Outcome-Based AI Pricing: A Buyer's Field Guide

Bessemer Venture Partners observed a structural shift in AI pricing: vendors are moving from cost-plus pricing (we charge per token, you worry about outcomes) to outcome-based pricing (we charge per outcome, we worry about costs). This shift matters enormously for your P&L.

As a buyer, outcome-based pricing flips your negotiating position. Instead of paying for inputs (tokens, seats, conversations), you pay for outputs (resolved tickets, processed claims, completed transactions). The vendor bears the economic risk that the AI actually delivers.

The Five Pricing Models (And How They Stack Against Each Other)

1. Per-Token Pricing (Cost-Plus)

You pay $3 per million input tokens, $15 per million output tokens. The vendor collects revenue on consumption. The vendor does not care if your AI resolves your customer's problem; the vendor gets paid either way.

Pros (for you):

• Transparent. You see exactly what you are paying for.
• Scales with volume. If you do not use it, you do not pay.

Cons (for you):

• Vendor has no incentive to be efficient. More tokens = more revenue.
• You bear all the cost risk. If the AI generates 10 outputs to solve one problem, you pay for all 10.
• Hard to budget. You do not know how many tokens a task will consume until you try.

Example vendor: OpenAI (per-token), Anthropic Claude (per-token)

2. Per-Seat Pricing (SaaS Model)

You pay $500/month per user seat. The vendor does not care how much each user uses the tool.

Pros (for you):

• Easy to budget. You know the cost upfront.
• No per-usage measurement overhead.

Cons (for you):

• You pay even for unused licenses.
• Vendor has no incentive to make the tool valuable. Once you buy 100 seats, the vendor stops caring about outcomes.
• Does not map to AI outcomes at all. You bought 100 seats but maybe only 10 are active on any given day.

Example vendor: None of the good AI vendors price this way anymore. It is a legacy SaaS model.

3. Per-Conversation Pricing

You pay $0.50 per conversation handled by the AI (whether or not it resolved the issue).

Pros (for you):

• Simple. Easy to forecast volume and cost.
• Scales with usage.

Cons (for you):

• Vendor still has no incentive to resolve the issue. More failed conversations = more revenue.
• Does not align with business value. A conversation that resolves the issue is worth more than one that escalates.

Example vendor: Decagon charges $0.25-$0.50 per conversation.

4. Per-Resolution Pricing (Outcome-Adjacent)

You pay $0.50 per resolved conversation. If the AI escalates to human, you do not pay.

Pros (for you):

• Vendor now has skin in the game. More resolutions = more revenue for the vendor. Vendor optimizes for outcomes.
• You do not pay for failures.

Cons (for you):

• Vendor might be conservative (only resolve easy cases, escalate anything hard) to avoid losses.
• Vendor might misclassify resolutions (mark escalations as resolutions) to hit targets.
• Revenue recognition is messy for both you and the vendor (when does a "resolution" become final?).

Example vendor: Intercom Fin charges approximately $0.99 per resolution. Klarna claims $0.19 per resolved ticket.

5. Outcome-Based / Value-Based Pricing

You pay based on the actual business value delivered. For customer support, this might be: you pay $0.50 per ticket resolved and customer satisfaction score is 4+ out of 5. For claims, you might pay per claim adjudicated and error rate is below 2%.

Pros (for you):

• Vendor bears all the risk. If the AI does not deliver, you do not pay.
• Aligned incentives. Vendor is motivated to optimize for your actual outcome, not just token consumption.
• Best price discovery. Vendor quotes a price only when it is confident the economics work.

Cons (for you):

• Requires deep partnership. Vendor needs to integrate with your systems, understand your workflows, measure outcomes.
• Revenue recognition is complex (when do you recognize revenue if a customer is dissatisfied?).
• Vendor may be conservative on pricing (only sign contracts where they are very confident).

Example vendor: Sierra quotes per-resolution pricing with SLA guarantees. Some custom vendors quote outcome-based pricing for specific use cases.

How To Evaluate Competing Bids

Three vendors pitch you on AI customer support. How do you compare them on pricing?

Vendor A: OpenAI GPT-4 via API

• Token cost: $0.11 per conversation (input + output)
• You supply infrastructure, observability, human review, integration
• Your total cost: $0.11 (tokens) + $0.20 (infrastructure) + $0.15 (human review) = $0.46/conversation
• Outcomes: You own the outcomes. If the AI fails, that is your problem.

Vendor B: Intercom Fin

• Per-resolution pricing: $0.99 per resolved conversation
• Intercom supplies infrastructure, observability, integration
• Your total cost: $0.99
• Outcomes: Intercom owns the cost structure. You only pay if it resolves.

Vendor C: Klarna (custom deployment)

• Revenue share: Klarna takes 15% of customer support cost savings
• Klarna supplies infrastructure, handles all integration
• Your baseline: human support costs $3.00 per conversation
• If AI costs $0.50 and resolves 90%, you save $2.25 per conversation
• You pay Klarna 15% of savings: $2.25 × 15% = $0.34/conversation
• Your net cost: $0.50 - $0.34 = $0.16/conversation
• Outcomes: Klarna is incentivized to maximize resolution and minimize cost.

On a per-resolution basis:

• Vendor A: You assume the cost and the risk. You own all variables.
• Vendor B: You pay $0.99 for each resolution. The vendor assumes risk that it resolves.
• Vendor C: You pay $0.16 for each resolution (net of Klarna's share), and Klarna is incentivized to reduce cost.

Vendor C looks cheapest, but Klarna only works if customer support costs are already high and the volume is predictable. Vendor A looks cheapest if you can operate it efficiently. Vendor B is the simple default.

Revenue Recognition Implications

The pricing model affects your revenue recognition (when and how you record the expense):

Per-token pricing: You expense as incurred. When you use the tokens, you expense the cost. Clear and simple.

Per-conversation pricing: Same as per-token. You expense when the conversation completes.

Per-resolution pricing: You need to define what "resolved" means. Does it mean the customer said they were satisfied? Does it mean the ticket closed? Does it mean the customer did not reopen it within 30 days? Your CFO and auditor need to align on the definition upfront.

Outcome-based pricing: Most complex. If you sign a contract with Klarna that says "you pay 15% of cost savings," you need to measure the cost savings reliably and consistently. Your auditor will ask for controls over the measurement of "savings." This is doable, but requires clean data and documented methodology.

What To Do Next

When evaluating AI vendors, ask for total cost of ownership, not token cost. What are you paying for? What are you supplying? What happens if the AI fails?

If a vendor quotes per-resolution pricing, ask: Who defines "resolution"? Who measures it? What disputes will be resolved? Get these in writing.

If you can get outcome-based pricing, take it. It is the only model where vendor incentives and your incentives are fully aligned.