An artificial intelligence agent is a system that uses a language model to decide which steps to take and which tools to use until it meets a goal. Unlike a chatbot, which answers, or a classic automation, which follows fixed rules, the agent observes the situation, picks an action, checks the result and repeats until it finishes or asks a person for help. It makes sense when the inputs vary and call for judgement; it does not when the process is always the same.
Chatbot, automation and agent: three different things
The three terms get mixed up in sales presentations, and that leads to buying the wrong tool. The difference is not in the technology underneath but in who decides the next step.
| Tool | Who decides the next step | Useful when | Example |
|---|---|---|---|
| Chatbot | The user, with each question | There are frequent questions and the answer is in your documents | "What is the return period?" |
| Classic automation | Nobody: the rules are written in advance | The process is always the same and the data arrives in order | When an order comes in, create the invoice and send it |
| Agent | The model, within set limits | Inputs vary or arrive messy and require judgement | Read a quote request and prepare the offer |
Classic automation remains the best option when the process is predictable: it is cheaper, faster, deterministic and easy to audit. If you can draw the process as a diagram with three or four branches and it never leaves them, you do not need an agent. An agent adds value when there is variability: incomplete requests, messages in natural language, documents in different formats or decisions that an experienced person makes today.
Anthropic distinguishes between workflows (the code fixes the order of the steps and the model fills in each one) and agents (the model decides the order). Their recommendation, which I share, is to always start with the simplest solution that works and add autonomy only when the problem demands it.
How an agent works inside
Strip away the marketing and an agent is a loop with five parts:
- Goal. What it has to achieve, written in plain language: "prepare a draft quote from this email".
- Context. What it needs to know: the current price list, lead times, examples of previous quotes, the company's rules.
- Tools. Actions it can run: search the CRM, read a PDF, look up a price sheet, create a draft email. Each tool has its own limited permissions.
- Decision and execution. The model picks a tool, uses it, reads the result and decides the next step.
- Stop. It finishes when it meets the goal, when it detects that something is missing or when it hits a limit on steps or cost. At that point it hands the result to a person.
Point 5 is the one most often neglected. An agent without clear stop conditions can go round in circles, burn tokens and produce rubbish with great confidence. In my projects I always set a maximum number of steps, a budget per task and an explicit rule for saying "I cannot do this, passing it to a person".
A concrete example: quotes in an industrial company
Imagine a company that manufactures custom parts. Every day it receives between ten and twenty quote requests by email. Some come with a PDF drawing, others with a photo, others with a paragraph of approximate measurements. Today one person spends two or three hours a day reading them, asking for what is missing, looking up prices and writing the offer.
An agent for that process would do this:
- Read the email and attachments and extract material, measurements, quantity and desired lead time.
- Check against a list of mandatory fields. If something is missing, draft a reply asking for exactly that piece of information and stop.
- If everything is there, look up the price list and the volume discount rules and calculate the price, showing the working.
- Search the history to see whether that customer has asked for something similar before and flag it.
- Leave a draft quote in the system with a summary of what it has assumed. A person reviews, corrects if needed and sends.
What it does not do, at least in the first phase: send anything to the customer without review, change prices in the price list or commit to factory lead times. Those limits are defined before a single line of code is written.
Cases that make sense in an SME
- Preparing quotes and offers. Reads the request, organises the data, detects what is missing, proposes a price with its calculation and leaves a draft for approval.
- Classifying and answering enquiries. Tells sales, support and billing apart, prepares the reply from your documents or routes it to the right person.
- Extracting data from documents. Turns supplier invoices, delivery notes or contracts into structured records and flags the ones that do not add up.
- Searching internal knowledge. Answers questions with citations to company documents. For this, augmented search is usually enough, as I explain in what RAG is and when search is enough.
- Preparing recurring reports. Gathers data from several sources, detects anomalies and drafts a report for someone to validate.
When it does not make sense
- When a simple process can be solved with fixed rules. A well-designed form and an automation without AI win almost every time.
- When there are no reliable data or documents to start from. The agent inherits the company's mess.
- When a mistake would have serious consequences and there is no way to review it before it happens: payments, decisions about people, legal commitments.
- When the volume is so low that the cost of building and maintaining the agent is never recovered. If the process takes two hours a month, it is not worth it.
How to start without taking risks
- Pick a process with volume, a clear outcome and a person who knows it well and wants to improve it.
- Measure how it works today. Hours, errors, response time and cost per case. Without a baseline you cannot prove the improvement. I have a template in how to measure the baseline before automating.
- Build the smallest version that shows improvement, with a person approving every result. A workflow with fixed steps is usually enough at the start.
- Evaluate with real cases. Take thirty or fifty past cases, run them through the system and compare with what the person did at the time.
- Expand gradually, only when results have been stable for several weeks. More volume first, more autonomy later.
What to measure to know whether it works
| Indicator | What it tells you | How to get it |
|---|---|---|
| Approval rate without changes | How many results the person accepts as they are | Approve versus edit button in the review panel |
| Time per case | How long the process takes from start to finish | Timestamp at entry and at closure |
| Cost per case | Tokens, tool calls and review time | Consumption log per task |
| Escalation rate | How often the agent says "I cannot do this" | Counter of stops due to missing data |
| Errors caught late | Failures that got past review | Later complaints or corrections |
An agent that escalates 30% of cases and gets the remaining 70% right is better than one that attempts everything and gets 85% right. The first one knows when to keep quiet.
Risks you need to control
Beyond general compliance, the European Artificial Intelligence Act classifies systems by risk level and sets different obligations for each. The cases I describe here (quotes, email classification, extracting data from invoices) usually fall under limited or minimal risk, with transparency obligations. For sensitive uses such as recruitment, credit decisions or access to essential services, the regulation treats them as high risk and it is worth consulting a specialised legal adviser before starting.
What to ask whoever is selling it to you
- What happens when the agent does not know what to do? If the answer is "it always knows", be suspicious.
- How is quality evaluated, and with how many real cases?
- Which data leaves the company, to which provider and under what contract?
- How much does each processed case cost, not only the monthly licence?
- Who can change the instructions, and how are those changes logged?
If you want to study a specific case, the AI implementation service starts with a one-week discovery phase and a measurable pilot. And if your process is predictable, an automation without AI will probably suit you better: it costs less and fails less.