AI Glossary / AI and machine learning / Beginner
Action
Definition
In AI, an action is anything an agent can do that changes its environment or its relationship to it: moving, sending, writing, booking, buying. Actions are how an agent's decisions become real.
What it means
The agent loop ends in action. Perceiving and deciding are private; acting is where the system touches the world. A thermostat's action is switching the heating; a software agent's action might be sending an email or updating a record.
Actions define what an agent is. Two systems with identical intelligence but different action sets are entirely different tools: one that may only draft text is a writing aid; the same system with the power to send is a spokesperson.
Every action has three business properties: reversibility, cost and visibility. Sending a draft to a reviewer is reversible and cheap; paying an invoice is neither. Mapping an agent's action set along these lines is basic safety work.
The set of available actions is a design choice, not a technical given. Permissions, API connections and tools decide the action set, and narrowing it is the cheapest risk control there is.
For a business owner, the action inventory answers the only question that matters about a new AI system: what can it actually do, and to what?
How it works
In the standard model, the agent's decision process selects an action, and actuators execute it: motors for a robot, API calls for software. The environment then changes, and the next percept reflects the result.
Action selection respects constraints: which actions exist, which are permitted in the current state, and what each costs. These constraints are where safety rules live in code.
Effects are uncertain in the real world, so careful designs confirm outcomes: did the payment clear, did the record update, did the email send? The confirmation becomes part of the agent's next percept.
Logging every action with its context turns a system's behavior into an auditable record, which is the foundation for trust, debugging and accountability.
Three practical examples
- Retail: a pricing agent's action changes a listed price; a mistaken one is visible to every customer at once.
- Operations: a scheduling agent's action books an appointment; a double-booking is a real person kept waiting.
- Finance: a reconciliation agent's action posts an adjustment; a wrong one moves money in the accounts.
Think of it this way
Think of the difference between a consultant and a signatory. The consultant advises and the advice can be ignored; the signatory's signature moves money. Giving an AI system actions is making it a signatory, however junior.
The limit: a junior signatory understands shame and consequences. An agent understands neither, so the signing rules have to carry all the caution the person would supply.
Worked example
inventorying one agent's actions
Suppose a firm deploys a customer-service agent. Before launch, the team lists every action the system can take: read tickets, read account data, draft replies, send replies, issue refunds, change addresses.
Each action is scored for reversibility, cost and visibility. Drafting is free and reversible; refunds move money; address changes are reversible but fraud-sensitive. Permissions follow the scores: drafting is free, sending is template-bound, refunds cap at a small amount, address changes always need a person.
The list is reviewed monthly, because new integrations quietly add actions. The scenario is invented; the action inventory with three-way scoring is the practice.
When an incident happens, the first question is which action fired, and the inventory makes the answer findable.
Illustrative case study
A fictional company connects an AI assistant to its email, calendar and payments in one enthusiastic afternoon. A confused sequence of tool calls later, a supplier receives three identical purchase orders.
The post-mortem finds no bug: the agent had the actions and used them within its instructions. The fix is to narrow the action set to read-only plus drafts, and to re-add sending and ordering one at a time with caps.
The invented lesson: every action you connect is a commitment. Connect them one at a time, with limits, and earn each expansion with clean logs.
Common misconceptions and mistakes
- Thinking is the risky part. Thinking is private; actions are where damage happens, which is why permissions matter more than prompts.
- More tools mean a more capable agent. They also mean a larger blast radius; add actions as they are earned, not because they exist.
- Logs are optional. The action log is the only reliable account of what the system did when something goes wrong.
Business use
Before deploying any AI system, write its action inventory: what it can do, to which systems, with what reversibility, cost and visibility.
Narrow the set deliberately: read-only first, reversible drafts next, money-moving last, each behind caps and approvals, with every action logged.
People also ask
Is generating text an action?
Internally it is output; it becomes an action when someone or something can act on it, which is why sending is a separate permission.
Who answers for an agent's actions?
The deploying organization, exactly as for a junior employee's acts within their authority.
What is the first safety step?
Shrink the action set: remove every capability the current use case does not strictly need.
Check your understanding
An assistant connected to email, calendar and payments issues three duplicate purchase orders. What is the root design failure?
Answer: The action set was wider than the use case required; payments should have been added last, behind caps and approvals, after read-only and draft stages proved clean.
Related terms
Learn next
AI Agents and Automation for Small Business covers permission design and approval steps for agent actions: https://aiforbusinessandlife.com/academy/ai-agents-automation-small-business. [C4]
More in AI and machine learning
- Ablation - An ablation is a controlled experiment that removes one part of a model or system and measures what changes, to learn how much that part actually contributes.
- Abstractive Summarization - Abstractive summarization is generating a summary in new words that capture the meaning of the source, rather than selecting and copying the source's own...
- Accumulated local effects - Accumulated local effects, or ALE, describe how a fitted model's predictions change with a feature by averaging local prediction differences and...
- Action model learning - Action model learning is the process of learning descriptions of what actions require and how they change an environment.
- Active learning - Active learning is a machine-learning approach where the system chooses which examples it wants labeled next, instead of learning from a fixed or random set.
- Adam - Adam is the most widely used algorithm for training neural networks: an optimization method that adjusts each of the model's millions of internal numbers...
Sources
[S1] Russell and Norvig, AIMA glossary. Full URLs are in the source ledger. Verified 11 October 2026.
Previous: Acquiescence bias Next: Action cost function