Early-stage startup valuations are often based on qualitative factors rather than financial modeling, since early-stage companies have limited historical data. Investors evaluate the founding team’s experience, the size and urgency of the market, early signs of product-market fit, technology defensibility, and the competitive landscape. They also benchmark comparable deals and evaluate ownership requirements for their investment.
As a company begins generating predictable revenue, that revenue becomes a powerful anchor for valuation by providing a quantifiable signal of traction. AI startups reach this stage far faster than previous generations of software companies. AI products can scale quickly, onboard users at remarkable speed, and monetize through usage. Even very young companies can show steep revenue curves. This accelerates both growth and valuation.
Annual Recurring Revenue (ARR) remains the core metric for SaaS and AI subscription businesses. ARR captures contracted, predictable revenue expected each year and excludes one-time fees, overages, or experimental spend. ARR matters because it indicates the business’s growth rate, the durability of its revenue base, the depth of customer commitment, and is one cornerstone that sets company valuation.
In this post, we will define experimental revenue, explain why it behaves differently from ARR, and offer a framework to convert revenue into something more recurrent and predictable.
What is ERR?
Experimentation is fundamental to AI development. Training a model is inherently iterative, combining data, algorithms, and configuration in rapid learning loops until the model meets business requirements. GenAI development follows this same pattern. Fine-tuning models with proprietary data demands repeated cycles of testing and refinement.
For decades, most organizations operated with Software 1.0. The widespread adoption of ChatGPT pushed enterprises into Software 2.0, triggering large-scale experimentation across systems, workflows, and products.
This experimentation, combined with massive demand from organizations new to AI, generates large volumes of non-committed, highly volatile revenue. We define this as Experimental Run-Rate Revenue (ERR), which is neither annual nor recurring. Usage often fluctuates dramatically from one month to the next. Many customers experiment for a few months, spike usage during development cycles, and then reduce or eliminate spend once the experiment concludes.
Key Characteristics of ERR
The challenge with early AI startups is that their growth often appears smooth at the top line. Companies can grow 10-15% month over month, or three to five times year over year, creating the illusion of predictable momentum.
New customers continue to sign up every month, reinforcing the perception of steady, reliable growth. At a glance, ERR can resemble consumer-driven revenue patterns. For instance, Netflix subscribers pay monthly without contractual commitments, yet remain for more than four years on average, while Spotify subscribers stay around two years and Disney subscribers about 18 months.
A closer look at customer behavior reveals a very different picture. AI usage is highly non-linear. A customer may start with modest usage for two months, surge for the next two months as they process proprietary data or fine-tune a model, and then drop to zero after the internal project is deprioritized. The average duration of these experimental cycles is often far shorter than 12 months, partly because the category is still young and historical data is limited.
These surges typically reflect short-term development cycles rather than durable production workloads. AI project failure rates remain high across studies, often estimated at 70 to 85%, with some reports placing them as high as 95%. This volatility is a core characteristic of ERR, highlighting why early AI revenue should be interpreted differently from traditional SaaS revenue.
From ERR to ARR
To increase the likelihood that AI projects succeed and that ERR converts to ARR, startups must understand the customer’s use case at a deeper level. Key questions include:
- What business problem is my customer trying to solve?
- What does business success look like and how do you measure it?
- Who is the solution for and what is their level of technical background?
- Who is the customer internal champion?
For example, selling an AI development platform to a company that is building an external-facing agent for its customers tends to have a higher chance of becoming ARR, because the solution ties directly to revenue. In contrast, internal productivity use cases may have vague ROI, bottom-line influence and often stall or get deprioritized.
Founders also need to know where the customer is in their AI journey, including their level of maturity and technical sophistication. Have they already tried building the solution using open-source tools and hit limitations, or are they just beginning to explore? This shapes expectations and risk.
AI projects fail for many reasons: poor data quality, limited resources, weak execution, unrealistic expectations, and unclear business success criteria. The last two can be validated early in the sales cycle. Clarifying the customer’s problem, what they have already attempted, and where they are stuck helps reduce the risk of churn and improves the likelihood that experimental usage becomes sustained usage.
Unlike AI benchmarks that focus solely on accuracy, real project success requires translating model performance into business outcomes. For instance, if a company is trying to detect fraud, larger fraud amounts may be prioritized, and minimizing false positives becomes critical to avoid alienating customers. Success requires defining the business KPIs and thresholds before development begins.
Startups should also segment usage into buckets such as committed versus uncommitted, production versus experimentation, and stable versus volatile. Analyzing retention and conversion rates across these segments allows teams to build weighted models that estimate which portions of ERR are likely to convert to ARR. This disciplined approach gives both founders and investors a clearer picture of true revenue durability.
The Valuation Impact
AI startups generally operate with lower gross margins in the 50-60% range and higher capital expenditures than traditional B2B SaaS companies, which often achieve 70-85% with minimal infrastructure costs. This gap is largely driven by the higher compute, storage, and infrastructure costs associated with training and running AI models, costs that are minimal in classic SaaS models.
In theory, these economics should produce lower valuation multiples than traditional SaaS. In practice, AI startup valuations are soaring. Investor enthusiasm has pushed AI-native companies to command substantial premiums and raise larger rounds more quickly, at elevated multiples.
According to Finro, seed-stage AI companies are valued at an average of 20.8 times revenue while generating about $2M in ARR. Series A and B companies achieve even higher multiples at 39.0 and 31.7 times revenue. These valuations reflect strong momentum, early traction, and market positioning rather than mature financial metrics.
Consider a seed-stage startup generating $1.4M in monthly revenue. Traditional B2B SaaS logic would annualize this to $16.8M ARR ($1.4M × 12) and apply a 15 times multiple, resulting in a $252M valuation.
Segmenting the revenue tells a different story. Suppose $0.4M of the monthly revenue comes from customers in production, while $1M comes from noncommitted experimental usage.
- Production ARR: $0.4M × 12 = $4.8M
- Experimental ARR: $1M × 12 = $12M
Applying differentiated multiples, the valuation becomes:
- Production at 15 times: $4.8M × 15 = $72M
- Experimental at 5 times: $12M × 5 = $60M
The adjusted valuation becomes $132M, far below the $252M headline, illustrating how ERR can distort valuations if not properly segmented.
The market has not fully priced in the lower gross margins and high levels of ERR typical of early-stage AI companies. Over time, as business models mature and margins improve, we expect valuation multiples to normalize.
Conclusion: A Useful but High-Velocity Metric
ERR is becoming a defining metric for the AI era because it reflects the unprecedented pace at which AI startups can grow. When used transparently and paired with segmentation and conversion metrics, ERR provides a more complete view of revenue quality. The combination of ERR and ARR shows both short-term momentum and long-term durability.
As AI adoption accelerates, the ability to correctly interpret and communicate ERR will become increasingly essential for founders, operators, and investors.
