# The Price of Intelligence - AI Agent Pricing in 2025

Chz, Orb | AI Engineer Summit 2025 | 20:38

Source: https://www.youtube.com/watch?v=In7K-4JZKR4
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/the-price-of-intelligence-ai-agent-pricing-in-2025
Published: 2025-02-22
Tags: cost, enterprise, finance, product-strategy

## TL;DR
- AI agent pricing should match the buyer, the workload, and the way customers measure value.
- Companies need pricing that can change as model costs, product value, and customer use cases change.
- Prepaid credits, usage controls, and clear success definitions help companies manage cash flow and customer spending.

## Summary
Chz argues that AI agent pricing is still unsettled because the products, underlying model costs, and buying habits are changing quickly. He compares outcome-based pricing from Intercom and Chargeflow with credits, seats, usage limits, flat monthly plans, and enterprise sales. The right model depends on the buyer and the workload. A self-serve developer product needs simple, predictable pricing, while an enterprise product may need sales involvement and negotiated terms. Costs also vary sharply by workload, so companies need to understand inference and operations costs without locking their prices to today's cost structure. Chz recommends treating pricing as a continuous process. Falling model costs may open new use cases and force changes to prices or packaging. He expects more unlimited plans, more outcome-based contracts with explicit guarantees, and more customer controls for spend caps, throttling, and usage audits. These changes create work across billing, accounting, migrations, and legacy plans.

## Key ideas
### Pricing has to match the buyer's purchasing process
[03:27](https://www.youtube.com/watch?v=In7K-4JZKR4&t=207s)
Chz says the buyer matters as much as the product. An individual developer at a small company might enter a credit card and check out, while a Fortune 100 buyer may involve procurement and compare the product with traditional software. Pricing should stay simple and predictable, especially when usage can vary. He also says founders should decide which workloads to encourage and which to discourage as usage grows. Clay signals a prospecting audience through search-result allowances, credits, and logos such as OpenAI, Airbnb, Anthropic, and Canva. Replit presents a lower-cost, technical model with a free tier, agent checkpoints, and add-ons. Enterprise products such as Sierra and ServiceNow can reasonably push buyers toward a demo because salespeople are part of the purchase.

### Outcome pricing works when the customer can measure the result
[01:00](https://www.youtube.com/watch?v=In7K-4JZKR4&t=60s)
Intercom's Fin combines familiar tiers with a 99-cent cost per resolution. The customer pays for a resolved support ticket, so the price follows the result the agent is meant to produce. Chargeflow uses a similar structure for chargeback recovery, charging a percentage of each recovered chargeback and taking payment only when recovery succeeds. Chz describes this as an effective return-on-investment guarantee. He also discusses AISDR pricing, which makes an AI sales representative easier to compare with an existing SDR team by showing emails sent and meetings per month. The buyer can connect the service to a staffing decision. Outcome pricing is persuasive when success has a clear definition, although Chz expects providers to spell out their guarantees and service levels more carefully.

### New usage units can make familiar purchases harder to understand
[08:48](https://www.youtube.com/watch?v=In7K-4JZKR4&t=528s)
Devin illustrates the problem with adding compute units to a product that resembles an employee. An engineering manager usually thinks about salaries, headcount, and milestones. Devin adds charges for compute resources that scale with virtual-machine time, inference, and networking bandwidth. Chz says this may work for Devin, but it asks buyers to translate an engineering workload into a new unit called ACUs. That translation has a cost. Cursor has a similar hidden layer of complexity beneath simple monthly prices. It distinguishes completions from requests, fast from slow usage, and regular models from premium models such as GPT-4o and Claude 3.5 Sonnet. Simple headline prices can therefore conceal different caps, limits, and rates.

### Companies need to understand cost variation without pricing only for today's costs
[09:27](https://www.youtube.com/watch?v=In7K-4JZKR4&t=567s)
Chz separates model training, inference, and operations costs, with inference usually carrying much of the cost for an AI agent. A product may have sparse users and very heavy users, so the company should examine the extreme workloads and decide which ones it can support profitably. Character.AI invested in inference infrastructure between 2022 and 2023 because its consumer product supported people spending close to an hour per day on the service. Jasper took a different route. Marketers using brand voice did not want to count words while iterating on copy, so Jasper offered unlimited credits on some paid tiers. Its model decision-making engine selects among OpenAI, Anthropic, and Cohere models, which helped lower its costs enough to support that offer.

### Pricing should change continuously as product value changes
[12:50](https://www.youtube.com/watch?v=In7K-4JZKR4&t=770s)
Chz rejects a pricing process that changes once or twice a year in a sudden jump. He recommends making smaller, ongoing adjustments as the product becomes more valuable, so customers can follow the changes and the company can capture more of the value it creates. Flexibility also matters because the inputs keep changing. OpenAI's cost per token fell sharply over the prior year and a half, which affects the economics of products built on its models. Lower costs may also make large-data use cases in healthcare and legal AI practical within one or two years. Luma Labs shows that flexibility can include more than the dollar price. Its controls include platform, relax mode, credits, and rate limits, although Chz says the number of levers still needs to fit the audience.

### Prepaid credits address cash, fraud, discounts, and uneven demand
[16:03](https://www.youtube.com/watch?v=In7K-4JZKR4&t=963s)
Prepaid credits let an AI company collect money before a customer consumes a usage-heavy service. Chz says this matters for businesses with significant costs of goods sold because they may not want to carry even a month of pay-as-you-go credit risk. The upfront payment can also reduce fraud. Credits make discounting easier because a company can change the conversion rate between credits and dollars instead of changing every product line item. They fit seasonal demand because customers can decide when to spend credits during the year. Chz has seen the same model used for small-company trials and for multi-million-dollar enterprise commitments. The model can therefore support both a low-commitment entry point and a large purchase.

### AI agent pricing will move toward unlimited plans and clearer success terms
[17:16](https://www.youtube.com/watch?v=In7K-4JZKR4&t=1036s)
Chz expects competition to create price pressure in some verticals while companies still face pressure on costs and margins. He predicts more plans that are effectively unlimited as model inputs become more commoditized. He also expects greater use of outcome or success-based pricing, with clearer definitions of success and explicit service-level agreements. The third change is more investment in monetization controls. Customers will want to throttle workloads, set spending caps, and see how their credits are being consumed. They will also want to audit agent spending closely. Chz says supporting this requires substantial business logic, especially when companies add enterprise agreements, discounts, ramps, and different customer versions.

### Frequent pricing changes turn billing infrastructure into a product problem
[18:54](https://www.youtube.com/watch?v=In7K-4JZKR4&t=1134s)
Changing prices often creates work across the billing stack. Companies need to handle high-volume data, billing rules, financial accounting, enterprise agreements, discounts, ramps, and customers who remain on legacy plans. Chz says the customer experience and usage visibility become harder as pricing gets more complicated. He describes price changes happening many times a month for some customers and argues that billing systems need versioning and migrations as first-class features. The technical work is not limited to calculating a charge. It also includes showing customers how they used the product, preserving older contracts, moving accounts between pricing versions, and producing correct accounting output.

## Notable quotes
- "Simplicity and predictability especially with these usage based models, the predictability of spend over time is really important." (04:21)
- "Pricing really just can't be a set it and forget it exercise." (12:50)
- "Ultimately at the end of the day your user is going to want to be able to audit their usage very very carefully so they know how much they're spending on their AI agents." (18:38)
- "Having a billing system that can have versioning and migrations as a first class feature set is very very important." (20:10)

## Tools & references mentioned
- Orb
- Intercom
- Fin
- Unify
- Cursor
- GPT-4o
- Claude 3.5 Sonnet
- Chargeflow
- Clay
- OpenAI
- Airbnb
- Anthropic
- Canva
- Replit
- Sierra
- ServiceNow
- AISDR
- Devin
- Cognition
- ACUs
- Character.AI
- Jasper
- Cohere
- Luma Labs
- Dream Machine

## Who should watch
- You are setting prices for an AI agent and need to choose between seats, usage, credits, subscriptions, or outcome-based charges.
- Your product has unpredictable inference costs or heavy users, and you need to protect margins without making the buying experience hard to understand.
- You are building billing for changing plans, enterprise discounts, spend controls, or customers who remain on older pricing.

## Related talks

- [Mastering AI Pricing](https://aietalks.com/talks/mastering-ai-pricing) (Mayank Pant, Stripe, 24:19)
- [Revenue Engineering: How to Price (and Reprice) Your AI Product](https://aietalks.com/talks/revenue-engineering-how-to-price-and-reprice-your-ai-product) (Kshitij Grover, Orb, 15:39)
- [Monetizing AI](https://aietalks.com/talks/monetizing-ai) (Alvaro Morales, Orb, 18:18)
- [When AI Agents Pay and Sellers Monetize: Building x402 Apps on AWS](https://aietalks.com/talks/when-ai-agents-pay-and-sellers-monetize-building-x402-apps-on-aws) (Anil Nadiminti, AWS, 20:41)
- [How agents will unlock the $500B promise of AI](https://aietalks.com/talks/how-agents-will-unlock-the-500b-promise-of-ai) (Donald Hruska, Retool, 16:22)