The terms, in plain language

An agentic workflow is a business process in which an AI agent carries out some of the steps, deciding what to do next based on what it finds. It fits work that follows a pattern but needs judgment along the way, and it needs clear limits on what the agent may access and do.

Where it comes up: Agentic workflow automation , Generative AI, agentic AI and intelligent automation

See also: AI agent , Least privilege

An AI agent is a software program that uses a language model to decide which steps to take toward a goal, and that can call tools such as search, databases or other systems to carry them out. It matters because it can handle a multi-step task instead of only answering a question, which makes its permissions and oversight important design choices.

Where it comes up: Agentic workflow automation , Generative AI, agentic AI and intelligent automation

See also: Agentic workflow , Human in the loop

AI evaluation is the practice of testing an AI system against a set of realistic questions or tasks with known good answers, and scoring the results. It turns a hunch that the system works into evidence, so a team can see whether a change to the prompt, data or model made it better or worse.

Where it comes up: MLOps and LLMOps , Knowledge assistants , MLOps, LLMOps and AI platform engineering , Quality engineering and reliability assurance

See also: Groundedness , LLMOps

An AI gateway is a single service that sits between an application and the language models it calls, handling routing, access control, logging and limits in one place. It keeps those concerns out of each application and makes it easier to switch or add models later.

Where it comes up: AI features in your product , MLOps and LLMOps , MLOps, LLMOps and AI platform engineering

See also: Model routing

An application programming interface, or API, is a defined way for one piece of software to ask another for data or an action, without needing to know how the other works inside. APIs let separate systems, such as a website and a billing tool, work together through a stable, documented contract.

Where it comes up: Systems integration , Web apps and portals , Digital product engineering and app development

See also: Data contract

An architecture decision record is a short written note that captures one important technical decision, the options considered and the reasons for the choice. It preserves the thinking behind a system, so people who join later can understand why it was built that way.

Where it comes up: MVP to production , Legacy modernization , Technology advisory and solution architecture

See also: Technical debt

Build versus buy is the decision between writing custom software and adopting an existing product or service for a need. The right answer depends on how central the need is to the business, what each route costs over time and how much control is required.

Where it comes up: MVP to production , Technology advisory and solution architecture

See also: Total cost of ownership

Change data capture is a technique that detects inserts, updates and deletes in a database as they happen and passes them on to other systems. It lets a copy of the data stay current without reloading everything, which supports reporting and integrations that depend on fresh data.

Where it comes up: Real-time data , Systems integration , Data engineering and modern data platforms

See also: Event streaming

Chunking is the step of splitting long documents into smaller passages before they are indexed for search by an AI system. Good chunks are small enough to be relevant and large enough to keep their meaning, which strongly affects whether the right passage is found.

Where it comes up: Knowledge assistants , Generative AI, agentic AI and intelligent automation

See also: Embedding , Retrieval-augmented generation

Container orchestration is the automated management of containers, which are lightweight packages that bundle an application with what it needs to run. It handles starting, scaling, updating and replacing them across many machines, to help applications stay available as demand and releases change.

Where it comes up: Cloud foundation , DevOps automation , Cloud, DevOps and platform engineering

See also: Infrastructure as code

Continuous integration and delivery, known as CI/CD, is a practice in which every code change is automatically built and tested, and approved changes are released through a repeatable pipeline. It reduces manual steps and can make releases routine.

Where it comes up: DevOps automation , Cloud, DevOps and platform engineering

See also: GitOps , DevSecOps

A data contract is an agreement between the team that produces data and the teams that use it, describing the fields, formats, meaning and expectations for that data. It makes changes visible before they break the reports or integrations that depend on it.

Where it comes up: Systems integration , Data engineering and modern data platforms

See also: Application programming interface , Change data capture

Data governance is the set of rules, roles and practices that decide who owns data, who can use it and how it is kept accurate and protected. It matters because people can only trust and safely share data when responsibility and access are clearly defined.

Where it comes up: Modern data platform , Data governance, cybersecurity and responsible AI

See also: Data lineage , Data quality

A data lake is a central store that holds large amounts of raw data in its original form, structured or not, at low cost. It gives analysts and data scientists one place to draw from, though it needs good organization so it does not become a pile of unusable files.

Where it comes up: Modern data platform , Data engineering and modern data platforms

See also: Lakehouse , Data warehouse

Data lineage is a record of where data came from and every step that changed it on the way to a report or model. It lets people trace a surprising number back to its source, and it helps check what a change would affect before it is made.

Where it comes up: Modern data platform , Data governance, cybersecurity and responsible AI , Data engineering and modern data platforms

See also: Data governance

A data mart is a focused subset of an organization's data prepared for one team or subject, such as finance or sales. Because it is organized around the questions that team asks, reports built on it can be quicker to create and easier to understand.

Where it comes up: Reporting and dashboards , Business intelligence, analytics and decision support

See also: Data warehouse , Semantic layer

Data mesh is an approach in which individual business domains own and publish their data as products, supported by shared platform tools and common standards. It suits larger organizations where one central data team has become a bottleneck.

Where it comes up: Modern data platform , Data engineering and modern data platforms

See also: Data lineage , Lakehouse

A data pipeline is an automated series of steps that moves data from its sources, cleans and reshapes it, and delivers it where it is needed. It replaces manual exports and copying, so reports and applications receive consistent data on a schedule or in real time.

Where it comes up: Modern data platform , Real-time data , Data engineering and modern data platforms

See also: ETL and ELT , Data quality

Data quality describes how accurate, complete, consistent and timely data is for the purpose it serves. Checking it as data flows through a system catches problems early, before they reach reports and decisions.

Where it comes up: Modern data platform , Reporting and dashboards , Data engineering and modern data platforms , Quality engineering and reliability assurance

See also: Data pipeline , Data governance

A data warehouse is a database designed for analysis, holding cleaned and organized data from many sources in a consistent structure. It gives a business a single reliable place to answer questions across the whole organization.

Where it comes up: Modern data platform , Reporting and dashboards , Data engineering and modern data platforms , Business intelligence, analytics and decision support

See also: Lakehouse , Data mart

DevSecOps is a way of working that builds security checks into every stage of software development and delivery, rather than adding them at the end. It helps teams find and fix weaknesses earlier, when they are easier to correct.

Where it comes up: DevOps automation , Cloud, DevOps and platform engineering , Data governance, cybersecurity and responsible AI

See also: Continuous integration and delivery

An embedding is a list of numbers that represents the meaning of a piece of text, an image or other data, so that similar items end up with similar numbers. Embeddings let software find content by meaning rather than by exact words, which is the basis of semantic search.

Where it comes up: Knowledge assistants , Generative AI, agentic AI and intelligent automation

See also: Vector search , Chunking

ETL and ELT are two patterns for moving data into an analytics system, either extract, transform and load or extract, load and transform. The difference is whether data is reshaped before it is stored or after it arrives, which affects flexibility, speed and cost.

Where it comes up: Modern data platform , Data engineering and modern data platforms

See also: Data pipeline

Event streaming is the continuous capture and delivery of events, such as clicks, payments or sensor readings, as they occur. It lets systems react within moments instead of waiting for a nightly batch of data.

Where it comes up: Real-time data , Data engineering and modern data platforms

See also: Change data capture , Event-driven architecture

Event-driven architecture is a way of designing software in which components announce that something happened, such as an order being placed, and other components react to it. It keeps systems loosely connected, so parts can change or scale without being rewired to each other.

Where it comes up: Real-time data , Data engineering and modern data platforms , Digital product engineering and app development

See also: Event streaming

Fine-tuning is the process of training an existing AI model further on a smaller, specialized set of examples so it behaves in a particular way or style. It can help when prompting and retrieval are not enough, but it adds cost and upkeep, so it should be used only where justified.

Where it comes up: AI features in your product , Generative AI, agentic AI and intelligent automation

See also: Foundation model

FinOps is a practice that brings engineering, finance and business teams together to understand and manage the cost of cloud usage. It makes spending visible and accountable, so cost decisions are made alongside decisions about performance and speed.

Where it comes up: Cloud cost optimization , Cloud, DevOps and platform engineering

See also: Rightsizing , Total cost of ownership

A foundation model is a large AI model trained on broad data that can be adapted to many different tasks, such as writing, summarizing or answering questions. A business AI feature can start from one of these models instead of training a new model from scratch.

Where it comes up: AI features in your product , Knowledge assistants , Generative AI, agentic AI and intelligent automation

See also: Large language model , Fine-tuning

GitOps is a way of managing infrastructure and application deployments in which the desired state is stored in a Git repository and applied automatically. Changes can then be reviewed, recorded and reversed, like any other change to code.

Where it comes up: DevOps automation , Cloud, DevOps and platform engineering

See also: Continuous integration and delivery , Infrastructure as code

Groundedness is how well an AI system's answer is supported by the source material it was given, rather than by guesses. Measuring it helps a team spot made-up statements and decide whether an assistant can be trusted for a given task.

Where it comes up: Knowledge assistants , MLOps and LLMOps , Quality engineering and reliability assurance

See also: AI evaluation , Retrieval-augmented generation

Human in the loop is a design in which a person reviews, approves or corrects an automated system's work at defined points. It keeps accountability with people for decisions that carry risk, while still saving effort on routine steps.

Where it comes up: Agentic workflow automation , Document intelligence , Data governance, cybersecurity and responsible AI

See also: AI agent

Hybrid search combines keyword search, which matches exact words, with vector search, which matches meaning, and merges the results. It can find relevant content that either method alone would miss, such as when a query includes both a product code and a general idea.

Where it comes up: Knowledge assistants , Generative AI, agentic AI and intelligent automation

See also: Vector search , Reranking

Infrastructure as code is the practice of defining servers, networks and other cloud resources in version-controlled files that tools apply automatically. It makes environments repeatable and reviewable, so they can be rebuilt or copied without manual setup.

Where it comes up: Cloud foundation , DevOps automation , Cloud, DevOps and platform engineering

See also: GitOps , Landing zone

Intelligent document processing is the use of software, often including AI, to read documents such as invoices or forms and pull their key information into structured data. It reduces manual re-keying and can route unclear cases to a person to check.

Where it comes up: Document intelligence , Generative AI, agentic AI and intelligent automation

See also: Structured output

A key performance indicator, or KPI, is a measure chosen to show how well a business is doing against a specific goal. A small set of clearly defined KPIs gives teams a shared view of progress and keeps dashboards focused on what matters.

Where it comes up: Reporting and dashboards , Business intelligence, analytics and decision support

See also: Semantic layer

A lakehouse is a data platform that combines the low-cost, flexible storage of a data lake with the organized, reliable querying of a data warehouse. It lets analytics and machine learning work from one copy of the data instead of separate systems.

Where it comes up: Modern data platform , Data engineering and modern data platforms

See also: Data lake , Data warehouse

A landing zone is a pre-configured, secure foundation in a cloud account, covering things like account structure, networking, access rules and logging. It gives new workloads a consistent starting point, so each team does not have to design the basics again.

Where it comes up: Cloud foundation , Cloud, DevOps and platform engineering

See also: Infrastructure as code , Least privilege

A large language model, or LLM, is an AI model trained on very large amounts of text that can understand and generate human language. It powers features such as drafting, summarizing and question answering, and it can produce fluent answers that are wrong, so its output needs checking.

Where it comes up: AI features in your product , Knowledge assistants , Generative AI, agentic AI and intelligent automation

See also: Foundation model , Retrieval-augmented generation

Least privilege is the security principle that a person, program or AI agent should have only the access it needs to do its job, and no more. It limits the damage if an account is misused or taken over.

Where it comes up: Cloud foundation , Agentic workflow automation , Data governance, cybersecurity and responsible AI

See also: Secrets management

LLMOps is the set of practices for deploying, monitoring and improving applications built on large language models. It covers concerns such as prompt versions, evaluation, cost and safety, so these applications stay reliable after launch.

Where it comes up: MLOps and LLMOps , MLOps, LLMOps and AI platform engineering

See also: MLOps , AI evaluation

A minimum viable product, or MVP, is the simplest version of a product that real users can use and learn from. It tests the core idea early, so a team invests further only in what proves useful.

Where it comes up: MVP to production , Digital product engineering and app development

See also: Proof of concept

MLOps is the set of practices for taking machine learning models from experiments to dependable production systems and keeping them working. It covers deployment, monitoring and retraining, so a model keeps delivering value after its first release.

Where it comes up: MLOps and LLMOps , MLOps, LLMOps and AI platform engineering

See also: LLMOps , Model drift

Model drift is the gradual or sudden loss of a machine learning model's accuracy as real-world data or behavior changes from what the model was trained on. Monitoring for it tells a team when a model needs to be retrained or replaced.

Where it comes up: MLOps and LLMOps , MLOps, LLMOps and AI platform engineering

See also: MLOps

Model routing is the practice of sending each request to the most suitable AI model, based on factors such as the task, the quality needed, speed and cost. It lets an application use a capable model only when needed and a lighter one for simpler requests.

Where it comes up: AI features in your product , MLOps and LLMOps , Generative AI, agentic AI and intelligent automation

See also: AI gateway

Observability is the ability to understand what is happening inside a system by examining the logs, metrics and traces it produces. It helps teams find the cause of a problem quickly and can help spot trouble before users notice it.

Where it comes up: DevOps automation , Cloud foundation , Cloud, DevOps and platform engineering , Ongoing engineering and platform support

See also: Service-level indicators and objectives

Prompt injection is an attack in which hidden or malicious instructions in text, a web page or a document trick an AI system into ignoring its intended rules. It matters most when the system can take actions or reach private data, so access should be limited and inputs treated as untrusted.

Where it comes up: Agentic workflow automation , Knowledge assistants , Data governance, cybersecurity and responsible AI

See also: Least privilege

A proof of concept, or PoC, is a small, time-boxed build that tests whether an idea is technically feasible and worth pursuing. It reduces risk by answering the key question before a larger commitment is made.

Where it comes up: Agentic workflow automation , Real-time data , Technology advisory and solution architecture

See also: Minimum viable product

Recovery time objective is how quickly a service must be restored after an outage, and recovery point objective is how much recent data a business can afford to lose. Together they set the targets that backup and disaster recovery plans are designed to meet.

Where it comes up: Cloud foundation , Cloud, DevOps and platform engineering

See also: Service-level indicators and objectives , Observability

Reranking is a second pass in search that takes the top results from a first search and reorders them by how well they match the question. It helps put the most relevant passages first, which shapes the quality of an AI assistant's answer.

Where it comes up: Knowledge assistants , Generative AI, agentic AI and intelligent automation

See also: Hybrid search

Retrieval-augmented generation is a way of answering questions with a language model by first searching a trusted set of documents and giving the model the relevant passages as context. It keeps answers tied to your own information and makes it possible to show the sources behind them.

Where it comes up: Knowledge assistants , AI features in your product , Generative AI, agentic AI and intelligent automation

See also: Embedding , Vector search , Groundedness

Rightsizing is the practice of matching cloud resources, such as server size or storage, to what a workload actually needs. It reduces waste from oversized resources, provided performance needs are measured first.

Where it comes up: Cloud cost optimization , Cloud, DevOps and platform engineering

See also: FinOps

Secrets management is the secure storage and handling of sensitive values such as passwords, API keys and tokens that software needs to run. It keeps them out of code and shared files, controls who can read them and makes them easy to replace.

Where it comes up: Cloud foundation , DevOps automation , Data governance, cybersecurity and responsible AI

See also: Least privilege

A semantic layer is a shared set of business definitions, such as what counts as revenue or an active customer, placed between raw data and reporting tools. It makes every report calculate the same measure the same way.

Where it comes up: Reporting and dashboards , Business intelligence, analytics and decision support

See also: Key performance indicator , Data mart

A service-level indicator is a measurement of how a service is performing, such as the share of requests that succeed, and a service-level objective is the target set for it. Together they turn reliability into something a team can track and discuss in plain numbers.

Where it comes up: Cloud foundation , DevOps automation , Ongoing engineering and platform support

See also: Observability , Recovery time and recovery point objectives

Structured output is a response from an AI model in a fixed, machine-readable format, such as the fields of a form, instead of free-flowing text. It lets other software use the result directly and makes the output easier to check automatically.

Where it comes up: Document intelligence , AI features in your product , Generative AI, agentic AI and intelligent automation

See also: Intelligent document processing

Technical debt is the accumulated cost of shortcuts and outdated choices in software that make future changes slower and riskier. Like financial debt, it can be a reasonable trade-off at first, but it needs to be managed or it keeps growing.

Where it comes up: Legacy modernization , Digital product engineering and app development , Technology advisory and solution architecture

See also: Architecture decision record

Total cost of ownership is the full cost of a technology over its life, including build or purchase, running, support and eventual replacement, not just the price at the start. Comparing options on this basis helps avoid choosing something that looks cheap today but costs more later.

Where it comes up: Cloud cost optimization , Technology advisory and solution architecture

See also: Build versus buy , FinOps

Vector search finds items by comparing embeddings, which are numeric representations of meaning, rather than matching exact words. It lets a search for one phrase return content that says the same thing in different words.

Where it comes up: Knowledge assistants , Generative AI, agentic AI and intelligent automation

See also: Embedding , Hybrid search