How does an AI agent work?
An AI agent works in a loop: it receives a goal, decides the next step, uses a tool to take it, checks the result and repeats until the goal is met or a limit stops it. In a business setting, the loop has six parts:
- Goal and instructions. People define what the agent should achieve, the rules it must follow and when it must stop or hand the task to a person.
- Reasoning and planning. The language model reads the input, works out what is being asked and decides the next step.
- Tools and data access. The agent calls the tools it has been given, such as reading a document, searching the ERP or querying the CRM, through an account with limited permissions.
- Actions. It does the work: fills in a field, drafts a reply, creates a record, routes a request. It can take only the actions you allowed.
- Checks and human approval. The agent checks its output against rules and expected formats, and anything that touches money, a customer or a legal document waits for a person to approve it.
- Memory and logging. It keeps the context of the task in progress, and every run is recorded with its inputs, steps and outputs, so any error can be traced.
A back-office example: a supplier emails an invoice to the accounts inbox. The agent reads the PDF, extracts the supplier, the amount and the order number, finds the purchase order in the ERP and compares the two. If they match, it prepares the invoice for booking and queues it for the accountant to approve. If they do not, it sends the invoice to the buyer with a note on the difference. The accountant works through a short queue of checked invoices instead of an inbox. What a task like this costs to run is worked through in how much AI implementation costs.
What is the difference between an AI agent, a chatbot and automation?
The difference is who decides the next step and whether the software can act on your systems: a chatbot answers, workflow automation follows a fixed path, a copilot suggests and an AI agent decides and acts within limits. The word "agent" is now attached to almost every AI product, which makes it hard to know what is being sold. Four categories cover most of what companies are offered.
| Type | Who decides the next step | Can it act on your systems | Typical use |
|---|---|---|---|
| Chatbot | The user, one question at a time | No, it only answers | Answering questions from a website or a knowledge base |
| Workflow automation | A fixed rule written in advance | Yes, along a fixed path | Copying an order from a form into the ERP |
| Copilot | A person, with suggestions from the model | Only when the person accepts | Drafting an email or a summary inside a tool your team already uses |
| AI agent | The model, within limits you set | Yes, through the tools and permissions you give it | Reading an inbound email, finding the order, drafting the reply and routing it for approval |
In practice the categories blend. A good agent is usually a workflow with a few decision points handed to a model, not a free-roaming assistant. The more of the path you can fix in advance, the more reliable the result. We treat autonomy as something a workflow earns, one step at a time, after it has been measured.
Where agents fit in a company of 10 to 250 people
The tasks that work share three traits: they repeat often, their inputs are written (emails, documents, tickets, records), and someone in the company can say what a correct result looks like. Five examples we see most often:
- Inbound email triage. The agent reads each message to a shared inbox, identifies the request type and the customer, attaches the relevant order or contract and routes it to the right person, or drafts a reply for the common cases.
- Document extraction. Invoices, delivery notes, contracts and forms arrive as PDFs or scans. The agent pulls the fields you need into the ERP or a spreadsheet and flags the documents it could not read with confidence. We cover this case in AI document processing for invoices, delivery notes and contracts.
- Drafting replies from a knowledge base. The agent searches your policies, product sheets and past answers, drafts a reply with the sources it used, and a person approves or edits it before it is sent.
- Ticket classification. Support or internal IT tickets are labelled by type, urgency and owner as they arrive, so the queue is sorted before anyone opens it.
- Report preparation. The agent gathers figures from the systems that hold them, fills the recurring report and writes a first commentary, which the owner of the report reviews and signs.
None of these replaces a team. Each removes a slice of manual, repeated work from people who have better uses for their time, and each keeps a person at the point where the outcome matters.
What agents do badly today
Language models are good at reading and writing and uneven at everything that depends on context they were never given. Four kinds of task are a poor fit today, whatever the vendor demo shows:
- Open-ended judgment. Deciding whether to extend credit to a difficult customer, or how to handle a sensitive complaint, depends on relationships and history the agent cannot see. It will produce a confident answer anyway.
- Rare processes. A task done a few times a year has too few examples to test against and too little volume to repay the build. A checklist for a person is cheaper and safer.
- Tasks with no definition of a correct result. If two experienced people in your company would disagree on the right output, there is nothing to measure the agent against, so there is no way to know whether it works.
- Irreversible actions without approval. Sending payments, deleting records, signing or sending contracts, changing prices. An agent may prepare these actions. A person should approve them.
Before an agent goes to production: the checklist
A demo proves that an agent can do a task once. Production means it does the task every day, on real data, with a way to catch it when it is wrong. These are the six conditions we put in place before any agent goes live:
- An evaluation set. Real examples from your process, each with the correct answer, agreed before the build. Every change to the prompt, the model or the workflow is run against it and the result is reported as a number.
- Human approval on high-impact actions. Anything that touches money, a customer or a legal document is prepared by the agent and approved by a named person, inside the tools that person already uses.
- Permissions and least privilege. The agent gets its own account with access only to the data and actions its task needs. Read access where reading is enough; no administrator rights.
- Logging. Every run is recorded with its inputs, the steps taken, the tools called and the output, so any error can be traced and turned into a new evaluation case.
- Cost limits. Model usage is billed per request. We estimate the cost per task from your real volumes, set a budget ceiling with alerts and choose the smallest model that passes the evaluation set.
- Data location and GDPR. You decide which data the agent may see, where it is processed, how long it is kept and which provider terms apply. Personal data is minimized before it reaches a model, and the decisions are written down for your DPO.
If a vendor cannot show you these six for an agent they propose, you are buying a demo. Our AI readiness checklist covers the conditions on the company side.
What the EU AI Act changes for agents
Most business agents, such as the five examples above, are not high-risk under the EU AI Act. Two points still matter. First, the Act's transparency rules apply from 2 August 2026: AI systems that interact directly with people must be designed so that those people are informed they are dealing with an AI system (European Commission, Quick Facts on transparency rules). A customer-facing agent that answers email or chat should say so.
Second, some uses are listed as high-risk, including AI used in recruitment and selection of people and in assessing the creditworthiness of individuals (AI Act, Annex III). Following the 2026 AI Omnibus amendments, the Commission states that rules for these areas apply from 2 December 2027. An agent that screens candidates or scores customers is a different project, with obligations to check with counsel before it is built. The Act was amended in 2026, so read the current text rather than older summaries.
When to start with process optimization instead
An agent automates a process. If the process is not defined, the agent automates the confusion. We recommend starting with process optimization when the steps differ from one person to the next, when nobody owns the process end to end, when the data it needs still sits in chat threads and personal spreadsheets (moving it into one structured place is then the first part of the project), or when a simpler rule-based automation would do the same job with no model at all.
This is the question our Diagnosis is built to answer: which processes deserve an agent, which need fixing first, and which should be left alone. Sometimes the honest answer is that no agent belongs there yet. When an agent does fit, our AI integration service builds it with the checklist above from the first week.
Summary
- An AI agent uses a language model to pursue a goal, deciding its next step and acting through tools within limits set by people; a chatbot answers, workflow automation follows fixed rules, a copilot suggests.
- It works in a loop: goal and instructions, planning, tool calls, actions, checks and human approval, logging.
- Agents fit repetitive, written work with a clear correct result: email triage, document extraction, drafted replies, ticket classification, report preparation.
- They fit badly where judgment is open-ended, the process is rare, correctness is undefined or the action cannot be undone.
- Production needs an evaluation set, human approval, least privilege, logging, cost limits and written data decisions.
- If the process itself is unclear, fix the process first.
