What is an AI agent? A plain answer for the SME office

What is an AI agent?

An AI agent is a language model that has been given tools and a goal, and works towards that goal in steps. It decides what to look up or do next, uses a tool, reads the result and carries on until the goal is met or it gets stuck.

A chat window in the browser answers a question and stops. An agent acts inside the systems it has access to. In a small business office that can mean sorting an inbox, preparing a reply or finding an answer spread across three systems.

Anthropic, which builds such models, draws the line this way: workflows are systems where models and tools run through predefined code paths, while agents dynamically direct their own processes and tool usage. In short, in a workflow you fix the path in advance, and in an agent the model chooses it.

• ModelReads and writes

  Understands the email, the note, the question in plain language.

• ToolsWhat it may touch

  Read the mailbox, search the customer list, create a draft.

• GoalWhen it is done

  For example: every new request sorted and a draft prepared.

How an agent differs from chat and workflow

Three terms get mixed up in everyday speech. Chat means you ask, the model answers and the exchange ends. Workflow means the same chain of steps on every run, often without the model choosing the route. Agent means the goal is fixed and the path can change from case to case, because the model picks the next tool.

Digitalisation, automation and AI explains the difference between a rule, a workflow and a model with a supplier invoice. Which process in your firm deserves attention first, agent or not, is the subject of which process to automate first.

What an AI agent can do in an SME office

An agent is useful where the next step depends on what the previous one found. Three jobs fit that description in many offices.

• Triage: the agent reads new mail in the shared info mailbox, decides whether it is an order, a complaint or a question, and assigns it to the right person with a one-line summary.
• Drafting: for a customer question it looks up the last order and the price list and prepares a reply that someone reads before it goes out.
• Looking things up across systems: «Where is the delivery for this customer?» means checking the order list, the shipping confirmation in the mailbox and a note in the CRM. The agent collects the three answers and writes one.

In all three the path changes from case to case. That is where a model choosing its own steps earns its cost. Copilot in Microsoft 365 or similar assistants in your suite can help with single steps. An «agent» in the narrower sense ties several tools to one goal without every intermediate decision being scripted.

What an AI agent should not do unattended

An agent should not make payments, send binding replies, delete data or take any action that no person checks. Binding means anything a customer can hold you to: a price, a delivery date, an acceptance of terms.

The OWASP list of risks for language-model applications calls this «excessive agency»: damaging actions that follow from unexpected, ambiguous or manipulated model output. It names too many functions, too many permissions and too much autonomy as the causes, and recommends that a human approves high-impact actions. Agents that read mail face a further risk: text inside an email or a file can carry instructions that the model then follows. OWASP lists this as indirect prompt injection.

Data and shadow AI in a Swiss office

Agents need access to real data: mail, customer lists, contracts. Customer and internal figures do not belong in free chat accounts without a contract and clear retention rules. Swiss data protection law expects appropriate technical and organisational measures, and which tools your firm allows is a business decision for management.

If staff already use consumer chat tools for office work, you do not have an agent programme, only shadow AI with no audit trail. An agent run by the firm therefore needs boundaries: which systems, which accounts, who approves, what gets logged.

Who should wait before buying «an agent»

Not every SME needs an agent to make AI useful. If your problem is a fixed path with few exceptions, a rule or simple workflow is often enough. If only one person holds IT access, fix the basics first. If the job comes up once a quarter, setup and review cost more than the benefit.

The mistake: an agent for a job a fixed rule does reliably

The most common mistake is using an agent where a fixed rule already does the job every time. «Every PDF from this supplier goes into this folder and to this approver» has one correct path. A rule follows it at almost no cost per run, and afterwards you can see why it did what it did. An agent adds variation exactly where you want none, and each run costs more.

If your team wants to work with agents

If the team first needs to understand what agents can do, try them on real tasks and agree on rules for their use, the AI in Business workshop is the format for that, including how to chain several tools into agentic workflows. It is a half or full day for up to 20 people.

If you already have one specific task in mind, such as the inbox that gets sorted by hand every morning, the workshop is the wrong format. That one task is what the Process Check is for.