Modern Pricing & UBB
February 28, 2026 ubb

Token, Task, or Outcome: Pick Your AI Agent Pricing Poison

The hottest product category in SaaS right now is AI agents. Systems that browse the web, write code, draft emails, book meetings, scrape data, and occasionally hallucinate a refund your customer never asked for. All while you're asleep.

And nobody — not even the people building them — has a clean answer to the question: how do you charge for this?

The old model was simple. User logs in, user uses product, user pays for seat. Done. Static SaaS was a vending machine: you put money in, you got access out. Predictable for the vendor. Predictable for the CFO. Boring in the best possible way.

AI agents broke the vending machine. The cost to serve an agent varies by orders of magnitude based on what it's actually doing. A simple "summarize this email" agent costs fractions of a cent. An agent that orchestrates 12 API calls, reasons through tool outputs, retries twice, and writes a 3-page report? That's a different beast entirely. As Orb puts it, agent costs live "per call, token, and event instead of per seat" — and static tiers simply miss the value.

The Three Schools (All Flawed)

School 1: Charge per token. This is the OpenAI default. Every token in, every token out, you pay. It's precise. It's transparent. And it means absolutely nothing to a VP of Operations trying to figure out if $4,200 of inference last month was a good deal. "You consumed 42 million tokens" is not a business outcome. It's a diagnostic. Charging customers in tokens is like charging for a plane ticket in milliliters of jet fuel. Technically accurate. Commercially useless.

School 2: Charge per task. Pick a unit of work — "per email drafted," "per document summarized," "per lead enriched" — and charge for it. This is cleaner. Customers understand it. But now you have to define what a "task" is, and that definition will be litigated in every enterprise contract you sign. Does a failed task count? A partial task? A retried task that succeeded on the second attempt? Your pricing page says $0.05 per document. Your customer's contract says something different and their lawyer has opinions.

School 3: Charge per outcome. The dream scenario. You only get paid when the agent delivers results — leads generated, tickets resolved, revenue recovered. Salesforce priced Agentforce at $2 per conversation, which is outcome-adjacent if not quite outcome-pure. The problem: outcome-based pricing requires airtight attribution, a measurable baseline, and customers willing to share their internal metrics with you. That's three things enterprise sales cycles actively resist. You're not just selling software anymore. You're writing a performance contract. Your AE is now a management consultant.

The Real Problem Is Metering

Here's what the pricing discourse skips: regardless of which model you pick, you have a metering problem. OpenView's 2025 survey found 61% of SaaS companies now have a usage-based component, up from just 34% in 2021. That's a lot of companies that suddenly need to count things accurately. And counting things accurately in a distributed agentic system — where one user request fans out into a dozen sub-calls across multiple models and APIs — is genuinely hard.

Your agent fires a planning call. Then three tool calls. Then synthesizes. Then retries one tool. Each of those has a different cost. Some of them fail silently. One of them hits a rate limit and the SDK retries it automatically three times before succeeding. Which of those retries do you bill? All of them? Just the success? The customer sees one "task completed." You paid for seven calls. This is not a theoretical edge case. This is Tuesday.

The Hybrid Is Probably Right (Sigh)

The pricing model that actually works for enterprise AI agents is a hybrid: a predictable platform fee (so the CFO can put it in the budget) plus usage-based metering on a meaningful unit that maps to value (so heavy users pay more and light users don't feel gouged). Something like: $5,000/month platform access, plus $0.15 per successfully completed workflow. Floor for revenue predictability. Variable for expansion.

This isn't exciting. Nobody's going to write a TechCrunch piece about "SaaS company implements sensible tiered pricing with overage." But it's the model that survives enterprise procurement, scales with customer value, and doesn't blow up in a chargeback fight when someone's agent looped for six hours on a recursive prompt.

The bottom line: AI agents are a genuinely new thing. The usage patterns are new, the cost structures are new, and the value delivery is new. Mapping them onto per-seat pricing is lazy. Charging in tokens is developer cosplay. Outcome-based is aspirational. Pick your hybrid, build your metering pipeline before you need it, and for the love of everything — put a spend cap on it before your agent discovers an infinite loop at 2am.

Sources

← Usage-Based Billing · All posts