Lately, there’s a new sermon being preached in the SaaS wilderness. You’ve seen the LinkedIn posts and the glossy product pages: billing vendors claiming they’ve built "ML models" to forecast your usage-based revenue or "simulations" to tell you exactly how to price your AI app.
If you’ve been evaluating modern usage-based billing platforms, you’ve probably seen the pitch:
“Use historical usage data and ML to simulate pricing.”
“Forecast revenue with confidence.”
“Let AI recommend how you should price your product.”
It sounds reasonable. It sounds modern. It sounds inevitable.
And it’s mostly wrong.
I’m not saying this as someone allergic to ML. I spent over five years building AIOps and ML-driven products in the observability and SRE space at SignifAI and New Relic. I’ve shipped anomaly detection, forecasting, root cause analysis, and decision-support systems that ran at scale in production environments.
I’ve seen what ML is good at.
I’ve also seen where it systematically fails.
Pricing sits squarely in the second category.
If you’re an engineer or product leader evaluating these "predictive" billing tools, here is the technical reality check you won't get from a sales demo.
1. The Fragility of the Time-Series Trap
Usage-based revenue is, at its core, a time-series forecasting problem. The best OG usage-based revenue companies like Snowflake, Datadog and AWS know this. This is why they lean more than 80% of their revenue to pre-committed contracts. In the observability world, we also learned the hard way that most time-series algorithms are incredibly fragile.
Standard models (like ARIMA or even more modern LSTMs) rely on Stationarity—the assumption that the statistical properties of the data (mean, variance) remain constant over time. But B2B usage data is the opposite of stationary. It is plagued by:
- Structural Breaks: A single deployment of a more efficient LLM can drop your "token spend" by 40% overnight.
- Anomalies vs. Trends: ML models struggle to distinguish between a "planned spike" (a customer running a massive one-time migration) and a "real trend" (organic growth).
Claiming you can forecast revenue with just "6 months of billing data" is a statistical pipe dream. Academic research suggests that for meaningful financial forecasting, you often need 2-3 years of clean data to account for true seasonality and market cycles.
In observability, this assumption was already fragile.
Even there, academic and industry research repeatedly shows:
- Time-series forecasting degrades sharply under concept drift
- Anomaly detection struggles with seasonality, interventions, and regime shifts
- Models trained on “normal behavior” fail precisely when humans care most
This is well-documented in work like Gama et al., A Survey on Concept Drift Adaptation, Laptev et al., Generic and Scalable Framework for Automated Time-series Anomaly Detection, and Chandola et al., Anomaly Detection: A Survey.
And that’s for systems where:
- The metrics are machine-generated
- The feedback loops are indirect
- The system behavior is at least governed by physics and software constraints
Pricing is none of those things.
2. Pricing is not a time-series problem - It’s a socio-technical system with adversarial dynamics
Usage-based billing vendors often frame pricing as:
“You have usage data. We can model it. We can simulate changes.”
This framing is deeply flawed.
Pricing is not a function of usage alone. It is a second-order human system that includes:
- Buyer psychology
- Budget ownership and approval flows
- Perceived value vs. perceived risk
- Competitive anchoring
- Contractual constraints
- Procurement behavior
- Internal politics on the customer side
In ML terms: pricing introduces strategic agents into the loop.
The moment you change pricing:
- Users change behavior
- Buyers renegotiate
- Champions lose leverage
- Competitors react
- Usage distributions shift
This is the Lucas Critique applied to SaaS pricing:
models trained on historical behavior become invalid once the policy changes.
No amount of gradient boosting fixes that.
This is the Lucas Critique applied to SaaS pricing:
models trained on historical behavior become invalid once the policy changes. No amount of gradient boosting fixes that.
3. High Cardinality: The Silent Model Killer
In observability, we fear "high cardinality"—having too many unique dimensions (like ContainerIDs or IP addresses). Billing data is the ultimate high-cardinality nightmare.
Pricing isn't a single-dimension problem like CPU or memory. It’s influenced by:
- Human Intent: Did the user stop using the tool because of a bug, a budget cut, or because they moved to a competitor?
- External Variables: Competitor price drops, changes in LLM provider costs, or shifting ICPs.
When you feed high-cardinality, "noisy" behavioral data into a model without business context, you get Complexity Creep. The model generates "fancy, granular, and wrong" projections because it's trying to find signals in what is essentially white noise.
4. The "Backtesting" Sin: Simulation $\neq$ Reality
Some vendors claim they can "simulate" a price change by replaying your historical usage through a new pricing model. In finance, this is called Backtesting, and it is notoriously prone to Selection Bias and Hindsight Bias.
A simulation can tell you what you would have made last year if you charged $0.01 more per token. It cannot tell you how many customers would have churned or downgraded because of that change.
"Pricing strategy happens in the minds of your customers, not in a database of past invoices."
Research into Concept Drift shows that when the underlying environment changes (like a price hike), the historical data becomes irrelevant. You are essentially trying to drive a car forward while looking exclusively through the rearview mirror.
Here’s the core logical flaw:
You are using data generated under Pricing Model A
to predict behavior under Pricing Model B.
In observability, this would be like:
- Training an anomaly detector on traffic patterns
- Then changing your entire deployment architecture
- And expecting the model to still be valid
We already know this fails.
Academic work on policy-dependent data, counterfactual inference, and off-policy evaluation exists for a reason. Most real-world systems cannot reliably estimate counterfactual outcomes without controlled experiments and strong assumptions.
Pricing simulations rarely meet those requirements.
They implicitly assume:
- User elasticity is stable
- Demand curves are smooth
- Behavioral shifts are linear
- Buyers act independently
None of these assumptions hold in enterprise SaaS or AI products.
6. You are not simulating the system. You are replaying the past.
What most pricing “simulations” actually do is replay historical usage, apply a new pricing formula, sum the result, present a range (optimistic / realistic / pessimistic)
This is not simulation.
This is accounting with uncertainty bands.
It ignores:
- Feature gating effects
- Behavioral suppression or amplification
- Strategic downgrade behavior
- Contract renegotiation
- Plan cannibalization
- Long-term trust erosion
In observability, we learned this the hard way:
models that look “accurate” on dashboards often fail catastrophically when used for decision-making.
6.1 ML breaks hardest when humans react to the output
There’s a deep irony here. The more “actionable” the ML output becomes, the more fragile it is.
In AIOps, this shows up as alert fatigue, automation-induced outages, feedback loops that amplify noise
In pricing, it’s worse.
The model doesn’t just observe the system.
It changes incentives.
Once customers realize pricing is algorithmically tuned:
- They game usage
- They restructure consumption
- They push for caps and commitments
- They move spend off-platform
Your model is now chasing a moving target that actively resists being predicted.
No training dataset fixes adversarial dynamics.
6.2 The dangerous part is not the tooling. It’s the false confidence.
The real risk isn’t that someone uses ML dashboards.
It’s that leaders:
- Delegate pricing decisions to models
- Replace judgment with pseudo-science
- Overfit to historical comfort
- Miss the structural risks hiding underneath
In observability, we learned that confidence is the most expensive failure mode.
Pricing is no different.
7. Monetization is Infrastructure, Not Strategy. Pricing is a product decision, not a data science problem
The job of a monetization infra vendor is to get you live, manage change, and handle the brutal complexity of implementation safely. That is a hard, honorable engineering problem.
The job of a monetization infra vendor is to get you live, manage change, and handle the brutal complexity of implementation safely. That is a hard, honorable engineering problem.
But telling a Fortune 500 company or a high-growth AI startup "how to price"? That’s a different profession entirely. It requires deep qualitative work: understanding value perception, ROI, and competitive positioning—variables that don't exist in a SQL table.
When a billing provider claims their software can "spit out the right price," they aren't selling software—they’re selling snake oil.
The Bottom Line
I love ML. I’ve built my career on it. But the fastest way to misuse ML is to apply it where it looks mathematically convenient and feels strategically reassuring.Pricing is one of those traps.
If your pricing strategy depends on an algorithm trained on yesterday’s behavior to predict tomorrow’s incentives, the problem isn’t the model.
It’s the premise.
Don't let the "AI-powered" buzzwords distract you. In B2B SaaS, the graveyard is full of pricing analytics tools that promised to automate strategy but failed because they lacked the "full breath of context" that influences human decisions.
If you want to price your product correctly, talk to your customers. If you want to bill them accurately, find a vendor that focuses on composable infrastructure, not "magic" simulations.
.png)
