Pricing AI Agents: The Labor Cost Replaced

Charge for an AI agent against the fully loaded cost of the work it removes, then subtract what the buyer still pays to supervise it. A US customer service representative earns a median $21.53 an hour. Wages make up 70% of what employers spend on compensation, so the loaded hour costs about $30.76. The going rate for a resolved support conversation sits between $0.50 and $2.00. Pricing AI Products argues that founders anchor on the per-resolution rate and leave most of the labor cost on the table.

Agents are paid from the labor budget

A seat priced access to a tool a person used, and one more user meant one more unit of value. An agent breaks that link, because it closes the tickets of ten people without asking for ten seats, so seat revenue falls as the agent gets better. The owner of the support headcount compares an agent with a salary. Salesforce now calls its agents digital labor, so a founder who charges $40 a seat for an agent sells labor at a software price.

Published resolution prices sit far below the human cost per contact

The support category shows the gap best, because four vendors publish a price per resolved conversation. HubSpot cut its Customer Agent to $0.50 per resolution in April 2026, Intercom bills $0.99 per Fin outcome, Zendesk charges $1.50 per automated resolution on committed volume and $2.00 without it, and Salesforce launched Agentforce at $2 per conversation in 2024. Gartner put the cost of a live-channel contact at $8.01 in 2019, and wages have risen since.

Strandable cost sets the ceiling on the price

The fully loaded salary sets the most an agent can be worth to the buyer, and the price has to sit well under it. A buyer can only release the part of the budget the agent removes, and an agent that handles 80% of a role rarely lets anyone delete 80% of the headcount, because exceptions, oversight and escalations still need a person. Call the releasable part strandable cost, the loaded cost of the work the agent retires minus the supervision the buyer keeps paying, and price against it.

Take a team of 10 representatives at the loaded $30.76 an hour, or $639,808 a year, that works an illustrative 60,000 tickets in the modeled example below.

Table 1 · Strandable cost for a modeled support team

At $4.00 the buyer still keeps about $120,000 a year and the vendor collects nearly three times the market-rate bill. The work also has to pay for itself on your side. In the book’s unit-economics example a routine ticket costs about $0.38 to resolve and a complex enterprise ticket about $1.20, so a $1.50 price earns 75% on the first and 20% on the second.

Price the job, then the result

Buyers sign off on a meter for the raw calls or tokens an agent burns with little argument, and a rival can win the account by metering the same actions for a little less. Chapter 33 of the book sets out three models above that floor. A per-agent price is a monthly fee set against the salary the agent offsets, and it suits an agent that maps to a role. Workflow pricing charges for a finished job, such as an underwritten loan application or a reconciled invoice, tiered by complexity, while outcome pricing charges only for a measured result and works once you can attribute that result to the agent.

Cognition ran the whole arc in four months. It sold Devin at $500 a month in December 2024, as a hire with a salary-shaped price, then moved to a plan starting at $20 in April 2025, with work billed in compute units at $2.25 each. Sierra sits at the far end, where it says it gets paid only when it completes a task.

Metrisque and an agentic GTM company moved to outcomes and credits

I have made this switch at two companies. Metrisque, the AI visibility product I work on, used to charge one price per business, so the account that asked for the most work paid the same as the one that asked for the least. It now charges by outcome, for each piece of work delivered, and the change lifted revenue to about 2.5 times what the per-business price brought in. Per-business pricing had capped what the heaviest accounts paid, and outcome pricing removed that cap.

At an agentic go-to-market software company whose board I sit on, the price was per user, which charged for logins while the agents did the work. We repackaged it on the Kano model, with features grouped into must-haves, performance features and delighters, and moved usage onto credits. A credit lets heavy usage pay for its own compute, which a per-user price never did.

Meter underneath every outcome price

Outcome pricing makes your best customer your most expensive one. The account that generates the most resolutions also burns the most compute. Set a margin floor and turn it into a cost ceiling: at a 40% floor, a $0.99 resolution price allows about $0.59 of cost per resolution, and any account above it moves to a higher rate, accepts a cap or leaves. Track compute per outcome per account, since a blended gross margin can read healthy while your largest logo loses money on every ticket.

Start the conversation at the loaded cost of the role, and let the negotiation settle downward from there. Pricing AI Products works these numbers through for agents, credits and outcome pricing.

Questions

How much should I charge for an AI agent?

Price it against the strandable cost of the work it removes: the fully loaded cost of the labor retired, minus the supervision the buyer still pays. Pricing AI Products by Nitin Kumar treats the loaded salary as the ceiling on value.

Should an AI agent be priced per seat?

Rarely, because an agent does a team’s work without a login per person, so per-seat revenue falls as the agent improves. Seats still suit software whose value grows with the number of people using it, such as a CRM.

What is outcome-based pricing for AI agents?

Outcome-based pricing charges only for a measured result, such as a resolved conversation. Intercom bills $0.99 per Fin outcome and HubSpot $0.50 per resolution.

How do you protect margin on an AI agent?

Know the fully loaded cost per successful action (inference, tools and any human review) and meter it per account even when you bill on outcomes. Then cap or tier the exposure so a heavy account cannot turn a healthy margin negative.