The short answer
An AI agent is a program that uses a language model to reach a goal: it understands the request, chooses among the tools it has been given, uses them, checks the result and replies. Unlike a chatbot, it does more than talk: it searches a database, reads a document, fills in a form. In a company it works with a limited set of tools, clear permissions and a person who confirms the important actions.
- Agent = language model + tools + goal + rules.
- A chatbot answers; an agent acts on company systems.
- Read-only tools are safe; those that write ask for confirmation.
- Checks that can be done with a rule are done with a rule.
- It pays off most where people today search, copy, check and re-enter information.
The definition, without jargon
An AI agent, that is, an artificial intelligence (AI) agent, is a program built around a language model, with four ingredients:
- A goal, expressed in natural language: “find the contracts that are about to expire and prepare the list for the sales team”.
- Tools, meaning functions it can call: searching an archive, reading a file, querying an external service, creating a record.
- A loop: the model decides which tool to use, reads the result and decides the next step, until the goal is reached or it stops to ask.
- Rules: what it can do, on whose behalf, with which confirmations and with what log.
Take away the tools and you are left with a chatbot. Take away the rules and you are left with a risk.
Agent, chatbot, automation: the differences
| What it can do | Chatbot | Traditional automation | AI agent |
|---|---|---|---|
| Understands requests in plain language | Yes | No | Yes |
| Reads unstructured documents | Partly | No | Yes |
| Acts on company systems | No | Yes, always in the same way | Yes, choosing the route |
| Handles unforeseen cases | Answers in words | Gets stuck | Adapts or asks |
| Needs checks and confirmations | Not much | Not much | Yes, always |
It is not a contest: in well-built systems the three coexist. The agent is given what requires understanding; traditional rules are given what requires certainty.
How an agent works: a real example
In the CRM we use every day an operator writes “Add this new customer” and pastes in a VAT number.
- The agent uses the search tool to check that the customer does not already exist.
- With a second tool it searches the web for the company details and turns them into the fields of the customer record.
- It shows the summary to the operator and waits for an “OK, create it”.
- Only after the confirmation does it use the tool that creates the customer. If the VAT number already exists, a rule written in code blocks the operation, whatever the model says.
Four steps, two tools that read and one that writes, one confirmation, one hard rule. The result is a customer record filled in within a few minutes, with no duplicates. On the customer’s page, the same work can be done starting from a check of the VAT number on the European VIES service.
What they can do in a company
- Sales: find customers and contracts with a single sentence, prepare draft quotes, summarise a customer’s history before a visit, flag renewals at risk.
- Purchasing: compare price lists and reference prices, check indexation clauses, prepare requests for quotation.
- Administration (accounts): read orders, order confirmations and documents in PDF, pre-fill accounting entries, check consistency.
- Support and quality: answer on the basis of manuals and cases already solved, classify incoming reports.
- Senior management: answer questions about company data and market data in the same conversation, as in our FT CEO.
The six rules for using them safely
- A limited, described set of tools. The agent does not “get into the system”: it has a closed list of functions.
- Reading is free, writing is an exception. The tools that change data are few and ask for confirmation.
- Everything starts as a draft. What the agent creates is reviewed by a person.
- Same permissions as the user. AI is not a shortcut for getting round roles.
- Checks in code. Duplicates, formats, limits, calculations: what can be verified with a rule is not left to a model.
- A log. Requests, tools used and outcomes all stay on record.
To these, add one design rule: what is irreversible stays with people. Creating credentials, sending payments, deleting data: the agent takes the user to the right screen and stops.
What it costs and how long it takes
It depends on the process, not on the technology. There are three cost items: the initial project, usually a fixed-scope pilot lasting a few weeks; model usage, which is paid for as it is used and for office tasks is generally a modest item; and maintenance, because processes and systems change. The right way to assess it is to compare the cost with the time given back to people and with the errors avoided, measured before and after.
Where to start
With a single process, well chosen. We have set out the full path in the guide How to bring AI into your company, and the way we apply it on the page AI agents in business processes.
Sources
- Regulation (EU) 2024/1689 (AI Act), Article 3: definition of AI system; Article 50: transparency towards people
- European Commission, guidelines on the definition of an AI system (February 2025)
Checked on 12 October 2026. Rules and deadlines change: always check the official source before you decide.