Explaining AI to the people who sign off on it.
Name the job first. Then the simplest thing that does it.
Fig. 1 — Five of the seven terms sit inside each other. RAG and agents are missing on purpose: they describe how a model is wired into your business, not a bigger kind of model. Tap a region.
“AI” covers everything from a chess engine to an email drafter. When a vendor says it, ask which of the circles inside they actually mean.
Two of the terms are wiring, not models.
Retrieval and agents connect a model to your documents, tools and permissions. That is where most of the business value sits, and most of the security risk.
Software that handles tasks we would normally say need judgement.
A chess engine and an email drafter both count.
Before you signWhich of the six terms below are we actually buying?
Where it bites“AI-powered” on a vendor slide says nothing about where your data goes or who can see it.
Learns patterns from past data, then predicts, scores or classifies.
Your bank flagging a card payment as unusual.
Before you signIs our history good enough, and representative enough, to predict from?
Where it bitesSkewed or stale data gives confident wrong answers. Models drift as the world moves.
A branch of machine learning built on layered neural networks, suited to messy input like speech, images and text.
Meeting audio turned into a transcript.
Before you signCould we explain an outcome if a customer or the regulator asked?
Where it bitesAccurate is not the same as explainable, and explaining is now part of the job. See below.
Produces new text, images, audio or code on request.
A first draft of a board paper.
Before you signWho checks the draft against the brief and the evidence before it leaves the building?
Where it bitesWhat staff paste in. Shadow AI is a data-handling problem before it is a technology one.
Generative models trained on vast amounts of text. They produce plausible language, which often reads as reasoning.
Summarising a report or comparing two contract drafts.
Before you signFluent is not correct. What gets verified, and by whom?
Where it bitesThey can follow instructions hidden in the content they read. Untrusted input is untrusted instructions.
Fetches relevant documents first, then has the model answer from them.
A policy assistant answering from this week’s HR policies.
Before you signAre the sources current, and does it respect who is allowed to see what?
Where it bitesIt can surface documents a user should never see. A poisoned source poisons the answer.
A model given a goal, tools and permissions. It plans the steps and takes the actions.
Finding a receipt, lodging the claim, chasing the approval.
Before you signWhat may it do alone, what needs a person, and how do we switch it off?
Where it bitesEvery permission is attack surface. An agent that reads email and can pay invoices is one planted instruction from fraud.
Understanding what an agent can do is only the beginning. The Intent Gap examines how consequential actions can be independently constrained, while the Overlap Map addresses who is accountable for authorising and assuring those controls.
From 10 December 2026, APPs 1.7 to 1.9 of the Privacy Act require an organisation’s privacy policy to say when a computer program uses personal information to make, or substantially help make, decisions that could significantly affect someone’s rights or interests.
The OAIC’s guidance, released on 30 September 2026, reads this broadly. Machine learning and generative AI used in significant decisions are likely in scope unless there is extensive human oversight. A rubber-stamp review probably does not count.
OAIC: transparency for AI and automated decisions ↗
A field guide for leaders, not legal advice.
So the leader check for every term above has a second half: if this shapes a decision about a person, can we describe it in plain words in our privacy policy?
Then agree what it may touch, who checks its work, and which decisions stay with people. That conversation is worth more than any definition on this page.