AI agent glossary for reading a price page

IC Guides · AI Agents · Start here

By IC Lab Desk · Edited by Yajush Gupta · How we report

Price pages for AI agents use words that sound alike and mean different things. This glossary covers the ones that decide your bill. Each links to the live examples in the AI Agent Pricing Tracker.

Seat

A fixed price per user. Predictable, but you pay for people who barely use it.

Outcome

A result the vendor defines and bills for, such as a resolved ticket. The definition is the vendor’s.

Resolution

An outcome in support, where the agent closes a conversation. Vendors differ on what counts, including whether silence counts.

Conversation or session

One exchange with a customer. Billed whether or not it solved the problem.

Credit

A prepaid unit of capacity. Different actions use different amounts, so credits are not tasks.

Token

A chunk of text a model reads or writes. Developer platforms bill by tokens, and agents use many per task.

Execution

One run of a workflow. Some tools bill per execution, others per step inside it.

Module

One step in a workflow. On some tools every module is a billable unit.

Hosting minute

A minute of voice agent runtime. Often excludes the model, voice and phone line.

Escalation

The agent hands the case to a human. Whether that is billed varies by vendor.

Hybrid pricing

A platform fee plus usage. Most enterprise deals end up here.

Overage

Usage beyond your plan. It may be billed at a higher rate or blocked.

Commitment

A volume you agree to pay for in advance, usually for a lower rate.

Concurrency

How many calls or runs can happen at the same time. Sometimes priced separately. Retell includes 20 concurrent calls free and charges $8 a line a month beyond that; Vapi includes 4 and charges $10 a line.

Cached input

Input tokens the provider has seen recently and bills at a lower rate. On OpenAI’s gpt-6.1-sol, cached input is $0.10 per million against $2.00 for fresh input.

Orchestration fee

A percentage charged on top of credits by some builder platforms. Gumloop lists 8% on Pro.

Pooled credits

Credits shared across a workspace rather than owned by one seat. One heavy user can draw down the team’s allowance. Lindy pools 5,000 per seat.

Excessive agency

OWASP’s term for an agent given too many tools, too many permissions or too much autonomy, so that a wrong or manipulated output causes damage. See permissions and safety.

Workflow versus agent

Anthropic’s distinction is that a workflow follows predefined code paths with model calls inside, while an agent lets the model direct its steps and tool use.

Read next Which AI agent, and what it really costs.

One workload, different meters

Imagine 1,000 customer requests in a month. Suppose each request triggers one conversation, five workflow module actions and one full workflow run, and 700 requests are resolved. These are illustrative assumptions, not vendor prices.

Billing unit	Usage counted in this example	What to verify
Conversation	1,000 conversations	When a conversation starts and resets
Resolution or outcome	700 successful outcomes, if all qualify	What counts as success and whether reopened cases are charged
Execution	1,000 workflow runs	Whether retries count as additional runs
Module action	5,000 actions	Which modules consume billable units
Credit or token	Cannot be determined from request count alone	Rates per action, model usage and context length

This is why two vendors with the same advertised unit price can produce different invoices. Map the work to the billable event first, then multiply by the published rate and add fixed fees.

Apply these terms See how action credits compare in Copilot Studio vs Agentforce, how workflow steps affect charges in n8n vs Make, and how outcomes differ from quotes in Fin vs Ada.

Related guides What is an AI agent? · Which AI agent, and what it costs · Free AI agents · Build or buy · Fin vs Zendesk vs HubSpot · AI Agent Pricing Tracker · All AI agent guides