at extremely high risk of growth going in reverse; three times as many are shrinking compared to growing quickly.
Looking at retention and ARR makes the picture even clearer. The companies that make it to $5 million ARR have substantially better gross and net revenue retention rates compared to their early-stage peers. Some of this could be selection bias; a surprisingly large piece is because the breakout winners manage to transform themselves as they grow.
This is already noticeable within AI-native companies, although admittedly the sample size starts getting small (~50 per bucket) which makes the data directional rather than statistically bulletproof. The durable AI-native companies have 2x the GRR and 2.5x the NRR compared to their early-stage peers.
What to do about churn
Over the years I’ve worked with dozens of companies to improve retention. It’s also something I’m actively thinking about as a newsletter writer with a $15 per month premium subscription.
In my experience the best companies treat churn as a business problem, not a customer success problem. How AI companies can prevent churn:
Go after valuable workflows and use cases with real budgets. AI products don’t inherently have high churn; consumer-focused AI wrappers do. Moving upmarket usually means (a) being more deeply embedded in business workflows, (b) being integrated with more tools, and (c) creating more differentiation compared to general-purpose LLMs.
Deliver services in addition to AI. Palantir famously hires forward deployed engineers to prototype and implement AI solutions for high-value prospects. Many AI companies, including OpenAI, are following suit. FDEs close that gap between what’s possible with AI and what customers actually realize. As a16z recently wrote, the FDE model trades margin for moat and tends to be worth it in the long run.
Stop overselling on the first deal. There’s so much promise of what AI could do. It’s usually better to start with a sharp problem, prove value quickly, and then unlock growth opportunities in the future. This also leads to faster sales cycles and dramatically narrows scope for FDEs, allowing them to work with more companies.
Narrow the gap between product delivery and adoption. Brian Balfour from Reforge recently told me that his top priority for 2026 is how do we accelerate product adoption? The pace of AI releases has been dizzying; customers can’t keep up and probably aren’t even aware of what’s changed.
Sell more annual plans. Moving from monthly to annual plans can be controversial. But you’re usually better off because you have more time to impress the customer and the customer is more motivated to implement the product. Data shows the median NRR is 10 to 20 percentage points higher for annual plans compared to monthly ones. How to do it: make annual the default, ask for annual upgrades early (months 2-3), and consider offering a bigger discount (Grammarly’s is 60%).
Excellent post, Kyle. So much to unpack.
1. Did you use each company's 2024 year end ARR as the starting cohort size and then looked at retention each month from Jan to Sep-25?
2. "AI-native products that sell for >$250 per month see 70% GRR and 85% NRR. This is essentially the same as B2B SaaS." --> How are GRR and NRR defined here? Over what period of time?
3. Very surprised to see ~60% of companies with NRR 80-99% have negative to only 24% y/y growth.
Where did you get the initial numbers for revenue for each company? You wrote:
“Then we compared gross revenue retention (GRR) and net revenue retention (NRR) rates”
Where did this data come from?