AI automation glossary: plain definitions.

In AI projects, the same word can mean different things to different people. When we work together, I use these terms with the meanings below.

Agents and automation

AI agent

Software that plans and runs several steps towards a goal on its own, using tools. On this site, the agent prepares and a person approves.

Workflow automation

A workflow that starts on a trigger and follows predefined steps. Less flexible than an agent, but predictable.

Human in the loop

A critical step does not run until a person has checked and approved it. Used for everything that goes outside.

Shadow mode

A trial mode where the agent runs but its output is not used. Its results are compared with the work people do.

Autonomy level

How much work an agent can do without a person. It grows step by step: shadow mode, suggest, limited and supervised authority.

Prompt

The instructions and context given to an AI model. A good prompt states the role, the task, the constraints and the output format.

Prompt library

Versioned, tested prompt templates for a team's recurring tasks.

RAG (retrieval-augmented generation)

The model finds relevant passages in company documents before answering, bases its answer on them and cites the source.

Embedding

A vector of numbers that represents the meaning of a text. Used to find texts with similar meaning.

Context window

The amount of text a model can take into account at once.

Hallucination

A model producing fluent but wrong or unsourced information. Source-based answers and human approval reduce it; they cannot remove it entirely.

MCP (Model Context Protocol)

An open protocol that lets AI applications connect to tools and data sources in a standard way.

A2A (Agent2Agent)

An open protocol that lets different AI agents discover each other and request work.

Exception queue

The list where work the automation cannot handle lands, owned by a named person. Recurring exceptions are reviewed to improve the process.

Measurement and method

Baseline

The measurement taken before a change. Results are discussed against this value.

Outcome vs activity metric

An activity metric counts work done (drafts written); an outcome metric shows its effect on the business (shorter proposal time).

Acceptance threshold

A measurable limit, written up front, that AI output must meet to be accepted; for example the share of drafts approved without edits.

Drift

Output quality degrading over time as the model, data or usage changes. Regular measurement catches it.

Cycle time

Total time from the start to the end of a task, waiting included.

Manual touches

How many times a task passes through human hands before it is done. Fewer touches mean a simpler flow.

First time right (FTR)

The share of work completed correctly on the first attempt, without rework.

OEE (overall equipment effectiveness)

Availability × performance × quality for a machine. Shows where losses occur in production.

8D

An eight-step method for solving quality problems: team, description, containment, root cause, corrective action, verification, prevention and closure.

Pilot trap

A pilot that works technically but never goes into use, because it has no owner, metric or budget.

Lean

A way of working that measures and removes steps that add no value for the customer (waiting, rework, unnecessary movement). It comes from the Toyota Production System.

Privacy, security and compliance

Prompt injection

An attempt to steer an agent through instructions hidden in a document or email it reads.

Excessive agency

The risk of giving an agent more access or tools than its task needs.

Data minimisation

Using only the data fields a task actually needs.

Masking and pseudonymisation

Hiding personal details such as names and phone numbers, or replacing them with codes, before processing.

Local model

An AI model running on the company's own server, so data does not leave the company.

Vendor lock-in

A system so tied to one provider or person that nobody else can maintain it.

Audit log

A tamper-proof record of every step taken by the agent and its users, so it can be traced later.

DPIA

Data protection impact assessment: a written assessment of risks and safeguards before high-risk processing of personal data.

EU AI Act risk classes

Prohibited, high risk, limited risk (transparency obligations) and minimal risk. Obligations depend on the intended use, not the technology.

ISO/IEC 42001

The AI management system standard. It defines how an organisation governs AI, its risks and its responsibilities.

OWASP Top 10 for LLMs

An open list of the ten most important security risks specific to large language model applications.

Visibility

AI search visibility (GEO)

Work that helps assistants such as ChatGPT, Gemini and Perplexity recognise a person or business correctly and recommend them.

llms.txt

A short text file at the site root that gives AI tools a summary of the site and its key pages.

Structured data (Schema.org)

Marking up information such as people, services and Q&A on a page in a format machines can read.

sameAs

Links in structured data that point to a person's or organisation's profiles on other sites, helping them be recognised as one identity.

Evidence ladder

A scale showing an expert's level of online evidence step by step, from just a name and title up to named case studies and references.

Results matter more than terms.

Most of these terms are covered in more detail on the lean AI automation, data privacy and assurances pages.