When you deploy Devin to write code autonomously or Intercom's Fin to resolve support tickets, you're not paying for a tool. You're paying for a digital employee—complete with a base rate, expenses, utilization tracking, and project attribution requirements.
This shift from tools to teammates happened in 2025. When you use GPT-4 through an API, you pay for a tool—measured in tokens, predictable in consumption. But AI agents are packaged, priced, and billed like workers. Most billing infrastructure predates this model.
Last Updated: January 2026
The Contract Worker Mental Model
AI agents as contract workers: AI agents are contract workers leased from model providers—with base rates (subscriptions), expenses (token costs), utilization tracking, and project attribution requirements. Managing them requires the same financial discipline as managing a distributed contractor workforce.
The parallel is exact:
| Contractor Dimension | AI Agent Equivalent |
|---|---|
| Day rate / retainer | Agent subscription ($20-$500/month) |
| Expenses | Underlying model costs (tokens consumed) |
| Utilization | Active work vs. idle time |
| Project attribution | Which customer triggered which work |
| Performance tracking | Cost per resolution, cost per task |
Most companies track agent subscriptions but ignore the underlying model consumption. Or they bury agent costs in a general "AI/ML" line item with no attribution. This is like hiring contractors without tracking which projects they worked on.
The Two-Layer Cost Structure
Here's what makes AI agent billing distinct: agents have their own AI costs.
Two-layer cost structure: Every AI agent has two cost layers—Layer 1 is the agent fee (subscription, per-resolution, or per-unit), and Layer 2 is the underlying model costs (tokens consumed to do the work). True cost-per-outcome requires visibility into both.
When Devin writes code, it calls underlying models. When Fin resolves a ticket, it consumes Claude or GPT tokens behind the scenes.
Some providers bundle these costs (Intercom's $0.99/resolution includes model costs). Others expose them separately. Either way, understanding both layers affects whether pricing is sustainable, whether you could build cheaper, and how to model costs when switching providers.
The Seven Agent Pricing Models
The market hasn't converged on a standard. Seven distinct approaches have emerged:
| Model | Example | Billing Unit | Typical Cost |
|---|---|---|---|
| Agent-as-Employee | Devin | Monthly + ACU | $20/mo + $2.25/ACU |
| Per-Resolution | Intercom Fin | Resolution | $0.99/resolution |
| Per-Conversation | Salesforce Agentforce | Conversation | $2/conversation |
| Hourly | Microsoft Copilot for Security | Hour | $4/hour |
| Compute Unit | Various | Credit/ACU | Varies |
| Hybrid | Enterprise platforms | Seat + usage | $85/seat + variable |
| Outcome-Based | Chargeflow | % of value | 25% of recovered |
Each model creates different billing challenges. Per-resolution pricing can spike unpredictably—a better knowledge base means more resolutions means higher costs. Your $3,000 monthly bill becomes $8,500 during a product launch.
Compute units like Devin's ACUs represent a new billing primitive that didn't exist before 2025. Comparing $2.25/ACU to $0.99/resolution to $4/hour requires normalizing to a common metric: cost per task completed, regardless of how the provider charges.
The Attribution Gap
If you're building a product that uses AI agents, you need to attribute agent costs to customers. This is where it gets hard.
Your SaaS uses Devin. Customer A triggers 50 sessions monthly. Customer B triggers 500. Questions your billing infrastructure needs to answer:
- What's the true cost of serving each customer?
- Is Customer B profitable at their current tier?
- Should you implement usage limits?
The attribution chain spans: customer action → product feature → agent invocation → subscription cost (amortized) → model costs → total cost-to-serve.
Most companies see agent costs as one lump sum and customer revenue as another. The connection between them is invisible. This creates the same dynamic that makes GitHub Copilot lose $20/month per power user—and the same power user problem that affects subscription businesses generally.
What This Means
Agent costs are neither purely fixed (like seats) nor purely variable (like API tokens). They're a hybrid requiring new infrastructure:
For cost structure: You need visibility into both agent fees and underlying model consumption, attributed to customers and projects.
For pricing: If your costs are outcome-based (per resolution, per task), should your pricing to customers reflect that? Intercom, Zendesk, and Salesforce are grappling with this same question.
For margins: Agent-heavy customers can be unprofitable even at premium pricing. Without attribution, you won't know which ones.
Companies that build agent-aware billing infrastructure—tracking both cost layers with customer attribution—will understand their true cost-to-serve. Companies without it will make pricing and build-vs-buy decisions based on incomplete data.
The agent era is here. Billing infrastructure needs to catch up.
Tracking costs across AI agents and models? Bear Lumen maps agent consumption to individual customers automatically—see how cost-to-serve tracking works.