AI agents are digital employees. Not metaphorically. Literally.
Devin has a monthly salary ($20) and expenses ($2.25 per compute unit). Fin has a per-task rate ($0.99 per resolved ticket). Agentforce bills per conversation. Copilot for Security bills hourly.
This pattern extends well beyond named products. Any AI workflow with a flat fee and a variable expense rate, whether it's a branded agent or a custom pipeline calling foundation models, carries the same cost structure: base rates, variable costs, utilization patterns, and attribution requirements.
The shift from tools to teammates happened in 2025. Most billing infrastructure hasn't caught up.
Last Updated: March 2026
The Shift: SaaS to Service-as-Software
Traditional SaaS made human workers faster. Agentic workflows execute the work with minimal human intervention. Andreessen Horowitz calls this "Service-as-Software".
The pricing implications are structural:
| Aspect | SaaS Model | Agentic Model |
|---|---|---|
| Pricing metric | Per seat / subscription | Per outcome / resolution / % of value |
| Value proposition | Do it faster | Do it for me |
| Growth driver | Headcount growth | Transaction volume |
| Cost structure | Predictable, linear | Bursty, recursive |
| Billing trigger | Calendar month | Work completed |
When an AI agent makes 10 support reps so efficient that 2 can handle the same volume, per-seat pricing loses 80% of revenue. Despite delivering more value than ever.
The Contract Worker Mental Model
AI agents are contract workers leased from model providers. The parallel isn't metaphorical. It's structural.
A contractor has a day rate. An agent has a subscription ($20-$500/month). A contractor has reimbursable expenses. An agent has token costs, the model consumption underneath every task. A contractor needs utilization tracking. An agent needs the same: how much compute went to active work vs. idle polling.
And just like contractors, agents need project attribution. When Devin writes code, which customer triggered that session? When Fin resolves a ticket, which account absorbed the model costs?
The analogy extends further than most companies realize. Contractors have SLAs, deliverable timelines, and performance clauses. If they underperform, you renegotiate or replace them. AI agents have none of this. If Fin resolves tickets poorly, your recourse is a support ticket to Intercom or switching providers entirely. There's no performance review, no renegotiation, no structured feedback loop that helps the vendor improve. The accountability infrastructure for digital employees doesn't exist yet.
Most companies track the subscription invoices. Almost nobody tracks the token consumption per customer. They bury agent costs in a general "AI/ML" line item, the equivalent of paying a dozen contractors and never recording which projects they worked on.
Two Layers of Cost, One Invoice
Every AI agent has two cost layers, and most companies only see one.
Layer 1 is the agent fee. Devin charges $20/month plus $2.25 per compute unit. Fin charges $0.99 per resolution. Agentforce charges $2 per conversation.
Layer 2 is the model cost underneath. When Devin writes code, it calls Claude or GPT to generate each function. When Fin resolves a ticket, it runs inference to understand the question, search the knowledge base, and compose the answer. Those tokens cost money, whether the vendor bundles them into the price or bills them separately.
Intercom's $0.99/resolution includes model costs. If Intercom's average inference cost per resolution is $0.30, they keep $0.69 of gross margin. If a harder ticket requires multiple model calls and costs $0.85 in inference, the margin drops to $0.14. The buyer never sees this.
Understanding both layers tells you whether the vendor's pricing is sustainable, whether you could build it cheaper, and what to expect when switching providers.
The variable cost layer adds a frustration that's unique to this era: it's unclear what the short-term vs. future cost will be. Model prices drop. New providers emerge. Usage patterns shift as customers discover features. Meanwhile, the CFO is looking at yet another AI line item with variable costs and no clear picture of long-term value. This is the financial reality of managing a distributed contractor workforce where the contractors' expenses change every quarter.
Seven Pricing Models, No Winner Yet
The market hasn't converged. 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 problems.
Per-resolution pricing spikes unpredictably. A better knowledge base means more resolutions, which means higher costs. Your $3,000 monthly bill becomes $8,500 during a product launch. And resolution-based pricing only works when "resolved" is binary. "Research this market" or "draft this contract" have no clean resolved/not-resolved boundary.
Per-action pricing penalizes thoroughness. An agent that takes 20 careful steps to get the right answer costs more than a sloppy agent that guesses in 3. The buyer pays for diligence, not value.
Compute units (Devin's ACUs, for example) are a billing primitive that didn't exist before 2025. Comparing $2.25/ACU to $0.99/resolution to $4/hour requires normalizing to one metric: cost per task completed. Nobody has standardized this yet.
Where Attribution Breaks Down
Your SaaS uses Devin. Customer A triggers 50 sessions monthly. Customer B triggers 500. They're both on the same $99/month plan.
Your CFO asks: what's our gross margin by customer? You pull the Devin invoice ($2,400/month total) and your Stripe revenue ($15,000 MRR across 150 customers). The per-customer cost is... somewhere between $4 and $400, depending on which customer you're looking at.
You know the total. You don't know the breakdown. This is the attribution problem, and it's the same dynamic that makes GitHub Copilot lose $20/month per power user. It's the same power user problem that affects subscription businesses generally.
Without per-customer cost visibility, pricing decisions are built on averages. Averages hide the customers who cost you more than they pay.
The market hasn't settled on how to price AI agents. It may not for years. But the billing infrastructure question is already answerable: track both cost layers, attribute costs to customers, and revisit pricing quarterly.
Patterns are emerging. Both buyers and sellers of AI are learning what's comfortable and what's not acceptable. These are hard decisions on both sides. Buyers want cost predictability from a technology that's inherently variable. Sellers want to capture value from a product that does the work, not just speeds it up. Neither side has found the right balance yet.
When agents make work cheap, demand for that work explodes. Low per-outcome prices generate massive volume. The billing infrastructure that worked for 50 seats doesn't scale to millions of micro-outcomes.
At the end of the day, it's the value being built that matters. Cost attribution doesn't answer the pricing question. It gives you the data to find your own tolerance in a space where nobody has the playbook yet. The teams that instrument this early will have pricing flexibility. The rest will have pricing constraints.
If you're building with agents, Bear Lumen gives you the cost attribution layer. See how it works.