AI agent glossary

Deployed AI

Deployed AI is the practice of building AI agents inside a company's real workflow and running them there. The team that builds the agent stays to measure it, fix it and hand it over.

Forward-deployed engineering

Forward-deployed engineering puts the engineer inside the customer's operation instead of behind a product. The engineer learns the workflow first hand and builds what that workflow needs.

AI agent

An AI agent is software that reads inputs, decides what to do next and takes an action within limits someone set. Unlike a script, it handles cases nobody wrote a rule for, which is why it needs an approval step and an eval.

Agentic workflow

An agentic workflow is a business process where agents do the steps that used to need a person reading and deciding. People stay at the points where judgement or sign-off matters.

Human-in-the-loop

Human-in-the-loop means a named person reviews or approves an agent's work before it takes effect. It is the deployment pattern that lets agents do real work in finance and operations without taking control of it.

Approval gate

An approval gate is the point in a workflow where an agent stops and waits for a person to accept, edit or reject its work. We design the gate before we build the agent.

Eval

An eval is a repeatable test that measures how well an agent does its job, such as the share of bank lines matched correctly. If an agent has no eval, nobody can say whether it works.

Held-out period

A held-out period is a stretch of real data, such as last month's books, that the agent never saw while it was built. Testing on it shows how the agent will do on data it has not met before.

Audit log

An audit log is a permanent record of every action an agent took, with its inputs, its output and who approved it. Auditors and controllers read it to see exactly what happened and why.

Company OS

A company OS is a company's standard procedures written down as a versioned layer that agents read and follow. Changing a procedure changes how every agent behaves, with a record of the change.

MCP (Model Context Protocol)

The Model Context Protocol is an open standard for connecting AI models to tools and data sources through one common interface. A system that speaks MCP can be used by any agent that supports the protocol.

Retrieval-backed copilot

A retrieval-backed copilot is an assistant that answers from a company's own documents and systems and cites the source it used. Access control decides which documents each person's questions can reach.

Agentic commerce

Agentic commerce is shopping done partly or wholly by AI agents that search, compare and sometimes buy on a person's behalf. For a brand it rewards clean product data more than clever marketing.

Agentic storefront

An agentic storefront is a store whose catalog, schema and feeds are structured so shopping agents can read and act on them. Shopify uses the term for its channel that makes merchant catalogs available to AI agents.

Answer engine optimisation (AEO)

Answer engine optimisation is the work of making a page the source an AI assistant quotes when it answers a question. It rewards direct answers, clear structure and facts with a source.

Generative engine optimisation (GEO)

Generative engine optimisation is the wider work of shaping how generative AI systems describe a brand across many answers. It covers what the models learn about a company from the whole web, not only from its own site.