Glossary · 74 terms
AI-native work glossary
Plain-language definitions of the terms leaders meet in AI transformation: forward-deployed engineers, AI discovery, agents, skills, harnesses, evals, shadow mode, cost per task and more.
A
- Agent experience (AX)Agent experience (AX) is how easily AI agents can understand and work inside your company: finding the right information, knowing who owns what, and using tools safely. It is the agent counterpart of user experience.
- Agent identityAn agent identity is an account of its own for an AI agent, separate from any person’s login, so its access can be limited to what its tasks need, reviewed, revoked and audited.
- Agent inboxAn agent inbox is one place where people see and answer everything their background agents need from them: notifications, questions and actions waiting for approval.
- Agent loopThe agent loop is the cycle an AI agent runs to do a task: the harness gives the model the context, the model chooses the next step, the harness carries it out and feeds back the result, and the cycle repeats until the task is done or a stop rule applies.
- Agent traceAn agent trace is the step-by-step record of one AI agent run: what came in, what the agent saw, each model call and tool call it made, what it produced, what it cost, how long it took, what failed and where a person stepped in.
- Agent washingAgent washing is selling an existing product, such as a chatbot, an AI assistant or robotic process automation, as an AI agent when it lacks real agentic capabilities.
- Agentic workflowAn agentic workflow is a workflow in which some steps are handled by AI agents that make judgment calls, alongside steps run by plain code and decisions kept with people.
- AGENTS.mdAGENTS.md is an open format for standing instructions to AI coding agents: a markdown file in a project that tells any agent how to work on it, like a README written for agents.
- AI acceptable use policyAn AI acceptable use policy sets out which AI tools people may use for work, what data each tool may handle, and how to ask for a new tool to be approved.
- AI agentAn AI agent is software that uses an AI model to work toward a goal by deciding on steps and taking actions, such as reading documents, calling tools or updating a system, instead of answering a single prompt.
- AI discoveryAI discovery is the work of finding where AI belongs in a company: which workflows, which steps within them, and what each change is worth. Done well, it is based on how work actually happens, not on what people remember in a workshop.
- AI governanceAI governance is the set of policies, controls and records that decide which AI tools, models and data may be used, what agents are allowed to do, and who approves what.
- AI opportunity assessmentAn AI opportunity assessment is a structured review, often run by consultants, that identifies where AI could create value in a company, scores the options and produces a prioritized roadmap.
- AI-nativeAn AI-native company designs its work around AI. Agents handle routine steps, people focus on judgment, review and improvement, and what works is captured as shared skills, SOPs and evals that keep getting better.
- Ambient agentAn ambient agent is an AI agent that listens to a stream of events, such as incoming emails, and acts on them without waiting for a prompt, involving a person only when it needs to. It is LangChain’s name for a kind of background agent.
- Audit trailAn audit trail is a record of what was done, by whom or by which agent, when, and with what inputs, kept so that any action can be reviewed and explained later.
B
- Background agentA background agent runs on its own when an event or schedule triggers it, does a defined job end to end, and only asks a person when it needs a decision. It is the opposite of a chat assistant you prompt step by step.
- BaselineA baseline is a measurement of how a workflow performs before a change, such as its cycle time, quality, cost per task and people’s time, taken with the same measures you will use afterwards.
- BenchmarkA benchmark is a fixed set of tasks used to score and compare performance. Public AI benchmarks test general skills; a personalized benchmark tests your own work.
- Bring your own AI (BYOAI)Bring your own AI (BYOAI) is when people use AI tools at work that their employer didn’t provide, such as a personal chatbot account or a browser extension they installed themselves.
- Build vs. buyBuild vs. buy is the decision between building a custom AI solution, buying an existing product, or doing neither and keeping the work as it is.
C
- Computer useComputer use is an AI agent’s ability to operate software the way a person does: it looks at the screen, moves the pointer, clicks and types to finish a task, even in apps that offer no other connection.
- Conformance gapThe conformance gap is the difference between how a process is documented, in an SOP or a process map, and how the work actually runs.
- Context engineeringContext engineering is deciding what information an AI model sees for each task, such as instructions, examples, documents and tool results, so it has what it needs and nothing that distracts it.
- Cost per taskCost per task is the full cost of getting one task done correctly with AI: the model usage across every attempt, plus the time people spend reviewing the output and fixing its mistakes.
- Cycle timeCycle time is the total time from the start of a process to its end, including the waiting between steps.
D
- Digital adoption platform (DAP)A digital adoption platform (DAP) is software that adds help on top of a company’s applications, such as tooltips, step-by-step walkthroughs and checklists, and shows where people get stuck.
- Digital orgA digital org is a living, machine-readable map of how your company works: people, roles, workflows, systems, documents and rules. Both people and AI agents use it to find their way around.
- DriftDrift is a change in how an AI agent performs after go-live, caused by something changing underneath it: the model, a prompt, a tool, a policy, or the data and cases it sees.
E
- Escalation ruleAn escalation rule says when an AI agent must stop work on a case and hand it to a named person: for example, when figures conflict, an amount crosses a threshold, or the agent fails after a set number of attempts.
- EvalAn eval (evaluation) is a repeatable test that checks an AI system’s output against what “correct” means for a task, using real cases and answers or grading rules the team agrees on.
- Event logAn event log is the record a business system keeps of what happened to each case: which activity took place, and usually when and by whom. It is the raw material for process mining.
- Exception loopAn exception loop is when a case is sent back to an earlier step, because of missing data, a failed check or a new approval, creating rework and delay.
F
- Forward-deployed engineer (FDE)A forward-deployed engineer (FDE) is an engineer who embeds with a customer’s team, learns how the work actually gets done, and builds and deploys software inside that team’s real workflows. Palantir pioneered the role.
- The four bucketsThe four buckets are a way to sort every step of a workflow before automating it: delete it, run it as plain code, give it to an agent, or keep a person on it.
- Frontier modelA frontier model is one of the most capable, and usually most expensive, AI models from the leading labs. It is used for hard reasoning, ambiguous judgment and long, multi-step work.
G
- Graduated autonomyGraduated autonomy means giving an AI agent more responsibility in steps as it proves itself on real work: from shadow mode, to suggesting while a person approves, to acting with spot checks, to acting alone on low-risk cases. It moves back down when quality slips.
- GuardrailsGuardrails are the automatic checks and hard limits around an AI agent that stop it from doing what it must never do, such as sending a message, moving money or editing a system of record, whatever the model decides.
H
- HarnessA harness is the code and setup around an AI model that turns it into a working agent: its instructions, tools, memory and context, permissions, retries, logging and checks.
- Human baselineA human baseline measures how well people do the same tasks an AI system is tested on, with the same cases and the same definition of correct, so the two can be compared fairly.
- Human-in-the-loopHuman-in-the-loop is a design in which a person reviews, approves or decides at chosen points in an AI-run process, usually where mistakes would be costly or hard to undo.
J
K
L
M
- MCP (Model Context Protocol)MCP, the Model Context Protocol, is an open standard that lets AI assistants and agents connect to tools and data sources, such as Slack, Google Drive or a CRM, through one common interface. Anthropic introduced it in November 2024.
- Moments of needMoments of need are the points at which people need help to do their work. The 5 Moments of Need framework, from Conrad Gottfredson and Bob Mosher, names five: learning something new, learning more, applying it, solving a problem and adapting to change.
O
- ObservabilityObservability is being able to see what AI agents and people are doing in live work: what each agent did and with which inputs, how well it performed, what it cost, and where a person stepped in.
- On-screen assistantAn on-screen assistant is an AI assistant that sees what is on your screen in the app you are using and helps in the moment, rather than in a separate chat window.
- Open-weight modelAn open-weight model is an AI model whose trained weights are published, so you can run it on your own infrastructure or through a provider of your choice.
P
- Personalized benchmarkA personalized benchmark scores AI on your own work: your real tasks, judged by your team’s own definition of correct, and compared with how your people do the same work.
- PRD (product requirements document)A PRD, or product requirements document, defines what to build and why: the problem, who it is for, the requirements, and what is out of scope.
- Process mappingProcess mapping is the work of recording how a workflow runs, step by step: who does each step, in which system, what they decide, and where the work waits, loops back or changes hands.
- Process miningProcess mining reconstructs how a process actually runs, its real steps, loops and delays, from the event data in business systems rather than from how it is documented.
- Progressive disclosureProgressive disclosure is the way AI assistants load skills in stages: first only each skill’s name and description, then the full instructions when a task matches, and supporting files or scripts only when the instructions call for them.
- Prompt injectionPrompt injection is an attack in which instructions hidden in content an AI reads, such as a web page, an email or a document, try to make it ignore its user and do something else.
R
S
- SandboxA sandbox is an isolated environment that limits what an AI agent can touch while it works, such as which files it can change and which networks it can reach.
- Shadow AIShadow AI is AI tools and agents that employees use for work without the company’s knowledge or approval.
- Shadow ITShadow IT is software, devices or cloud services that people use for work without the IT department’s knowledge or approval. Shadow AI is the AI form of it.
- Shadow modeShadow mode means running an AI agent alongside the real work without letting it send or change anything, then comparing its output with what people actually did.
- Sidekick agentA sidekick agent is an AI agent that works with a person in real time, inside that person’s tools: the person asks, the agent answers or acts, and the person steers and decides at every turn.
- Skill (AI skill)An AI skill is a packaged, reusable set of instructions, and sometimes files or scripts, that teaches an AI assistant or agent how to do a specific task the way your team does it.
- Small modelA small model is a compact AI model that is cheaper and faster than a frontier model. It handles well-defined tasks such as classification, extraction and routing, often as well as a much larger model.
- SOP (standard operating procedure)An SOP, or standard operating procedure, is a written, step-by-step description of how a task should be done.
- Staggered rolloutA staggered rollout gives a new tool or process to teams in waves, in an order set in advance, so the teams still waiting can serve as a comparison group.
- Stop ruleA stop rule is a limit that ends an AI agent’s run or pauses it for a person, such as a maximum time, number of steps or spend, or a condition such as two figures that disagree.
- Straight-through processingStraight-through processing is when a case, such as an invoice or an order, goes from start to finish without manual handling or exceptions. Its rate is a key measure of process efficiency.
- Swimlane diagramA swimlane diagram is a process map divided into lanes, one for each person, team or system, so you can see who does each step and where the work passes from one to another.
- System of recordA system of record is the authoritative system where a type of business data lives, such as a CRM like Salesforce or HubSpot for customers, or an ERP like NetSuite or SAP for financials.
T
- Task miningTask mining records what people do on their computers, such as clicks, keystrokes and data entry, and analyzes it to show how tasks are really done, including the work between the steps that business systems log.
- ThroughputThroughput is how many cases a process completes in a given period, such as invoices matched per week or customer issues resolved per hour.
- TokenA token is the unit of text an AI model reads and writes, roughly three-quarters of an English word. Providers price usage per million input tokens and per million output tokens.
- Total cost of ownership (TCO)Total cost of ownership (TCO) is the full cost of a system over the years you run it: what it takes to build or buy it, plus what it takes to run, check, maintain and eventually replace it.