What's an Outcome? The First Question Every AI Team Gets Wrong

We talk to AI teams every week about cost tracking. The conversation stalls at the same place every time.

"We want to track cost per outcome. But what counts as an outcome?"

Fair question. Most teams overthink it. Here's a practical framework based on what we've seen actually work, using customer service AI as the running example.

Start With What the Customer Pays For

If the product is AI customer support, the customer doesn't pay for API calls or tokens. They pay for resolutions. A customer inquiry that gets answered without a human stepping in. That's the outcome. A resolved inquiry.

Not a model inference. Not a conversation turn. Not a session. A resolution.

In our tooling this is the unit everything ties back to. Voice Agent: 3,180 resolutions/month. Chatbot: 7,420. Escalations: 1,800.

One Outcome, Many Costs

A single resolution isn't a single cost. It's a chain.

• Intent routing. The system classifies the incoming query. One agent call, about 420 tokens, $0.063.
• Knowledge retrieval. RAG pipeline fires. 3.2 vector DB queries per outcome, plus embedding lookup and reranking. Full RAG cost per outcome is $0.146.
• Action execution. If the resolution requires doing something (updating an account, creating a ticket) that's another agent.
• Response generation. Final agent synthesizes everything and responds. About 920 tokens, $0.12 per call.
• Infrastructure. Voice gateway on EC2, API layer on Lambda + API Gateway, audio storage on S3, networking underneath.

One resolution. Four agent calls. Multiple RAG queries. Infra overhead. Five different vendors. Total for Voice Agent: $1.98. For Chatbot: $0.58. Same product, same outcome definition, very different cost.

Mistakes We Keep Seeing

"Our outcome is a conversation." Too vague. A conversation could be one turn or twenty. There needs to be a success condition. The inquiry was resolved, the ticket was closed, the customer didn't call back within 24 hours.

"Our outcome is an API call." Too granular. That's an ingredient. There are dozens of API calls per actual business outcome.

"We track monthly spend per feature." Right direction, wrong denominator. $6,300/month on Voice Agent means nothing without knowing that's spread across 3,180 resolutions. The $1.98 per resolution is the number that matters for pricing.

"Our outcomes are all the same." They're not. P50 cost per outcome is $0.82 but P99 is $4.31. That's 5x. And 6.2% of outcomes cost over $3. Anyone pricing at $0.99 per resolution is losing money on every P90+ outcome. The distribution matters a lot.

A Simple Test

The easiest way to land on an outcome definition: what would go on the invoice? Charging per resolution? That's the outcome. Charging per conversation? That's it. Charging per document processed? There it is.

Then trace every cost that goes into producing one unit. Models, RAG, infra, data pipelines. All of it.

This Feeds Directly Into Roadmap

Once cost per outcome exists by feature, downstream decisions get easier. Chatbot at $0.58 with $28,800/month productivity value? High margin, expand to more segments. Voice Agent at $1.98 with strong adoption? Core bet, anchor for enterprise plans. Escalations at $0.83 with low adoption? Monitor before investing more.

The outcome definition is where everything starts. Get it right, and cost tracking, pricing, and roadmap decisions all follow.