How to Price AI Agents
Outcome-based pricing is not a pricing tactic. It is how value gets captured in the P&L.
At Directly, pre-LLM, the Instant Answers AI priced at ~$1 per resolution. That price started from what a resolved ticket was worth to the customer, not from our inference cost. EBITDA margin sustained above 22% over two years. The pricing model was the difference.
What is outcome-based pricing in AI?
Outcome-based pricing ties revenue directly to a measurable business result. In AI, that typically means: cost removed, revenue generated, or work completed.
To price that outcome, measure Customer EBITDA Created. The customer will compare the cost of the AI agent to the EBITDA it generates. That comparison determines whether revenue renews.
Revenue becomes a function of output, not access. That changes the economics of every function in the company.
Customer EBITDA Return (CER)
CER is CEC divided by your price. It answers the question every customer asks at renewal: is this worth what I pay?
Below 3x: the value is too thin. You will be rationalized.
3x to 10x: durable revenue. The customer sees clear return and has a strong incentive to stay.
Above 10x: you are underpriced. Margin is sitting on the table.
The 3x to 10x range is where AI revenue becomes structural. Price inside it.
Market pulse check
These companies have shifted to outcome-based pricing. The last column is the one the industry does not track yet.
| Company | Vertical | Outcome | AI Price | Human Equiv. | CER Range |
|---|---|---|---|---|---|
| Intercom | Support | Resolved conversation | $0.99/resolution | $5–$15/ticket | 4x–14x |
| Zendesk | Support | Automated resolution | $1.50–$2.00/res. | $5–$15/ticket | 2x–10x |
| Salesforce Agentforce | Sales/CRM | Per conversation / per action | $2/conv. → $0.10/action | $15–$30/interaction | 7x–14x |
| EvenUp | Legal | Demand letter | ~$300–$500/demand | $2K–$5K/demand | 3x–16x |
| AP Automation (category) | Finance/AP | Invoice processed | $2–$5/invoice | $15–$25/invoice | 3x–12x |
The pattern: companies pricing in the 3x to 10x CER range retain revenue. Above 10x, they are in land-and-expand mode or have not done the math yet. Below 3x, they will be rationalized at renewal. That is not a pricing problem. It is a CER problem.
Notes
– Salesforce launched at $2/conversation in late 2024, then introduced Flex Credits ($0.10/action) by mid-2025 after customer pushback on outcome ambiguity. If the outcome definition is not unambiguous, the pricing model breaks.
– EvenUp uses a hybrid AI + human review model. Base price ~$300; effective cost ranges higher depending on review tier. CER range calculated against base.
– Human equivalent costs are industry benchmarks: support ticket cost per Gartner/HDI ($5–$15), sales interaction cost per Salesforce/Forrester ($15–$30), demand letter cost per PI industry surveys ($2K–$5K), invoice processing cost per IOFM/ABBYY ($15–$25).
When outcome-based pricing works
Outcome-based pricing is underwritable when three conditions are true:
- 1/
The outcome definition is unambiguous. One line. No ambiguity. Example: resolve a support ticket without human intervention.
- 2/
Attribution is provable. You can demonstrate the AI caused the outcome and the outcome has economic value.
- 3/
Delivery is controllable at scale. You control enough of the system to produce the result consistently. If the customer controls too many variables, you are pricing volatility.
Minimum Productive Agent (MPA)
You can decide to offer outcome-based pricing anytime. MPA is the threshold where your AI agent reliably delivers enough value to support it. Three tests:
- 1/
You can measure Customer EBITDA Created (CEC).
- 2/
The customer sees 3x or better return on your price (CER).
- 3/
The customer agrees with both.
Check all three for three customers and you are at MPA. That is when outcome-based pricing becomes the conversation. Before MPA, price on usage. After MPA, price on outcomes. The bridge is not the destination.
For the full CEC and CER measurement frameworks, see The AI to EBITDA Playbook.
Usage-based pricing: the bridge
Usage-based pricing is the bridge when your AI agent cannot yet reliably deliver a measurable outcome. It buys time.
- 1/
Fixed-price consumption bands. Flat fee for a defined volume of agent activity. Predictable for the customer. Protects your revenue floor.
- 2/
Tiered pricing. Lower tiers reduce customer risk and accelerate trial conversion. Higher tiers capture expansion as usage grows.
- 3/
Hybrid with a platform fee. Fixed platform fee as the floor, usage-based upside above it. When the agent reaches MPA, the hybrid model migrates cleanly to outcome-based pricing.
When outcome-based pricing fails
Outcome pricing breaks when the product is still multi-purpose, the outcome depends on the customer's process, attribution is disputed, or delivery is inconsistent.
Symptoms: custom contracts, margin leakage, finance pushback, stalled deals.
Outcome pricing does not fix those problems. It exposes them. Get the workflow and the measurement right first. Then price.
How to get there
- 1/
Instrument. Measure the outcome inside the product. Track CEC from day one, even before you price on it.
- 2/
Shadow price. Run outcome economics internally alongside your current model.
- 3/
Pilot. Test with controlled customers. Validate that both sides can verify the outcome and agree on the return.
- 4/
Scale. Align GTM, contracts, and packaging around the outcome model.
What outcome-based pricing does to the org
When pricing anchors to customer outcomes, the entire organization aligns around it.
Engineering works backwards from the margin. Inference costs fall because they have to. GTM sells value, not features. CS optimizes customer impact, not usage volume. AI performance becomes a leading indicator for Finance.
EBITDA quality improves because revenue is earned through performance. That is what a buyer pays a durable multiple for.
Key takeaway
AI changes the unit of value. If you price outcomes without control, you take on risk. The goal: the AI reliably delivers a result, the result is measurable, pricing captures the value. That is how AI turns into EBITDA.
If your portfolio company has AI that isn't driving EBITDA, I've solved that problem twice.