AI agent for small business: do you need one? 7 questions
In short. An AI agent is a system where the model chooses the steps and the tools itself to reach a goal. It is useful when the path cannot be written in advance, and unnecessary (sometimes risky) when an automation or an assistant is enough. This guide offers seven questions to decide and four permission levels to start cautiously.
An AI agent for a small business is a system that receives a goal and then decides on its own which steps and tools to use, instead of following a list written in advance. The term is fashionable and is used for very different things. This guide helps the owner of a small company decide whether they need one, and on what conditions to try one. It draws on Anthropic’s article “Building effective agents”, the text of the GDPR and the CNIL page on deploying generative AI, all read on 29 October 2026.
What is an AI agent, compared with an automation?
Anthropic’s article separates two families. Workflows run a model and tools along paths written in code. Agents let the model direct its own process and choose its tools. The article adds that it is best to start with the simplest solution: a single call to the model, with your documents and a few examples, is often enough. Agent systems trade speed and cost for better performance on open-ended tasks, and autonomy raises the risk of errors that compound.
| Assistant | Automation (workflow) | Agent | |
|---|---|---|---|
| Who decides the steps | you, at each question | the code, written in advance | the model, along the way |
| Example | answer a question from your documents | extract PDF invoices into a spreadsheet each month | handle a customer request by looking things up in several tools |
| Predictability | high | high | lower |
| Cost of an error | low, you proofread | bounded, the steps are known | assess before granting rights |
| When to choose it | occasional questions and drafting | repetitive task with known steps | open-ended task whose path depends on what turns up |
Many small-business needs fall into the first two columns. Our article on PDF invoices to Excel is one example: a script, not an agent.
Do you need one? Seven questions to decide
Ask these for the specific task you have in mind, not for “AI” in general.
| Question | If the answer is… | Then |
|---|---|---|
| 1. Are the steps always the same? | yes | an automation is enough |
| 2. Would one request to the model, with your documents, solve it? | yes | an assistant is enough |
| 3. Does the path depend on what turns up along the way? | yes | an agent is worth considering |
| 4. What is the worst possible mistake, and can it be undone? | irreversible (sending, paying, deleting) | human approval is mandatory |
| 5. What data must the agent read? | personal or confidential data | settle purpose, access and retention first |
| 6. Who checks before the final action? | nobody | do not start |
| 7. How will you know it works? | no success criterion | write the criterion and keep a log before starting |
If you only answer “yes” to questions 1 or 2, you do not need an agent. If you answer “yes” to question 3 but have no clear answer to 4, 6 and 7, the project is not ready.
Which permission levels should you give an agent?
The method we suggest: move up a level only after watching the previous one on fictional or non-sensitive cases. Anthropic recommends testing in a sandboxed environment with guardrails, and planning checkpoints for human feedback and a stopping condition (for example a maximum number of iterations).
| Level | What the agent can do | What you control |
|---|---|---|
| 0. Read-only | read named documents, answer or summarise | the list of documents it can reach |
| 1. Draft | prepare a text, a table, a reply | you proofread and send it yourself |
| 2. Action with approval | propose an action and carry it out after explicit approval | every action goes through a person |
| 3. Bounded autonomous action | act alone within a closed list, with a cap and a log | a readable log, a stop button, written limits |
Level 3 is not a goal. For many tasks, staying at level 1 or 2 is the right choice: the agent saves time, and the decision stays with a person.
What do the GDPR and the CNIL say?
Two points are worth knowing before you hand a task to an agent.
-
Automated decisions. Article 22 of the GDPR sets out the right not to be subject to a decision based solely on automated processing, including profiling, when it produces legal effects or similarly significantly affects the person. Exceptions exist (contract necessity, law, explicit consent) and, in some cases, the controller must provide at least the right to obtain human intervention, to express a point of view and to contest the decision. An agent that alone sorted job applications or turned away customers would typically fall in this zone: get legal advice first.
-
A framework for use. In its page on deploying generative AI, the CNIL recommends starting from a concrete need, defining allowed and forbidden uses, training users and involving the data protection officer and business managers early. It also reminds readers that these systems can produce inaccurate but plausible results.
These pages do not cover every case. They give reference points, to be completed with a lawyer for your situation.
What does a reasonable first project look like?
Here is a fictional example, with no real client. A company receives quote requests by email, with attachments in varied formats. Instead of launching an autonomous agent, it starts like this:
- Level 0. The agent only reads the folder of incoming requests and writes a one-line summary for each.
- Level 1. It proposes a draft reply and a list of missing information. A person proofreads and sends.
- Measure. Over a fictional or real month, count the proofreading time and the corrections needed. If corrections stay frequent, stop or rethink the task.
- Level 2 only afterwards, for one precise action (for example filing the request in a spreadsheet), with explicit approval.
This scenario illustrates a method, not a measured result. The real cost depends on the model used, the volume and the time spent checking.
What are the limits of an AI agent?
-
It can be wrong with confidence. Mistakes in reading, calculation or judgement remain possible, and they can compound from one step to the next.
-
It costs more than a script. Anthropic notes that agent systems trade cost and latency for better performance: check this against your own volume.
-
It needs oversight. Log, access limits, emergency stop, proofreading: the checking time is part of the cost. If a second model does that checking, test it with planted errors first.
-
It does not replace responsibility. You remain responsible for what is sent or decided in your name.
-
It is not always the right answer. For a task with known steps, an automation will be simpler, cheaper and easier to verify.
How does Castel Studio work on agents?
We build and repair agent systems on Claude Code: agents, hooks, skills, memory between sessions and MCP tools that connect them to your systems. A first piece of work starts from 150 €: a one-page map of what each agent does, rules with tests (one action that must be blocked, one that must pass) or blocking of confidential folders. Connecting to your email, spreadsheets or messaging is priced by written quote. For tasks with known steps we also offer regular reports from your spreadsheets or an assistant that knows your business from 790 €. For eight practical starting points with prices, see AI for small business.
To understand the technical building blocks, read our articles on agents and skills in Claude Code and on hooks that block actions. If your agents handle confidential documents, see also our guide to private AI for business, and for a model that runs on your own hardware, the guide to a local LLM for small business.
Describe the task and the tools involved in a few lines, without attaching sensitive documents through the form. We answer in writing within 48 hours, including to tell you that a simple automation or an assistant is enough.
The starting price covers tidying or checking a system, not a whole project. Model fees, maintenance and validation time have to be planned separately.
Questions
What is the difference between an AI agent and an automation?
In the distinction Anthropic proposes, an automation (a "workflow") follows a path written in advance, while an agent decides the steps and the tools to use by itself. If you can write all the steps in advance, an automation is usually simpler to control.
Does an AI agent replace an employee?
That is not the right question to start with. An agent can take on a precise task, with limited rights and human validation. It does not replace the responsibility of the person using it. Possible errors, cost and control should be assessed on the task, not on a whole job.
What data should I give an AI agent?
The minimum the task needs. Start read-only, on non-sensitive or fictional documents. Once personal data is involved, define the purpose, the retention period and the access rights with your data protection officer or a lawyer.
What does the GDPR say about decisions made by an AI?
Article 22 gives a person the right not to be subject to a decision based solely on automated processing when it has legal effects or similarly significantly affects them. Exceptions exist, with safeguards including the right to obtain human intervention. A real case needs legal advice.
How much does it cost to set up an AI agent?
At Castel Studio a first piece of work on an agent system (a one-page map, tested rules, confidential files blocked) starts from 150 €. Connecting to your tools (email, spreadsheets, messaging) is priced by written quote. Model fees and maintenance are extra.