AI Business Glossary
Sixteen AI terms explained with business examples, practical limitations, and implementation guides.
Use these definitions to make a business implementation decision. Each term includes an example, a limitation, and a practical follow-up guide.
Retrieval-augmented generation (RAG)
A workflow that retrieves relevant source material and supplies it when generating an answer.
Business example: A policy assistant retrieves the current service policy before drafting its response.
What to check: A retrieved passage still needs the right authority, access permissions, and effective date.
Fine-tuning
Adapting model behavior through a training process using examples or other training signals.
Business example: A team evaluates training examples for a specialized classification task.
What to check: Training does not remove the need for a held-out evaluation set or current business evidence.
Semantic search
Retrieval that uses representations of meaning to find related material.
Business example: A question about faster replacements retrieves a passage titled expedited replacement.
What to check: Check exact identifiers and similar product names separately.
Hybrid search
Retrieval combining keyword and meaning-based signals.
Business example: A knowledge base must support product codes and paraphrased customer questions.
What to check: A combined method still needs evaluation against relevant passages.
Context window
The bounded input and output space a model can work with for a request, subject to its implementation limits.
Business example: A long document collection must be selected or summarized before use.
What to check: Fitting material into a request does not prove every relevant detail will be used correctly.
Context engineering
Designing the evidence, instructions, tools, and state available for a task.
Business example: A customer brief receives the verified account ID and current notes.
What to check: Keep source material distinct from operating instructions.
Structured output
A response organized to a defined field and type contract.
Business example: An intake result has a case ID, category, and explicit unknown values.
What to check: Correct structure does not establish factual accuracy.
Tool calling
A mechanism for a model to request operations from defined external functions or services.
Business example: An assistant requests a lookup using a verified work-order ID.
What to check: The tool service must enforce authorization and input validation.
Idempotency
A property allowing repeated execution of the same operation without multiplying its intended effect.
Business example: A retried task-creation request recovers the existing task.
What to check: A prompt asking the agent to avoid duplicates is not a durable duplicate-prevention mechanism.
Evaluation set
A collection of task cases with expected behavior or judging criteria.
Business example: An intake classifier is tested on routine, ambiguous, and missing-data requests.
What to check: Keep release cases separate from examples used to tune the implementation.
Regression test
A repeatable check for behavior lost or changed after an update.
Business example: A new model is tested against the old version on the same critical cases.
What to check: Average improvement can hide a severe failure in one category.
Abstention
Withholding an unsupported answer and returning a useful clarification or escalation.
Business example: A policy assistant cannot find an approved source and names the responsible owner.
What to check: Track unnecessary refusals as well as unsupported answers.
Human in the loop
A person participates at a defined review or decision point.
Business example: A manager approves the exact customer communication before sending.
What to check: Name what the person verifies and bind approval to the action version.
Evidence provenance
The traceable origin and version history of information used in an output.
Business example: A claim links to a document ID, version, and supporting passage.
What to check: A real citation can still fail to support the associated claim.
Batch processing
Submitting independent tasks for asynchronous processing and later reconciliation.
Business example: An overnight job reviews a backlog of completed support records.
What to check: Track success and failure per item instead of assuming submission equals completion.
Cost per accepted output
Total operating cost divided by deliverables meeting the acceptance standard.
Business example: The denominator excludes rejected drafts while costs include their review and rework.
What to check: State what costs and acceptance rules were included before comparing workflows.
Technical references
Explore OpenAI retrieval documentation and Anthropic context engineering guidance for deeper technical detail.