Advanced AI product glossary
Key terms from this course, in plain English.
- Activity log
- A record of every action an AI agent has taken, so users can review and undo them.
- Agent
- An AI system that uses a model, tools and a loop to pursue a goal, choosing its own steps.
- AI PRD
- A product requirements document that also specifies AI behaviour under uncertainty: context, quality, autonomy, budget and risk.
- AI-added
- An existing product that gains AI features.
- AI-native
- A product that only makes sense because of AI, often rethinking the workflow from scratch.
- Automation bias
- The tendency to accept an automated suggestion even when it is wrong.
- Batch processing
- Running non-urgent AI jobs together, usually at a discounted price and with slower turnaround.
- Capability curve
- The rapid improvement of model capabilities over time, which products should be designed to benefit from.
- Cost per task
- The total cost of completing one user task, including all model calls, tools and human review.
- Error analysis
- Grouping AI failures into types, counting them and fixing the most important first.
- Fair-use limit
- A stated usage allowance included in a plan, above which extra charges or limits apply.
- Gross margin
- Revenue minus the direct cost of delivering the product, as a share of revenue.
- Holdout group
- A group of users kept without a feature to measure its long-term effect.
- Intervention rate
- How often a person must step in to correct or complete an AI agent's work.
- Kill criteria
- Conditions agreed in advance for stopping or rescoping an experiment or bet.
- Marginal cost
- The extra cost of serving one more request or user.
- Metrics tree
- A structure linking business outcomes to user outcomes, adoption, quality and guardrail metrics.
- Moat
- A durable advantage that makes a product hard for competitors to copy.
- Model Context Protocol (MCP)
- An open standard that lets AI agents connect to tools and data in other products.
- Model deprecation
- A provider's scheduled retirement of a model version, after which it stops being available.
- Model routing
- Sending each request to the most suitable model — smaller for simple tasks, larger for hard ones.
- Novelty effect
- A temporary spike in usage of something new, which fades as users get used to it.
- Outcome-based pricing
- Charging per successful result, such as a resolved case, rather than per seat or request.
- Prompt caching
- Reusing repeated prompt content so it is charged at a reduced rate and processed faster.
- Risk-rated tools
- Classifying an agent's tools by the harm they could cause and setting oversight for each.
- Shadow test
- Running a new model or version on a copy of live traffic without showing its output to users.
- Stop condition
- A limit on an agent's steps, time or cost, after which it stops or escalates.
- Thin wrapper
- A product that adds little beyond a prompt on top of a general model.