BlogDelegation
AI agent: what it is, and how a small business should choose one
An AI agent takes a goal and acts in your tools to reach it, where an assistant only writes. A definition, selection criteria and a trial for small businesses.
Contents
- What is an AI agent, in practice?
- What is the difference between an AI agent, an assistant and an automation?
- How do you recognise a real AI agent?
- Can an autonomous AI agent act without oversight?
- What is an AI agent for in a small business?
- How should a small business choose an AI agent?
- What is the best AI agent for a business?
- Should you build your own AI agent or adopt one?
- Where do you start with an AI agent?
An AI agent is software you hand a goal rather than a list of instructions, and which chooses the steps to reach it by acting directly in the company’s tools: it checks the accounts, updates a client record, prepares a follow-up, notices what did not work and tries another way. To choose one as a small or mid-sized business, you trust neither the label nor the demo, but five behaviours you can verify and a trial run on your own cases.
France Num, the French government programme that helps small businesses go digital, uses a very similar definition: an AI agent is “an autonomous programme able to interact with its environment, analyse, decide and act to reach a goal”. The same guide separates it from the assistant, a tool that “assists a human” in a supporting role, and from generative AI, which produces content from a prompt. The distinction is not academic, because most companies that say they use AI are in fact using the last two.
France Num’s 2026 barometer, carried out by the research centre Crédoc across 9,655 companies, puts a number on it. It finds that 40% of French small and mid-sized businesses say they use at least one AI solution, against 13% in 2024, and that the share reaches 53% among companies with ten employees or more. Yet 34% use it to generate text, voice or images, while only 11% use it to automate tasks: AI already writes a great deal in small companies, and it still acts very little.
What is an AI agent, in practice?
Take an ordinary request, sent on a Monday morning: “chase the September invoices that are still unpaid”. An assistant writes you a reminder template, which you adapt and then send to each client yourself. An automation sends the same email to every client whose invoice is more than thirty days old, without checking whether one of them paid yesterday or disputed a line. An agent opens the accounts, finds the six invoices concerned, checks that no payment has arrived in the meantime, rereads the latest exchanges with each client and writes six reminders that take what it found into account; it puts them to you, then, once you have approved, sends them and logs the reminder in each file.
Three things make the difference, and every serious definition has them. The agent starts from a goal rather than a script, it has tools that let it read from and write to your software, and it works in a loop: it acts, looks at the result, then decides the next step from what it observed. IBM sums it up as a system “that autonomously performs tasks by designing workflows with available tools”, and Google Cloud stresses that it pursues goals and completes tasks “on behalf of users”.
Agentic AI is the name for this whole family of systems, as opposed to generative AI, which stops at producing an answer. The word is recent and the thing a little less so, but the essential point fits in one idea: an agent is judged by what it changes in your tools, and not by the quality of what it writes.
What is the difference between an AI agent, an assistant and an automation?
What separates them is the answer to a single question, namely who decides the sequence of steps and when. The table below sums up what follows from it for a company that has to choose.
| Assistant | Automation | AI agent | |
|---|---|---|---|
| What you give it | a question | a script written in advance | a goal |
| Who decides the steps | you, with every message | you, once and for all | the agent, while it runs |
| What it produces | text you copy elsewhere | the same action every time | an effect in your tools, which depends on what it found |
| When it is wrong | you see it as you read | it stops, and that shows | it produces something plausible that has to be checked |
| Typical example | writing an advert, summarising a document | copying each signed quote into the tracking sheet | preparing overdue-invoice reminders file by file |
Anthropic, which develops the Claude models, draws the line in the same place in its design guide: a workflow runs the model and its tools along paths defined in advance in code, whereas an agent directs its own process and its use of tools. Google Cloud adds a point that matters to a buyer, since in its words an assistant “can recommend actions but decision-making is done by the user”.
None of the three categories is better than the other two. A task whose every branch you can write in advance is better handed to an automation, which is cheaper and more predictable, and we set out the rule for choosing between an agent and an automation task by task. The agent earns its place when the path depends on what turns up along the way, which is often true of the tasks that get put off until tomorrow.
How do you recognise a real AI agent?
By its behaviour, because the word no longer guarantees anything and plenty of products use it for a renamed assistant or automation. Five questions are enough, and each one has a yes or no answer.
- Does it write to a system? If the result of the conversation has to be copied by hand into some software before it exists, you are dealing with an assistant, however good it is.
- Does it choose the order of its steps? If the vendor can draw the whole path on a whiteboard, branches included, it is an automation, and it deserves to be paid for as one.
- Does it notice when it has failed? An agent reads the result of what it has just done, sees that a message bounced or that a record is empty, and changes approach instead of carrying on.
- Does it start without being asked? A deadline, an incoming email or an anomaly in a spreadsheet should be able to put it to work without anybody writing to it.
- Does it remember from one time to the next? An agent that forgets everything between two conversations makes you explain the company again every Monday, and delegation never takes hold.
Something that ticks none of these is an assistant, something that ticks two or three is an automation with a language model inside, and none of those categories is anything to be ashamed of. The trap lies elsewhere: only two of these behaviours show up in a demo, writing and choosing the steps, because a demo is a smooth path started by a person on a fresh session. The questions to ask the vendor about the other three are in our piece on agent washing and how to spot a real agent.
Can an autonomous AI agent act without oversight?
It can act alone on everything that stays inside the company, and has to go through a person for everything that leaves it. Reading a file, filing a document, updating a record or preparing a draft need no approval, whereas a message to a client, a publication or a payment each need one, however good the agent is. A decision that closes a door, such as turning down an application or dropping a supplier, stays human even when nothing leaves the building.
The rule has an advantage that finer-grained matrices lack: it is mechanical, so it does not get renegotiated case by case. It also has a consequence people often forget, which is that autonomy is set action by action rather than agent by agent. The same agent reads, drafts, writes internally and prepares something to send in the course of a single task, and classing it as a whole would mean having every read approved.
France Num recommends the same precautions for an assistant connected to a small business’s internal tools: keep permissions to the strict minimum, test in a separate environment, keep a trace of every action and maintain active human oversight, “especially for sensitive tasks such as sending messages or editing documents”. Those limits still have to be built into the product, as a permission never granted or an approval step in the code, rather than written into an instruction, since a rule written in natural language can be talked around by another sentence. We have listed what an agent should not be able to do, with a way to check each item in thirty seconds.
What is an AI agent for in a small business?
For all the work that crosses several tools and whose course depends on what you find, which covers a good share of what piles up on a Friday afternoon in a small company. The trades differ, the shape of the work much less so.
In an accounting firm, the agent spots the clients whose documents for the month are still missing, finds what has already been asked of them and prepares a reminder that names the exact documents instead of a generic nudge. In an estate agency, it writes up a viewing, updates the listing and offers the owner a time slot. In an IT services firm, it sees contract end dates coming and matches the consultants about to become available with open needs at clients. For a tradesperson, it prepares a quote from the site notes and the current price list, which the tradesperson reads before sending.
In each of these cases the agent is only useful if it sits where the work gets decided. An agent that lives in its own interface means you have to think of it, go there and copy the context across, which cancels a good part of the gain, while an agent you can call on in the team’s chat gets work handed to it by reflex. That is the choice we made with Balt, the AI employee of small and mid-sized businesses: you write to him in Slack, Teams or WhatsApp as you would to a colleague, and he works in the tools the company already uses.
Starting without a human also changes the nature of the service. A scheduled task, such as the list of unanswered quotes that lands every Monday at eight, needs nobody to start it, and it is often the one that pays most, because it brings into existence work that nobody had time to do. It must, on the other hand, say that it found nothing when it found nothing, otherwise its silence cannot be told apart from an outage.
How should a small business choose an AI agent?
By checking six points in the order below, then trying it on your own cases. The first ones quickly rule out the products that will not do, the last ones separate those that remain.
Does it live where your team already works? Check first that the agent is available in the chat tool you use, before you even look at its features. Teams and Slack do not offer an agent the same possibilities, and availability promised for later on yours is worth no more than absence.
Does it write to the tools you open every morning? The advertised number of connectors matters less than two precise questions: are the four or five applications that run the company among them, and can the agent write to them or only read? A read-only connection can be built in a day, whereas a write connection commits the vendor to handling errors, duplicates and permissions. You can run the exercise on the list of tools Balt connects to.
Is approval built into the product? During the trial, make the forbidden request yourself, for instance “send it directly, don’t ask me”. If the agent complies, the limit does not exist, whatever the documentation says.
Does it act with the rights of the person asking? An agent with an account of its own bypasses every permission in the company, since it sees everything its account can open, whoever is talking to it. The test fits in one question: do two people with different rights who ask it the same thing get the same answer? If they do, it sees too much, and we explain why an agent should not have an identity of its own.
Can you read and erase what it remembers? Memory is essential, otherwise everything has to be explained again each week, but a memory that grows without a filter makes the agent less accurate over time, and a memory nobody can read cannot be corrected. Ask what expires and after how long, because what an agent forgets matters as much as what it keeps.
Are you paying for access or for work? A vendor paid per seat earns the same whether its agent works a lot or very little, which aligns it with the opposite of what you are buying. Bring every offer back to the cost of the same month of work, so many follow-ups, so many meeting notes, so many quotes, and look at what really moves an agent’s bill.
That leaves the trial, which is the only moment a product is measured on your material rather than the vendor’s. Take thirty or so cases from your recent history whose right answer you already know, write that answer down before you run the agent, and replay each case three times: an agent that gives three different answers to the same question is not ready, even if all three are defensible. The full protocol, week by week, is in our method for testing an agent before you sign.
What is the best AI agent for a business?
There is no best agent in the absolute, and a ranking that claims otherwise compares models rather than products. The best agent for a small business is the one that passes the six questions above and succeeds in the trial on its own cases, which depends on its tools, its habits and the way its data is actually entered.
Two rules nonetheless help to separate the candidates. The first is about how many: an agent that holds context about your company should be the only one, because each copy of that context ages at its own pace and the divergence never announces itself, whereas an agent that brings one-off expertise can be called on as often as you like. We have described what twelve agents that do not talk to each other cost inside a single company, and how the divergence sets in without a sound.
The second is about specialisation. A trade’s knowledge ends up sooner or later in general-purpose models, so a specialised agent is protected not by what it knows but by what it answers for: write access, compliance for a precise use, responsibility when the answer was wrong. The question to put to every vendor is therefore simple, and we ask it of ourselves: what would your product lose if the model became twice as good tomorrow?
Should you build your own AI agent or adopt one?
Building an agent has become easy, and that is exactly what makes the question misleading. Several platforms let you configure one without writing code, by giving it a role, instructions, examples and a few documents, and France Num publishes a step-by-step tutorial for small businesses that concludes that “creating a simple AI assistant is within everyone’s reach”.
The work starts afterwards, and it never really stops. You have to connect the agent to your software with write access, decide what it may do alone, give it a memory that stays accurate and keep its instructions up to date, the same guide recommending a review of instructions and knowledge bases at least every quarter. You also have to survive the replacement of the model it runs on, which comes sooner than people expect and breaks first the fixes written to compensate for the previous model’s flaws.
The decision rule therefore fits in two sentences. A narrow, repeated task that lives in a single tool is perfectly well built in-house, when it is not simply automated without AI at all. Work that crosses several tools and requires knowing the company is better served by an agent you adopt, because what you are paying for then is the upkeep of everything else, writing being within reach of any model. Anthropic gives developers the same advice: find the simplest solution possible, and only add complexity when it becomes necessary.
Where do you start with an AI agent?
With the work whose result you can check in ten seconds, rather than the work that takes the longest. A follow-up, a set of meeting notes or an updated record can be judged at a glance, which lets you see several times in a row that it was done well, and it is on those observations that trust builds before it gets spent on tasks that are harder to check.
Allow three to four weeks before delegating becomes a reflex. The agent can do the work from day one, but the team needs that long to think of it, and the best sign of adoption is the day somebody hands it a task without anyone having suggested it. The next question is therefore which task to hand over first, and the order that works, week by week, answers it.
Frequently asked questions
What is the difference between an AI agent and ChatGPT?
In the way most companies use it, ChatGPT is an assistant: you ask a question, it writes, and a person copies the result into the right software and decides what happens next. An AI agent starts from a goal and acts in your tools itself, with the access the company has granted it, picking the work up where it left off. The line is therefore not the model, which can be the same in both cases, but what the product is allowed to do.
How much does an AI agent cost a small business?
The price depends mostly on the volume of work handed over and the tools the agent is connected to, far more than on the number of people using it. Offers come as a monthly subscription or as usage-based billing, often in credits, and the only honest comparison is to ask each vendor what the same month of work costs with them: so many follow-ups, so many meeting notes, so many quotes prepared. A per-seat offer deserves one more question, since a vendor paid for access earns the same whether its agent does a lot or very little.
Do you need to code to build an AI agent?
No. Several platforms let you configure an agent without writing code, by giving it a role, instructions, examples and documents, and France Num publishes a step-by-step tutorial aimed at small businesses. The work that remains comes afterwards: connecting the agent to your software with write access, deciding what it may do alone, giving it a memory that stays accurate and revising its instructions as the model is updated.
Can an AI agent act without approval?
Yes for whatever stays inside the company, such as reading a file, filing a document, updating a record or preparing a draft. No for whatever leaves it, a message to a client, a publication or a payment, and no for a decision that closes a door, such as turning down an application. That limit has to be built into the product, as a permission that was never granted or an approval step in the code, because a rule written into an instruction can be talked around by another sentence.
Sources
- France Num, AI assistant: a beginner’s guide to automating tasks in a small business (updated October 2025, in French)francenum.gouv.fr
- France Num, 2026 barometer of digital and AI use in French small businesses (Crédoc for the DGE, 9,655 companies, September 2026, in French)francenum.gouv.fr
- IBM, What are AI agents?ibm.com
- Google Cloud, What are AI agents? Definition, examples, and typescloud.google.com
- Anthropic, Building effective agents (December 2024)anthropic.com

