Pricing AI agents by completed jobs: Sierra's $400M ARR ledger

You are choosing between seat pricing, API-credit pricing, and pricing by completed work. Sierra's ledger points to the third: the company had reached more than $400M ARR in October 2026, adding more than $150M ARR in Q3 2026. It bills customers at about $1.50 per resolved customer-service interaction instead of per seat or API credit. ADT, Sonos, WeightWatchers, and Casper are named among Sierra's disclosed customers. The buyer is paying for completed work, not access to a tool.

Price the resolved job, not the seat

When an agent closes a refund, a password reset, or a billing question, the buyer gets value at the moment the issue closes. A per-resolution unit gives operators a number they can compare against labor cost, escalation rate, and satisfaction. It also changes the sales conversation. Instead of arguing that the agent is useful, you can show what one completed job costs and what a batch of jobs saves. That makes the pricing unit easier to defend when the customer asks why the invoice is worth paying. Seat pricing suits a buyer that is buying a capability for a team. API-credit pricing suits a buyer whose main cost is consumption. Outcome pricing fits best when the buyer can verify the end state and the agent can own it.

Watch doubling time to see if usage compounds

Outcome pricing only looks durable if the work compounds. A resolved interaction should improve routing, reduce repeat contacts, and make the next batch of jobs easier to complete. When each job starts from scratch, the price may hold, but the growth story weakens. The ledger to watch is how fast ARR doubles as usage expands.

Sacra's model put Sierra's ARR at $26M in late 2024, above $100M by November 2025, and at $200M by May 2026. Sierra's ARR doubling time shortened from six months for the $100M-to-$200M step to five months for the $200M-to-$400M step. Raw ARR can be flattered by a large launch, a one-time migration, or a customer that buys more volume. Doubling time forces the question: is the next step faster because the system is learning, or because you are selling harder?

Let the multiple compress as revenue catches up

A high valuation can survive strong growth, but it needs the growth to keep arriving. Investors price in future ARR, so the multiple is a forecast. When the agent's work compounds, revenue can catch up; otherwise, the multiple becomes the story.

Sierra raised a $950M Series E in May 2026 at a $15.8B post-money valuation, led by GV and Tiger Global. That valuation was about 79x its roughly $200M ARR at the time and under 40x its October 2026 ARR. The gap between those multiples is the market's bet that the work keeps compounding. For a founder, that compression is the useful part: it turns a pricing model from a sales story into a measurable operating result. When the multiple does not compress, the company is still proving the model. It may need a narrower job, a better resolved-state definition, or a clearer link between saved labor and revenue.

Run the ledger before you commit

Start with a narrow job where completion is verifiable. Define the resolved state before you price it: if the customer still calls back, the outcome has not occurred. That definition protects the price and the ledger. Usage share is the quiet metric. When the agent handles a larger slice of the same support queue, the customer sees value without buying more seats. When the queue shrinks because the agent resolves more, the company must show that the saved labor is worth the invoice. Outcome pricing earns trust there. Keep the unit narrow enough that a customer can check it.

Track the ledger:

• Price per completed job.
• ARR doubling time and usage share.
• Valuation multiple as the work compounds.