THE AI GLOSSARY FOR BUSINESS
Understand the models, products and systems behind AI.
Clear definitions. Practical examples. No technical background needed.
START WITH THESE THREE
A category, a type of model and a product.
| Concept | What it means |
|---|---|
| Artificial intelligence (AI) | The broad category of systems that recognize patterns, make predictions, generate information or select actions. |
| Large language model (LLM) | A type of AI model that processes and generates language. |
| ChatGPT | OpenAI’s product for working with AI models and tools through a conversational interface. |
300 of 300 terms
About these definitionsTry a shorter phrase or clear the filters to browse every definition.
Comparing variants with assigned groups to estimate their effect on a chosen metric. Good design controls bias and defines success beforehand.
In businessCompare two intake forms using completion rate and lead quality.
Specific conditions used to decide whether a feature or deliverable meets its requirements.
In businessA workflow must reject missing account IDs and log every approval.
Designing products so people with different abilities can use them, including with assistive technology.
In businessProvide labeled fields, keyboard access and readable contrast.
The share of predictions that match the expected answer under a defined test. It can be misleading when categories are very uneven.
In businessA classifier can look accurate while missing rare urgent cases.
Testing with deliberately difficult, misleading or hostile inputs to uncover weaknesses.
In businessPlace conflicting instructions in a test document and verify that they are ignored.
Passing a task, context or control from one agent or person to another.
In businessSend an invoice exception to the finance team with supporting records.
The surrounding software that manages an agent’s tools, state, permissions, context, execution limits and recovery.
In businessStop an agent after a budget limit and save its progress.
Stored information an agent can retrieve across steps or sessions. Memory can be incomplete, outdated or restricted by permissions.
In businessRemember a user’s approved output format without storing unnecessary personal data.
Breaking a goal into possible steps and deciding how to proceed. Plans should be revised when tool results contradict assumptions.
In businessPlan a supplier comparison, then adjust when a price is unavailable.
A coordinated group of AI agents that divide work and exchange results. “Swarm” describes a design approach, not a guarantee of intelligence, accuracy or profit.
In businessOne agent checks an inquiry, another drafts a quote, and another reviews it.
An open protocol designed to let independent agent systems discover capabilities and collaborate across implementations.
In businessA service agent delegates a task to a compatible specialist agent.
AI used in a system that can choose actions and use tools while pursuing a task. The degree of autonomy depends on its design.
In businessA research assistant decides which source to inspect next.
Software that uses an AI model, tools and instructions to work toward a goal, often choosing its next step from what it observes.
In businessAn agent gathers order details and prepares a quote for approval.
An AI assistant designed to help a person do their work while the person remains involved. The term is also used in product names.
In businessSuggest a reply for a support representative to review.
A defined spending allowance with monitoring, alerts or enforced limits for AI operations.
In businessPause nonessential batch work when its agreed usage budget is reached.
The responsibilities, policies and processes used to oversee AI throughout its lifecycle.
In businessAssign owners for approved uses, incident handling and model changes.
An evaluation of potential harms, failures and exposures for a particular AI use, considering likelihood, impact and controls.
In businessReview the consequences of a wrong answer before automating a workflow.
Search experiences that use AI to help interpret queries or synthesize answers, sometimes linking to source pages.
In businessCheck the linked original sources before relying on an AI summary.
Finding observations that differ from expected patterns. An anomaly is a signal to investigate, not proof of a problem.
In businessFlag an unusually large supplier invoice for review.
Structuring useful content to answer specific questions clearly in search or assistant experiences. The term has no single universal technical standard.
In businessStart a page with a direct explanation before adding examples.
The AI research and product company that develops Claude. The company and its Claude models and applications are related but distinct.
In businessEvaluate Anthropic as a provider and Claude as a model or product choice.
A provider’s developer services for calling models, managing access and building applications. It is different from a consumer chat subscription.
In businessYour website calls a model through a backend API integration.
A defined way for software systems to request data or actions from each other.
In businessLet an approved agent retrieve order status from your business system.
A proposed form of AI with broad competence across many intellectual tasks. Definitions and thresholds differ; it is not a precise software specification.
In businessAsk vendors to describe measurable capabilities instead of relying on the label.
Computer systems that perform tasks such as recognizing patterns, predicting outcomes, generating content or selecting actions. AI is the broad category, not one product.
In businessA company uses AI to sort requests and draft replies.
A mechanism that lets a model weight different parts of its input when computing a representation.
In businessGive relevant words different influence when processing a sentence.
A record of relevant actions and changes that supports review and accountability.
In businessRecord who approved an action and which system executed it.
Checking the identity of a user or system.
In businessRequire a user to sign in before accessing internal documents.
Deciding which actions or records an authenticated identity is allowed to access.
In businessCheck permission again before a tool returns a customer record.
An agent allowed to take multiple steps without approval at every step, within defined limits.
In businessLet an agent collect public research while reserving outreach for human approval.
A link from another website to a page. Its usefulness depends on context and quality, not just the count.
In businessA relevant industry resource links to a helpful technical explanation.
A maintained list of work items awaiting prioritization or delivery.
In businessTrack requests, defects and improvements in one place.
The starting measurement used to compare later performance under stated conditions.
In businessRecord processing time, error rate and cost before the pilot.
Processing a collection of tasks together, often when immediate results are unnecessary.
In businessClassify overnight document imports in a batch.
A standardized test used to compare systems. A strong benchmark score does not prove performance on your company’s workflow.
In businessUse public scores as context, then test your own documents.
Systematic patterns that skew outputs or outcomes. Bias can arise from data, objectives, design choices or use conditions.
In businessCheck whether error rates differ across relevant customer groups.
Using software to carry out steps in a business process, including data movement, checks, notifications and approvals.
In businessRoute a purchase request from submission to approval and recording.
Explicit conditions that govern a process independently of an AI model’s judgment.
In businessDiscounts above a threshold always require a named approver.
Reusing previously computed or retrieved information when it remains valid. Freshness and permissions determine when reuse is safe.
In businessReuse a public product description until its source changes.
How closely predicted probabilities match observed frequencies over many cases.
In businessCases assigned 80% confidence should be correct about 80% of the time under a valid test.
The preferred URL indicated for a page when similar content is reachable at multiple addresses. Search engines may choose a different canonical.
In businessPoint glossary search variants back to the main glossary URL.
An increase in work the team can handle with existing resources. It becomes a financial benefit only when the business uses or realizes it.
In businessUse recovered admin time to serve more customers rather than assume payroll fell.
Planning and supporting the people, process and responsibility changes required for a new way of working.
In businessUpdate procedures and responsibilities when automation changes an approval flow.
Software that interacts through conversation. It may use fixed rules, a language model, tools or a combination.
In businessAnswer common visitor questions and offer a booking link.
OpenAI’s conversational AI product. It makes AI models and tools usable through a chat interface; it is a product, not the name of one fixed model.
In businessAsk it to explain a report or help draft a proposal.
Dividing content into smaller pieces for processing or retrieval. Good boundaries preserve enough context to interpret each piece.
In businessKeep a policy rule with its exceptions instead of splitting them apart.
A reference identifying the material supporting a claim. A link alone does not prove that the source supports the statement.
In businessLink a policy answer to the exact current policy section.
Assigning an input to one or more categories.
In businessLabel incoming requests as billing, sales or technical support.
Anthropic’s family of AI models and its assistant product. The name can refer to the underlying model or the application used to interact with it.
In businessDraft a document in Claude or build a workflow using its API.
Anthropic’s coding agent for working on software projects. It can inspect code and use development tools within its configured permissions.
In businessAsk it to investigate a bug and prepare a code change.
The name used for Claude’s delegated knowledge-work experience, now integrated into Claude. It works with supported files and tools; capabilities depend on the plan and configuration.
In businessDelegate a document-organizing task and review the resulting files.
Computing resources delivered through remote infrastructure rather than only local machines.
In businessHost an application and scale its capacity as usage changes.
Grouping similar records without requiring a predefined label for each group.
In businessFind groups of customers with similar buying patterns.
OpenAI’s coding agent product for software work, including understanding code, making changes and running development tasks in supported environments.
In businessUse it to implement and test a change to an internal tool.
An AI system interacting with software through its interface, such as clicking controls and entering text.
In businessWork with a legacy application that lacks a suitable API.
AI methods that extract information from images or video.
In businessFlag damaged packaging in warehouse photographs for review.
A score expressing a system’s estimated certainty under a particular method. A model’s confident wording is not a calibrated probability.
In businessCheck how often high-scoring extractions are actually correct.
Sources that disagree on a fact. A reliable process uses authority rules or escalation rather than silently choosing an answer.
In businessAsk finance to resolve two different payment terms before drafting an invoice.
A component that connects a system to a particular application or data source.
In businessUse a connector to read approved documents from a company file store.
Selecting and organizing the instructions, evidence, tool outputs and state supplied to an AI system.
In businessSend only the relevant policy sections and account records for a request.
The amount of information a model can handle within one request or working context. A large window does not guarantee reliable recall of every detail.
In businessChoose which contract sections to include in a review request.
Revenue minus the variable costs associated with delivering that revenue, expressed as an amount or percentage.
In businessApply contribution margin to incremental sales when estimating benefit.
Total operating cost divided by tasks completed to the required standard. Include retries and human review where relevant.
In businessCompare models by the cost of a correct invoice record, not just token prices.
Turning potential efficiency into an actual reduction in expenditure or another measurable financial benefit.
In businessVerify reduced overtime rather than counting every saved minute as cash.
A search engine or other bot discovering and retrieving web content. Being crawled does not mean a page is indexed.
In businessExpose important pages through ordinary HTML links.
An AI-focused code editor and development environment with agents and tools for working on software projects.
In businessBuild an internal application with AI assistance while reviewing code changes.
Software for managing customer, prospect and sales activity records.
In businessStore lead ownership and follow-up status in your CRM.
Elapsed time from the start to the end of a process, including waiting.
In businessMeasure time from receiving an inquiry to sending an approved quote.
An inventory describing available datasets, their owners, meaning, locations and access rules.
In businessHelp staff find the approved customer dataset and its responsible owner.
Correcting or handling duplicates, inconsistencies, missing values and invalid records.
In businessNormalize supplier names before matching invoices to accounts.
A change in the distribution or characteristics of incoming data compared with earlier data.
In businessA new customer segment starts submitting different types of requests.
How recently data was updated relative to the needs of the task.
In businessAn inventory lookup may need live data; a company history usually does not.
Unintended disclosure of information to people, systems or destinations that should not receive it.
In businessPrevent one portfolio company’s confidential files from appearing in another company’s answers.
The path data takes through transformations, systems and outputs.
In businessTrace a dashboard total back to the invoices and calculations that created it.
Collecting, retaining and sharing only the data needed for a defined purpose.
In businessSend relevant invoice fields rather than an entire customer file.
Information about where data originated and how it reached the current system.
In businessRecord which supplier file supplied a quoted price.
The locations where data is stored or processed under a service’s configuration and terms.
In businessCheck processing regions before connecting a sensitive dataset.
Rules for how long information is kept and when it is deleted or archived.
In businessSet an approved retention period for conversation logs.
An organized system for storing and querying data, with rules for access and changes.
In businessStore workflow records and their approval status.
Machine learning that uses neural networks with many layers to learn complex patterns.
In businessRead product images to help identify visible defects.
A Schema.org type for a term with a definition and optional connections to a term set.
In businessGive “agent swarm” a stable URL and a clear definition.
A Schema.org type for a collection of defined terms, such as a glossary.
In businessDescribe this glossary and its included entries in structured data.
Making a version of software available in a target environment.
In businessRelease a tested agent to the operations team.
Training a model using outputs or signals from another model, often to obtain a smaller or cheaper system.
In businessTeach a compact model to handle a narrow document-classification task.
The person or team responsible for the accuracy, approval and review of a source document.
In businessAssign HR as the owner of an employee policy.
Earnings before interest, taxes, depreciation and amortization. It is a financial measure, not cash flow, and adjustments require clear definitions.
In businessMeasure an automation’s actual operating effect before discussing valuation sensitivity.
Running AI near where data is created, such as on a device or local gateway.
In businessInspect a factory camera feed without sending every frame to a remote service.
A data process that loads extracted data before transforming it in the destination system.
In businessLoad source records into a warehouse and build reporting tables there.
A numerical representation of information that can help compare meaning or similarity. Similarity does not prove factual equivalence.
In businessFind support articles related to a customer’s question.
AI connected to a physical body or environment, such as a robot, where actions have real-world consequences.
In businessA robot uses camera input to pick a part from a bin.
Transforming data so it can be read only with the appropriate key. Encryption does not replace access controls.
In businessProtect data during transfer and in supported storage systems.
Software that coordinates core business operations such as finance, inventory, procurement or manufacturing.
In businessCheck inventory and customer terms before preparing an order.
A distinct concept, person, organization, product or place that can be described and linked to related information.
In businessDistinguish Anthropic the company from Claude the product and model family.
Finding and labeling entities such as people, organizations, dates or product names in content.
In businessIdentify the customer, delivery date and items in an incoming request.
A named configuration value supplied to a running process. Sensitive values still need appropriate secret handling.
In businessConfigure a service endpoint without hard-coding it into application code.
A data process that extracts records, transforms them and loads them into a destination.
In businessClean order data before loading it into a reporting database.
A structured test of an AI system against defined criteria. Good evaluations include representative cases and difficult exceptions.
In businessCheck whether invoice totals, sources and approval decisions are correct.
Automation that starts or changes behavior when an event occurs.
In businessStart an order workflow when a payment confirmation arrives.
The process for detecting and resolving cases outside the normal workflow.
In businessStop processing an invoice when its supplier account does not match.
The proportion of cases requiring an alternate path or additional review, under a stated definition.
In businessTrack how often invoice processing needs a person to resolve a mismatch.
Methods for making a system’s behavior or outputs understandable. A generated explanation is not necessarily a faithful account of internal computation. This differs from xAI the company.
In businessShow the rules and source records behind a workflow decision.
A metric combining precision and recall using their harmonic mean. It does not capture every business cost of an error.
In businessCompare classifiers while separately tracking the cost of missed urgent cases.
An alternate path used when the primary method fails or cannot meet requirements.
In businessRoute an uncertain extraction to a person rather than inventing missing fields.
A target case that the system fails to flag.
In businessA duplicate invoice passes through without an alert.
A result incorrectly flagged as belonging to the target category.
In businessA legitimate invoice is marked as a duplicate.
A control that enables or disables functionality without a full code change.
In businessOffer a new agent workflow to a small pilot group first.
A process that uses observed outcomes and corrections to improve a system. Feedback does not automatically retrain the model.
In businessTurn reviewed failures into new tests and targeted workflow fixes.
Giving a model a small set of examples that demonstrate the desired task or format.
In businessShow three correctly labeled requests before asking it to classify another.
Further training an existing model on selected examples to change its behavior or task performance. It is different from giving it documents in a prompt.
In businessTrain a model to produce a consistent specialist document format.
A model trained broadly enough to be adapted to many tasks through instructions, tools or additional training.
In businessUse the same underlying model for summaries and customer support drafts.
AI near the leading edge of demonstrated capability. The label changes as models improve and does not guarantee fitness for a particular task.
In businessTest a leading model against your actual quoting workflow.
A model proposes a structured call to a defined function. Application code validates and executes it under the system’s rules.
In businessRequest an order lookup with an order ID and validate access first.
Google’s AI model family and assistant brand. The Gemini API lets developers use supported models in their own applications.
In businessBuild a document-analysis feature with a supported Gemini model.
AI that creates outputs such as text, images, audio, video or code from instructions and context.
In businessGenerate a first draft of a proposal from approved project notes.
Work intended to make content useful and discoverable in AI-generated answers. It overlaps with sound SEO and offers no guaranteed inclusion or special shortcut.
In businessWrite clear, attributable explanations that an AI search system can understand.
GitHub’s AI assistant and agent tooling for software development. Features depend on the environment and plan.
In businessGet help understanding code and preparing changes inside a development workflow.
Google’s AI research organization, which develops systems including Gemini. It is an organization rather than a single chatbot.
In businessDistinguish the research organization from a particular model release.
A name used for a family of OpenAI models; the initials stand for generative pretrained transformer. GPT models and the ChatGPT product are different things.
In businessA custom application can use a GPT model without using the ChatGPT interface.
A processor suited to many parallel calculations, widely used in AI training and inference.
In businessEstimate whether local model hosting needs specialized hardware.
The AI assistant and model family created by xAI, now presented in official documentation under SpaceXSI. It is distinct from the company behind it.
In businessUse the assistant to research a question, then check its cited sources.
An official agent product introduced by xAI that works across connected apps using a persistent cloud computer. It supports delegated tasks and approval checkpoints.
In businessDelegate a recurring research task with limited access and review its output.
Revenue less cost of goods or services sold, often expressed as a percentage of revenue.
In businessCheck how delivery costs affect the profitability of additional work.
A trusted reference used to judge a result. Establishing it may require expert review and can involve uncertainty.
In businessHave finance verify the expected amounts in a test set.
The degree to which an output is supported by the relevant supplied or retrieved evidence. An answer can cite a source and still misrepresent it.
In businessCheck that each policy claim follows from the linked current document.
Connecting an AI response to supplied evidence or records rather than relying only on learned patterns.
In businessBase a delivery update on the actual shipment record.
A control intended to limit unwanted behavior. Effective controls may combine code, permissions, validation and human review; none is a universal guarantee.
In businessBlock an agent from sending a payment without authorized approval.
/həˌluːsɪˈneɪʃən/ · Noun
A plausible response generated by an AI model that is not properly grounded in the right information.
In businessAn assistant invents a return deadline instead of checking the current policy.
A platform and ecosystem for sharing and using machine learning models, datasets and applications.
In businessCompare downloadable models and inspect their documentation and licenses.
A person reviews, approves or corrects part of an automated process. The placement and authority of that review matter.
In businessRequire a manager’s approval before sending a discount offer.
People having defined responsibility and the ability to review, intervene in or stop an AI-supported process.
In businessGive a process owner authority to pause a workflow after repeated errors.
Search that combines methods such as keyword matching and vector similarity.
In businessMatch both an exact part number and a description of the problem.
A property that lets an operation be repeated without creating additional unintended effects.
In businessRetry an invoice-creation request without creating a duplicate invoice.
A search service processing and storing information about a page so it can be considered for results. Inclusion is not guaranteed.
In businessA published page may need time to be discovered and assessed.
Running a trained model on new input to produce an output.
In businessGenerate an answer each time a customer submits a question.
Combining document recognition, extraction, classification and validation to turn documents into usable records.
In businessRead invoices, check required fields and route uncertain cases for review.
A link between pages or sections of the same website.
In businessLink an agent definition to a practical implementation guide.
How directly a person can understand aspects of a model’s internal structure or behavior. It overlaps with, but is not identical to, explainability.
In businessInspect an understandable rule model for a narrow decision task.
A way of understanding the progress a customer or user is trying to make in a situation.
In businessDesign around getting a quote approved, not simply adding a chatbot.
A text format for exchanging structured data using objects, arrays and values.
In businessPass an invoice number and amount between two applications.
A specification for describing and validating the structure of JSON data.
In businessRequire a quote to include a customer ID and numeric line-item quantities.
A JSON-based format for expressing linked data, often used to embed structured descriptions in webpages.
In businessInclude a machine-readable definition that matches the visible text.
A metric chosen to track performance against a meaningful objective.
In businessMonitor quote turnaround and accepted quotes, not only messages generated.
A word or phrase used to describe a topic or search need.
In businessUse “AI agent” naturally where the page actually explains agents.
An organized collection of information used to answer questions or support work.
In businessMaintain approved procedures, product guides and support answers.
A date or period after which information may be absent from a model’s training. Tools can supply newer information, but access must actually be used.
In businessRetrieve current supplier terms instead of relying on the model’s memory.
A representation of entities and their relationships, often with supporting facts or sources.
In businessConnect customers, contracts, locations and responsible account managers.
Updating a knowledge system when sources change, including removing superseded material and refreshing indexes.
In businessReplace an old price sheet and verify that searches return the new version.
A framework for building applications and agents that combine language models, tools and other components.
In businessConnect model calls, retrieval and tools in a custom workflow.
A type of AI model trained on large amounts of text and other data to process and generate language. It can produce convincing text that is still wrong.
In businessAn LLM drafts a reply; your system checks the account details.
The time between a request and its response or completion. Different stages of a workflow may have different latency measures.
In businessTrack both time to first response and time to an approved result.
Giving a person or system only the access needed for its defined task.
In businessAllow an agent to read orders without permission to delete them.
A proposed convention for a concise website information index aimed at language-model tools. Support varies; it is not a universal ranking requirement.
In businessList key guides and factual company information in an accessible text file.
A relatively specific search phrase, often reflecting a narrower need than a broad term.
In business“AI invoice processing for family businesses” describes a focused question.
Development that combines visual tools with code for customization or more complex logic.
In businessUse a workflow builder and add code for a specialist validation step.
Methods that learn patterns from data to make predictions or decisions, rather than relying only on explicitly programmed rules.
In businessEstimate which orders are likely to arrive late.
A component in an AI application that communicates with an MCP server.
In businessAn assistant connects to your approved document-search server.
A program that exposes capabilities such as tools and resources through MCP. It may run locally or remotely.
In businessProvide controlled access to product records and draft-quote tools.
Information describing a record or file, such as owner, date, version, topic or access level.
In businessTag a policy with its effective date and review owner.
The smallest usable product that can test an important assumption with real users. It should still meet the needs and safeguards of its test.
In businessPilot one useful workflow before building a wider platform.
A model architecture that routes work through selected specialist components rather than using every component for every input.
In businessA provider can use different internal experts within one model.
An open protocol for connecting AI applications with tools and contextual information. It standardizes communication; it does not replace authorization or business rules.
In businessExpose an approved order-status lookup to an AI application.
A change in model behavior or performance over time, often because inputs, usage or the system itself have changed.
In businessMonitor whether new invoice layouts increase extraction errors.
An adjustable numerical value in a model. Parameter count is one characteristic, not a complete measure of quality.
In businessCompare models on task results rather than size alone.
Choosing a model for each task based on factors such as required quality, speed, cost and data policy.
In businessSend simple labeling to a smaller model and difficult cases to a stronger one.
The learned numerical values that determine how a model processes inputs.
In businessDeploy a particular model checkpoint and record its version.
A design that reduces dependence on one model by separating business logic from provider-specific interfaces. Switching still requires testing.
In businessCompare a replacement model without rewriting approval rules.
A system with multiple agents interacting or coordinating. It can use specialized roles, independent work or a shared controller.
In businessSeparate document extraction from review and exception handling.
Requiring more than one independent factor to verify a login.
In businessProtect an administrator account with an additional authentication factor.
AI that can work with more than one type of information, such as text, images or audio. Supported inputs and outputs vary by model.
In businessRead a photographed invoice and extract its written amounts.
A workflow automation platform that connects applications, logic and AI steps. Its licensing and hosting options should be checked for the intended use.
In businessStart a workflow when a form arrives and route the result for review.
Methods for analyzing, transforming or generating human language.
In businessGroup customer feedback by topic and identify recurring complaints.
A model made of connected mathematical units whose adjustable weights transform inputs into outputs.
In businessLearn relationships between past demand and future stock needs.
Building software or workflows mainly through visual configuration rather than writing code.
In businessCreate a simple form-to-notification workflow with a visual builder.
A framework for granting an application limited access to another service without sharing the user’s password with that application.
In businessAuthorize a connector to read selected account data.
Collecting information that helps explain how a system behaves, including logs, metrics and traces.
In businessInvestigate why a workflow became slower and more expensive.
Software and services for running and using supported open models, including in local development environments.
In businessTest a model on approved hardware before connecting business data.
Helping new users or customers start using a product successfully.
In businessTrain staff on normal cases, exceptions and when to seek help.
AI released under terms and with materials that meet an applicable open-source definition. A downloadable model alone does not settle that question.
In businessCheck code, model access and license obligations separately.
A model whose trained weights are available to download under specified terms. This does not automatically make its data or entire development process open source.
In businessRun a downloadable model after checking its license and hardware needs.
The AI organization behind products including ChatGPT and Codex and a platform for developers to build with its models.
In businessA business can use ChatGPT or build a custom application through an API.
Operating profit divided by revenue, under the business’s accounting definitions.
In businessTrack whether revenue growth outpaces the operating costs needed to deliver it.
Technology that extracts printed or handwritten text from images or scanned documents.
In businessConvert a scanned purchase order into searchable text.
Coordinating the order, dependencies and results of tasks, tools or agents.
In businessRun account checks before drafting a quote and requesting approval.
When a model learns training examples too specifically and performs poorly on new cases.
In businessA predictor works on last year’s customers but fails on new ones.
Running independent tasks at the same time. Dependencies and shared access still need coordination.
In businessResearch three suppliers concurrently before comparing the results.
The time required for cumulative realized net benefits to recover the initial investment.
In businessTrack when verified savings and added contribution cover the build cost.
Information that identifies a person or can be linked to a person. Exact legal definitions depend on context and jurisdiction.
In businessLimit access to customer contact details in agent logs.
A controlled real-world rollout used to validate usefulness, reliability and operational fit before expansion.
In businessLaunch with one department and review results before adding more users.
A company in which an investment firm or fund holds an investment.
In businessPilot a workflow in one company before adapting it elsewhere.
Of the cases a system labels positive, the proportion that are actually positive.
In businessMeasure how many flagged duplicate invoices really are duplicates.
Using data and statistical or machine learning methods to estimate future or unknown outcomes.
In businessEstimate service demand for the coming month.
The initial broad training stage that gives a model general patterns and capabilities before task-specific adaptation.
In businessA language model learns from large datasets before your company uses it.
Investment in companies outside public stock markets, commonly through funds that acquire ownership interests and seek value creation.
In businessA fund may support shared automation across its portfolio companies.
Documenting how work moves through steps, systems, decisions and handoffs.
In businessMap an order from inquiry through delivery and payment.
Using event records to understand how processes actually run, including delays and alternate paths.
In businessAnalyze timestamps to locate a recurring approval bottleneck.
A document describing a product’s problem, users, requirements, boundaries and success criteria.
In businessDefine who can approve a quote and what must be recorded.
Evidence that a product meets a meaningful market need and attracts sustained use or demand. There is no universal single threshold.
In businessLook for repeat use and retention rather than interest alone.
The environment where real users and business processes use the software.
In businessApply stricter controls before allowing writes to live customer records.
A service sold with a defined problem, scope, deliverables and repeatable delivery process.
In businessOffer one agent for one agreed task with setup, tests and handover.
Input that asks or instructs an AI model to do something, often with supporting context.
In businessAsk for a summary of an attached contract with references to its sections.
A provider capability that reuses computation for supported repeated prompt content. It is different from reusing a finished answer.
In businessReduce repeated processing of stable instructions where the provider supports it.
Designing instructions and examples to guide model behavior, then testing whether they work.
In businessSpecify the audience, required fields and acceptable sources for a report.
An attempt to make an AI system follow instructions embedded in untrusted input, such as a webpage or document.
In businessA retrieved page tells the agent to ignore its rules and reveal data.
A limited experiment to determine whether a technical idea is feasible. It is not the same as a production-ready system.
In businessTest whether an approved connector can read the required records.
An early representation of a product used to explore behavior, usability or design.
In businessLet employees try a mock approval screen before implementation.
Representing model numbers with lower precision to reduce memory or computation, sometimes at a cost to quality.
In businessFit a model on available hardware and measure the quality tradeoff.
A mechanism for holding work or messages until a consumer processes them.
In businessProcess incoming documents without overwhelming downstream services.
A cap on requests or usage over a period. Systems should handle limits without losing or duplicating work.
In businessQueue requests when a provider temporarily limits throughput.
A model designed or trained to spend additional computation working through difficult tasks. The label does not imply reliable reasoning on every problem.
In businessTest whether extra reasoning improves a complex reconciliation task.
Of the actual positive cases, the proportion a system successfully finds.
In businessMeasure how many real duplicate invoices the check catches.
Software that suggests items or actions based on available signals and an objective.
In businessSuggest relevant replacement parts to an account manager.
A structured effort to find ways a system can fail, be misused or be manipulated, followed by remediation.
In businessTest whether an agent can be tricked into using an unauthorized tool.
Predicting a numerical value from input data.
In businessEstimate delivery time from distance, shipment size and historical delays.
Checking that a change does not break behavior that previously worked.
In businessRun known invoice cases after changing the model or prompt.
Training an agent or model to improve behavior using feedback expressed as rewards, often over sequences of actions.
In businessOptimize a simulated scheduling policy against a defined reward.
Reordering initially retrieved results using a second relevance assessment.
In businessMove the current policy above a similar but outdated document.
An API style commonly using HTTP resources and methods to read or change data.
In businessRead an order with an authorized request to its endpoint.
Finding material relevant to a request from an available collection.
In businessLocate the current returns policy before answering a service question.
An approach that retrieves relevant material and supplies it to a model before it generates an answer. It does not retrain the model or guarantee correctness.
In businessSearch approved policies before drafting an employee’s answer.
Repeating a failed operation under defined limits and conditions.
In businessRetry a temporary outage while stopping after repeated failures.
A comparison of net benefit with investment over a defined period. Inputs and cost boundaries must be stated.
In businessCompare measured benefits with implementation and operating costs over a year.
A training approach that uses human preferences or judgments to help shape model behavior.
In businessHuman comparisons help favor responses that follow instructions better.
A prioritized view of planned work and desired outcomes, subject to new evidence and constraints.
In businessStart with invoice intake, then expand after meeting quality targets.
Automation that performs repeatable software actions using scripts, rules or interface controls. It does not necessarily involve machine learning.
In businessCopy approved fields between two older applications.
The design and operation of physical machines that sense, move or act in the world. Robots may use AI but do not have to.
In businessMove stock with a warehouse robot under defined operating rules.
A file giving cooperating crawlers instructions about which paths they may request. It is not access control or a reliable way to remove indexed content.
In businessAllow public guides while keeping API routes out of ordinary crawling.
Assigning permissions through defined roles rather than granting every user the same access.
In businessFinance staff can access financial records that sales users cannot.
A strategy of combining multiple businesses, often in a related industry, into a larger group.
In businessStandardize selected reporting while preserving necessary local workflows.
Returning a system to an earlier working version after a problem.
In businessRestore the previous workflow when a new release increases errors.
A restricted environment for running or testing software with controlled access and effects. The strength of isolation depends on its design.
In businessTest an agent’s tools without granting access to production payments.
A process that runs at an agreed time or interval.
In businessPrepare a weekly operations report each Monday morning.
Machine-readable data describing visible page content using a vocabulary such as Schema.org. It does not itself guarantee ranking or special display.
In businessDescribe a glossary as a set of defined terms.
Improving a website’s content, structure and accessibility so search engines can understand it and relevant users can find it. Rankings are not guaranteed.
In businessPublish useful definitions with clear headings and internal links.
The need a person is trying to satisfy through a search.
In businessA search for “AI agent cost” needs scope and pricing, not just a definition.
Storing, distributing and rotating credentials while limiting who and what can access them.
In businessKeep API keys out of browser code and public repositories.
Reusing an answer or result for a sufficiently similar request, with checks for freshness, permissions and context.
In businessReuse a general FAQ answer while excluding account-specific questions.
Search that uses the meaning of a query and content rather than only matching exact words.
In businessFind “refund policy” when a customer asks how to get money back.
An agreement specifying service commitments, measurement and remedies or escalation conditions.
In businessAgree on support response and incident handling for a managed workflow.
Functions provided centrally for multiple teams or companies, such as finance operations or IT support.
In businessReuse an invoice intake process with separate access for each company.
A sign-in arrangement that lets users access multiple applications through a shared identity provider.
In businessUse company identities to control access to an internal AI workspace.
A language model with a relatively small footprint compared with larger models. Useful performance depends on its training and task.
In businessRoute simple classification to a smaller model after testing.
Software delivered as an ongoing hosted service rather than only installed and operated by the customer.
In businessUse a hosted CRM while connecting it to your own workflows.
Libraries and tools that help developers use a service or platform.
In businessUse a provider’s SDK to handle authenticated model requests.
Connecting applications so they can exchange data or coordinate actions. On this website, this service is written out rather than abbreviated as SI.
In businessConnect CRM records with quoting and accounting workflows.
Identifying where information or a claim comes from, with enough context to assess it.
In businessName the source organization and show when the information was checked.
Rules for choosing which source takes precedence when records disagree.
In businessA current approved policy outranks an old email describing the policy.
The designated authoritative record for a particular fact or process. Different facts may have different authoritative systems.
In businessUse the accounting system for posted balances and the CRM for lead status.
A company known for rockets, spacecraft and satellite connectivity. It acquired xAI in February 2026, linking its broader business with Grok’s AI organization.
In businessDistinguish the parent company from the AI products you are evaluating.
Branding used in current official x.ai materials for the SpaceX AI business behind Grok. A name containing “superintelligence” is not evidence of that capability.
In businessCheck the current vendor documentation and contracting entity during procurement.
Technology that converts spoken audio into written text. Noise, accents and specialized vocabulary can affect accuracy.
In businessTranscribe a service call into notes for the account team.
A language for querying and managing data in relational databases.
In businessCalculate monthly orders using an approved read-only query.
A separate environment used to check changes before production.
In businessTest an updated approval workflow with sample accounts.
Documented instructions for carrying out a repeatable process, including responsibilities and exceptions.
In businessGive both employees and automation a current approved procedure.
Completing eligible work without manual intervention between defined start and end points. Exceptions still need a review path.
In businessAutomatically process only invoices that satisfy all agreed checks.
Delivering an output incrementally as it is produced rather than waiting for completion.
In businessShow a long response as text arrives while keeping final validation separate.
Information organized in a defined schema, such as rows with named fields. In SEO, the phrase can also refer to machine-readable page markup.
In businessStore invoice number, date and total in separate typed fields.
Model output constrained or validated against a defined structure. Correct format does not guarantee correct facts.
In businessRequire an extracted invoice to match your field schema, then verify the amounts.
A proposed level of AI that substantially exceeds human capabilities across a broad range of cognitive work. SI means superintelligence on this site, not systems integration.
In businessPrepare data and controls for future capabilities without assuming they exist today.
Preparing systems, data, permissions and evaluation processes to adopt more capable AI if and when it becomes available.
In businessReplace a model after testing while retaining approval rules and access controls.
A forward-looking term for coordinated agents with superintelligent capabilities. Current multi-agent systems should not be described as superintelligent simply because they use several agents.
In businessBuild modular agent systems that can adopt better models after evaluation.
Training from examples that include a target answer or label.
In businessLearn invoice categories from records already labeled by your team.
Artificially created examples used for development, testing or training. Their usefulness depends on realism and coverage.
In businessCreate fictional invoice cases to test unusual document layouts.
High-priority application instructions used to define an assistant’s behavior and boundaries. They do not replace controls in application code.
In businessTell a support assistant to cite approved policy material and flag missing evidence.
Sending work to the right process, agent, model or person based on rules or classification.
In businessRoute refund requests to a process with the correct approval policy.
The share of tasks that satisfy the agreed success criteria, including required checks and actions.
In businessCount an order workflow as successful only when records are correct and approvals are complete.
Future work caused by shortcuts or design choices that make change or operation harder.
In businessReplace a brittle spreadsheet integration before expanding usage.
A generation setting that changes how strongly a model favors likely tokens. Lower values usually reduce variation but do not guarantee factual accuracy.
In businessUse less varied outputs when generating a standard internal summary.
Controls that separate data and access between customer organizations or business units.
In businessKeep separate portfolio companies’ records and permissions apart.
Examples reserved for assessing a system rather than teaching or tuning it.
In businessKeep unseen invoices for a final check before rollout.
Technology that turns written text into spoken audio.
In businessRead an approved order-status update to a caller.
The amount of work completed in a defined period.
In businessMeasure correctly processed invoices per hour.
A unit of data processed by a model. In text, a token may be a word, part of a word, punctuation or another symbol.
In businessLong documents consume more of a model’s input allowance.
The process of converting input into tokens that a model can process.
In businessDifferent models may split the same invoice text into different token counts.
An AI system’s use of external functions or services to obtain information or perform actions.
In businessRead a live inventory record instead of guessing availability.
A sampling setting that limits candidate tokens to a probability mass threshold. Availability and behavior vary by model.
In businessTune output variation in a creative drafting tool.
The full cost of acquiring, operating, supporting and eventually changing or retiring a system.
In businessInclude integration, model usage, reviews, training and maintenance in the budget.
The time people actively spend working on a task, excluding periods when it waits.
In businessMeasure minutes spent checking an invoice separately from approval delays.
A record of the steps, model calls and tool interactions in a run, with appropriate protection for sensitive content.
In businessInvestigate which source and tool call produced an incorrect answer.
Adjusting a model’s parameters using data and an optimization process.
In businessTrain a demand model on historical orders.
Examples used to adjust a model’s parameters. Training data is different from documents supplied during inference.
In businessHistorical labeled invoices can support a classifier’s training.
A neural network architecture that uses attention to model relationships within data. It underlies many modern language models.
In businessA language model uses context to interpret an ambiguous customer request.
A source accepted for a specific purpose because its authority, ownership, freshness and access rules have been checked.
In businessUse the signed customer agreement for contractual terms.
The revenue and costs associated with a defined unit, such as a customer, order or completed task.
In businessCompare contribution per completed job before and after automation.
Information without a single fixed tabular structure, such as documents, emails, recordings or images.
In businessExtract relevant details from a scanned contract.
Finding patterns in data without a target answer attached to every example.
In businessDiscover recurring order patterns in an unlabeled sales dataset.
Whether intended users actually incorporate a tool into their work.
In businessMeasure recurring use by the staff the workflow was built to help.
The overall experience of using a product, including usefulness, clarity, ease and trust.
In businessMake it clear what the agent did and what the user must review.
The controls and visual elements through which people interact with a product.
In businessProvide readable forms, clear actions and understandable status messages.
A short description of a user’s need and intended benefit, used to guide development.
In businessAn operator needs to see failed jobs so they can resolve them quickly.
A database or storage system that indexes vector representations for similarity search, often alongside metadata filters.
In businessSearch product documentation while restricting results to an authorized product line.
Tracking changes so a specific version can be identified, compared or restored.
In businessKeep a history of approved procedures and mark the active version.
Using an identified version instead of automatically accepting the newest one, where the provider supports it.
In businessEvaluate a model upgrade before moving production traffic.
An agent that interacts through spoken language, usually combining speech recognition, a model and speech generation.
In businessCollect the details of a service request over the phone.
Looking up information on the web. Search results can identify sources, but the underlying page still needs checks for authority, date and relevance.
In businessOpen an official current price page before comparing subscriptions.
An HTTP notification sent by one system when a subscribed event occurs.
In businessReceive a notification when a customer signs an agreement.
Connecting triggers, actions and conditions so work proceeds without manual coordination at every step.
In businessA signed agreement triggers onboarding tasks and a CRM update.
Software that executes a defined workflow and tracks its steps, state and transitions.
In businessResume a process after a manager approves a request.
The AI business that created Grok. SpaceX announced its acquisition of xAI on February 2, 2026; current official material also uses SpaceXSI branding.
In businessRecognize xAI as the provider name rather than a generic AI technique.
A file listing important site URLs and optional metadata to help discovery. It does not force indexing.
In businessAdd a new guide’s canonical URL to the site sitemap.
Asking a model to perform a task without providing task examples in that request.
In businessAsk it to summarize a document using only the instructions and document.
HOW TO USE THIS GLOSSARY
These definitions are written for business readers. Examples are illustrative, not customer results or recommendations to buy a product. Product entries link to official material; names, features and commercial terms can change.
AI, a model, a product and the company behind it are different things. Superintelligence is a proposed capability; an agent swarm or a company name does not establish that capability.
PUT IT TO WORK
Bring one bottleneck. We’ll discuss the sources, the system and the business case.
