24 terms · Browse A–Z or search above
A software toolkit for building agents and coordinating their steps, tool calls and state. It supplies reusable building blocks; the team building the application still defines its business rules, permissions, testing and recovery behaviour.
In practiceA framework coordinates a research step, a draft and a review before a workflow continues.
Software that uses an AI model to choose steps and call tools in pursuit of a task. Its capabilities depend on the tools, permissions and limits it is given; a useful agent still needs checks on its work.
In practiceAn agent gathers overdue invoices and drafts a follow-up for a person to review.
Google Cloud: AI agents (opens in a new tab)API
Application programming interface
A defined way for software systems to request information or actions from one another. An API describes what a request can contain and what a response looks like; access may require credentials and specific permissions.
In practiceA dashboard requests the current stock level from an inventory system through its API.
MDN: API (opens in a new tab)A record of what happened, when it happened and who or what performed it. For an AI workflow, useful records include the inputs, proposed action, reviewer decision and execution result, with sensitive information handled appropriately.
In practiceA manager can trace a stock-transfer recommendation back to its source records and approval.
A sequence of tasks triggered by an event, a schedule or a person. Some steps follow fixed rules; others can use AI. Clear triggers, retry behaviour and stop conditions help prevent missed work or duplicate actions.
In practiceA new enquiry creates a CRM record, assigns an owner and prepares an acknowledgement.
The amount of information a model can process in a single interaction, usually measured in tokens. Instructions, conversation history, retrieved material and output compete for space, so selecting relevant information matters even when the window is large.
In practiceA document assistant selects relevant contract clauses instead of sending an entire document library.
CRM
Customer relationship management
The practice and software used to manage customer relationships, contacts, opportunities and interactions. A CRM gives a team a shared record, but the quality of its reporting depends on how consistently that record is maintained.
In practiceA sales team records a prospect's next step, deal stage and owner in its CRM.
How well data supports its intended use. Accuracy, completeness, consistency and freshness all matter. A model cannot reliably repair missing business facts merely by producing a plausible answer, so gaps need to be visible to the user.
In practiceA sales report flags unmapped products rather than guessing which category they belong to.
A numerical representation of an item, such as a passage of text, that helps software compare it with other items. Similarity can support search and grouping, but an embedding is not a guarantee of factual equivalence.
In practiceA search system finds an invoice-chasing guide when someone asks how to follow up on late payments.
A structured check of how well a system performs on representative tasks. Useful AI evaluations examine correctness, unsupported claims, tool use, cost and failure cases, using clear criteria and examples that resemble the work users actually do.
In practiceA proposal assistant is checked against past briefs for missing requirements and invented commitments.
Further training an existing model on selected examples to change its behaviour for a task or style. It changes model parameters. It differs from retrieving documents at question time, and still requires evaluation on examples outside the training set.
In practiceA model is trained on labelled support tickets to improve classification into a team's categories.
AI that generates content such as text, images, audio or code from learned patterns and supplied inputs. The output may be useful as a draft, but fluent or polished content can still contain errors or unsupported details.
In practiceAn assistant drafts a monthly narrative from a table of verified performance figures.
An AI output that presents incorrect or unsupported information as if it were established fact. This can include invented numbers, citations or events. Source checks, constrained tasks and human review can reduce the risk without eliminating it.
In practiceA generated account summary mentions a meeting that never appears in the underlying records.
A person participates at a defined point in an automated process, such as checking evidence, approving an action or handling an exception. The design should make clear what the person must decide and what happens when they decline.
In practiceAn account manager approves a drafted renewal email before the workflow sends it.
Running a trained model on new input to produce an output. It is the use of the model rather than its training. Response time and cost depend on the model, the input, the output and the serving setup.
In practiceA model receives a customer message and returns a suggested intent and draft reply.
A connection that lets systems exchange information or coordinate actions. Building an integration involves mapping fields, handling access and failures, and deciding how changes are synchronised. A vendor logo alone does not establish that a working integration exists.
In practiceAn approved order update is written back to the correct record in a business system.
LLM
Large language model
A model trained on large amounts of data to process and generate language. It can support tasks such as drafting, extraction and summarisation. Access to current company information or external actions requires additional context or connected tools.
In practiceAn LLM turns supplied call notes into a concise summary and proposed next steps.
MCP
Model Context Protocol
An open protocol for connecting AI applications with external tools and information. A shared interface helps applications discover and use capabilities, while authentication, permissions and approval policies still need to be designed and enforced by the systems involved.
In practiceAn AI application uses an MCP server to find documents in an authorised knowledge source.
MCP: protocol introduction (opens in a new tab)The instructions and context supplied to an AI model for a task. A useful prompt explains the goal, relevant facts, constraints and expected output. Better wording can help, but it cannot supply missing evidence or guarantee a correct result.
In practiceA prompt asks for three renewal risks, each supported by a supplied account record.
RAG
Retrieval-augmented generation
An approach that retrieves relevant material and supplies it to a model when generating a response. It can ground an answer in a document collection without retraining the model. Retrieval quality, permissions and checking the final answer still matter.
In practiceA policy assistant retrieves the current travel policy before answering an expense question.
IBM: retrieval-augmented generation (opens in a new tab)A response arranged in an agreed format, such as a JSON object with named fields. Structure makes information easier for software to process. A response can fit the required shape and still contain incorrect values, so validation remains necessary.
In practiceAn extraction step returns invoice number, currency, total and due date as separate fields.
Artificially created data used for purposes such as testing, demonstration or training. It can help show how a product behaves without presenting real customer records. Realistic-looking figures are not evidence of business results or automatically a guarantee of privacy.
In practiceA fictional automotive group demonstrates inventory decisions using generated sales and stock figures.
Finding items by comparing their numerical representations, often embeddings. It is useful for similarity-based retrieval where exact keywords may differ. The closest mathematical match is not necessarily the right business answer, so filtering and relevance checks are often needed.
In practiceA knowledge search finds a relevant maintenance procedure despite different wording in the question.
Systems that work with spoken language, including speech recognition, generated speech and conversational voice applications. A voice workflow may combine several services or a speech-to-speech model, with turn-taking, interruptions and escalation affecting the quality of the conversation.
In practiceA phone assistant collects the reason for a call and hands the context to a person.