GuidesAI at work
AI agent examples: seven jobs an agent can do in a small business
Seven AI agent examples for a small business, from prospecting to the phone: what sets each one off, and what always stays with a person.
Contents
- Which AI agent examples are most useful in a small business?
- How does an AI agent handle prospecting and winning clients back?
- What does an AI agent do with quotes, from request to follow-up?
- How does an AI agent track invoices and chase late payment?
- How does an AI agent write up a meeting, then follow it up?
- What can you hand an AI agent in recruitment or an IT services firm?
- How does an AI agent keep administrative files moving?
- Can an AI agent answer the phone for the business?
- How much does an AI agent cost, from custom-built to ready-made?
- Which AI agent example should you start with?
- Balt takes on these examples from the tools you already use
An AI agent in a business takes on a task that runs across several pieces of software and whose course depends on what it finds along the way: it chases an overdue invoice after checking the bank account, prepares a quote from past quotes, writes up a meeting or asks for the document missing from a file. The seven examples below take one business function at a time and set out, for each, what triggers the agent, what it does, which tools it touches and what stays with a person. An order of magnitude comes with an example whenever a source supports one, and only then.
These examples matter because most small businesses that use AI use it to write, and very few use it to get things done. The 2026 France Num barometer, run by Crédoc across 9,655 French small and medium-sized businesses, found that 40% of them use at least one AI tool, up from 13% in 2024, and 53% among the medium-sized ones. Yet 34% use it to generate text, voice or images, while only 11% use it to automate tasks. Insee, which only surveys companies with ten or more employees, found in 2025 that 15% of those with 10 to 49 employees used at least one AI technology, a gap that shows how much the measure depends on the question asked.
The definition itself, and the five behaviours that set an agent apart from an assistant, are in our article on what an AI agent is and how to choose one. This guide starts from the work instead, one function at a time. One rule runs through all seven examples: the agent prepares on its own whatever stays inside the company, and whatever leaves it goes through a person’s approval, given before sending or, on the phone, on the scope of what the agent may handle alone.
Which AI agent examples are most useful in a small business?
The seven examples below follow the functions of a small service business, from prospecting through to the switchboard, and the table sums them up before each one is covered in detail. We put what stays with a person in the last column, because that is what decides whether an example holds up once it is in production.
| Function | What triggers it | What the agent does | What stays with a person |
|---|---|---|---|
| Prospecting and win-back | the Monday review, a client with no order for a year | spots clients drifting away, prepares a list of accounts and a message | choosing who to approach, and every send |
| Quotes | a request by email, a quote with no reply | lists what is missing to price, starts from comparable quotes, drafts the follow-up | the price, the send and the signature request |
| Invoices and late payment | a missed due date | checks the bank, sets aside disputes, drafts one reminder per account | the send and the formal notice |
| Meetings | a video meeting in the calendar | attends, transcribes, writes the minutes and follows up the actions | the version sent to a client |
| Recruitment and IT services | an open role, the end of the month, a client need | searches the talent pool, prepares timesheets and skills dossiers | rejecting a candidate, approving a timesheet |
| Administrative follow-up | a new file, the morning review | checks each file against the documents expected, drafts the requests | signatures and decisions that call for judgement |
| Phone | an incoming call the team cannot take | takes the call, handles or passes on the request, files the transcript | the scope of calls it handles alone |
How does an AI agent handle prospecting and winning clients back?
An AI agent for prospecting starts with the clients you already have: it compares orders year on year in the CRM, picks out those who have stopped ordering and looks through the latest exchanges for the likely reason behind each silence. It is usually triggered by a fixed slot, such as a list that lands every Monday morning, or by a threshold you set, for instance twelve months without an order. For each client it keeps, it drafts a message that reflects what it found, a buyer who left or a competitor’s quote that was mentioned, and it presents that reason as a hypothesis to check, never as a fact.
New business comes next. Starting from your best clients, the agent searches a database such as Apollo for companies that look like them and hands you a list to review, with one contact at each. It touches the CRM, HubSpot or Salesforce, the prospecting database, the mailbox and the campaign tool, Brevo or Lemlist, where it can set up a whole campaign while leaving out recipients who should not get it, such as a contact who has left or a client in dispute.
Who gets approached, when and with what message stays with a person, and every send goes through them. French law draws one more line: according to the CNIL, a prospecting email to a professional can go without prior consent if its subject relates to that person’s job and lets them opt out, whereas a private individual must have agreed beforehand. Nor does the agent borrow a salesperson’s personal LinkedIn account to send invitations on their behalf, for reasons we set out when we worked out what automating LinkedIn really costs.
We give no figure for the gain here, because no public measure holds up. A precaution is worth more than a number anyway: a mass reminder sent to every dormant client is the quickest way to lose some of them, and the point of an agent is precisely to write to each one for a reason of their own. Balt does this work in the CRM and campaign tool you have opened to it, as the business development page shows.
What does an AI agent do with quotes, from request to follow-up?
An AI agent follows a quote from the first request through to signature. When a request arrives by email, it spots what is missing to price it, a quantity or a deadline for instance, and prepares the questions to put to the client before anyone writes down a price. When it is time to price, it finds your comparable quotes in Sellsy, Odoo or your Word templates, carries over your usual prices and margins, and flags anything out of the ordinary instead of deciding for you.
Follow-up is where the agent brings the most regularity. The schedule is set on the day the quote goes out and tied to its validity: a first message two or three days later, then at seven and fourteen days, and a last one right before it expires, each bringing something the previous one did not, as our guide to writing a quote that gets signed explains. Before every follow-up the agent rereads the last exchange with the client, so it does not chase someone who asked for a variant yesterday or is still waiting to hear from their bank.
The price, a discount outside your rules, sending the quote and starting the e-signature stay with a person. The agent sets up the request in YouSign, DocuSign or PandaDoc with the right document and the right signatory, then waits for your approval, because a quote that has gone out binds the company for as long as it is valid. It will not get a client who is still thinking to sign any faster, and we found no reliable public measure of what it would change to the signature rate. What it does change shows without a figure: no quote is left hanging for want of a follow-up.
How does an AI agent track invoices and chase late payment?
An AI agent for collections starts from the due date: every morning it picks up overdue invoices, checks the bank account for any payment that has come in, rereads the client’s history and sets aside the disputed accounts before writing anything. For the rest it drafts a reminder whose tone follows how late the payment is, friendly the day after the due date and firmer around the fifteenth day, then puts it to you. When a client replies, to say they have paid or to dispute an amount, the agent checks the claim and prepares what comes next.
It touches the invoicing or accounting software, Pennylane or Sellsy for example, the bank and the mailbox. The same agent can match the transfers received on Qonto against the invoices issued every Friday and tell you what still needs explaining, which stops it chasing a client who has already paid under another reference.
The stakes show up in cash. According to the 2025 report of the French Observatory of Payment Terms, presented by the Banque de France on 24 September 2026, late payments in France stood at 13.4 days at the end of 2025, and without them small and micro businesses would have had €13 billion more in cash. Between French businesses, every invoice paid late also entitles the supplier to a fixed recovery fee of €40, due once per invoice and even when part of it has been paid, as Service Public Entreprendre points out.
Sending each reminder, settling a dispute and deciding to send a formal notice stay with people, as does any payment or credit note the agent can prepare but never issue alone. The sorting rule fits in a sentence: an unpaid invoice tied to a dispute is not chased, it is settled first. The full schedule and the letter templates are in our guide to chasing an unpaid invoice, and Balt does this work in Pennylane and Qonto, showing you every reminder before it goes out.
How does an AI agent write up a meeting, then follow it up?
An AI agent for meetings is triggered by the calendar: as soon as a video meeting appears there, it joins, transcribes it with each speaker named and then sends minutes that keep the decisions apart from the actions to take. Useful minutes are short, and each action carries a single owner and a date, otherwise it is likely to still be open at the next meeting. The transcript then serves as a memory, since three weeks later you can ask what was decided about a given client.
The useful work starts once the meeting is over. The agent nudges each owner when their action falls due, prepares the next meeting from the actions still open and updates the client’s file when the meeting was about them. It touches the calendar, Google Calendar or Outlook, the video tool, the team’s messaging and, depending on the case, the CRM or Notion.
The time at stake is far from marginal: the Wisembly and Ifop barometer found in 2018 that French managers attended 3.5 meetings a week, which comes to 27 days a year spent in meetings. Internal minutes need no approval, whereas the version sent to a client commits the company and is read before it goes. A recording is also announced to participants before the meeting starts, and how to write the document itself is covered in our guide to meeting minutes.
Balt joins Zoom, Google Meet, Teams, Webex or GoTo meetings as soon as Google Calendar or Outlook is connected, and sends you the minutes as a direct message in Teams, in Slack or by email. Thirty minutes before the meeting, it also prepares a brief for you, as the meetings page shows.
What can you hand an AI agent in recruitment or an IT services firm?
In recruitment, an AI agent takes the steps where volume is high and the decision is light. An IT services firm also hands it two pieces of work that drive its invoicing and its sales: each consultant’s monthly timesheet, which French firms call the CRA, and the skills dossier sent to the client.
In recruitment, it searches and tidies up
As soon as a role opens, the agent searches the applicant tracking system’s talent pool first, then LinkedIn or Indeed, and returns a ranked list that gives the reason for each choice, with no score. It turns interview notes into a clean write-up, updates the candidate’s record and drafts follow-ups for those who have been waiting too long for an answer. Rejecting a candidate and choosing which profiles go to the client stay with people. Article 22 of the GDPR gives everyone the right not to be subject to a decision based solely on automated processing, and since 2 August 2026 an agent that talks directly with a candidate has to tell them it is an AI, under Article 50 of the EU AI Act. Our guide to AI in recruitment goes through each step.
In an IT services firm, it prepares timesheets before the invoice
A few days before the end of the month, the agent sends each consultant a timesheet already filled in with what the firm knows, the working days, public holidays and booked leave, then checks each return against the purchase order and the schedule before preparing the invoice. The declaration remains the consultant’s and the approval the client’s. A late timesheet costs more than it seems: with 45-day end-of-month payment terms counted from the end of the month of issue, one of the two methods Service Public Entreprendre describes, an invoice dated 30 November 2026 is payable on 14 January 2027, and the same invoice dated 2 December only on 14 February. The full collection calendar is in our guide to collecting consultant timesheets.
For a proposal, it prepares the skills dossier
When a client need comes in, the agent matches the consultants coming off assignment against the open needs in BoondManager, rewrites the chosen CV in the client’s template, anonymises it where required and lists what the profile lacks for that need. It selects, orders and rephrases but never adds, because the skills dossier is a sales document the consultant has to be able to defend line by line in front of the client. The consultant reads the dossier, then a person releases it, and what Balt does inside Boond is set out task by task in our guide to BoondManager and AI.
How does an AI agent keep administrative files moving?
An AI agent for administration checks every open file against the list of documents it should contain and drafts, for each client or supplier, a request that names the exact document, where a generic reminder would go unanswered. It is triggered when a file opens or by the morning review, in which it flags incomplete files and absences that have no cover yet. You give it the list of documents once, or it works the list out from completed files, and it warns you when a case is out of the ordinary.
The same work extends to schedules and suppliers. When someone is out, the agent cross-checks the schedules, the absences recorded in Lucca and the history of each site to suggest colleagues who already know the job or the client, and it compares promised delivery dates with actual ones to spot the supplier who keeps slipping. It touches the mailbox, the spreadsheets where schedules often live, Excel or Google Sheets, the absence tool and the task tool. This use is still rare in small businesses: France Num counts 13% of French SMEs using AI to analyse or classify documents.
Requests sent to clients and suppliers go through a person, as do a change to the schedule and the message that tells a stand-in to go. Signatures and decisions that call for judgement stay with people, and we sorted the tasks of an assistant’s job by what each requires, whether access, judgement, a signature or someone physically present, in our guide to what an administrative assistant costs. Balt does this follow-up from the tools you have opened to it, and your team leads can write to it from the field on WhatsApp, in Teams or in Slack, as the operations page shows.
Can an AI agent answer the phone for the business?
Yes, and this is the example where approval changes shape, because a live call cannot be read before it happens. An AI phone agent picks up when the team is busy or outside opening hours, understands the request, handles it when it falls within what the agent has been allowed to do and otherwise passes it on with all the context. It can also make outgoing calls, returning a missed call or confirming an appointment, then file the transcript and a summary in the right client’s record.
It touches the phone system, Aircall or Ringover for example, the calendar to book an appointment, the CRM to recognise the caller and the team’s messaging to alert whoever takes over the call. What a person approves is decided in advance: which calls it handles alone, and what it never promises. Anything that commits the company beyond that scope goes back to someone on the team.
Two rules frame this example. Since 2 August 2026, Article 50 of the EU AI Act requires that callers be told they are talking to an AI from the start of the first interaction, unless it is obvious, an exception the European Commission asks to be read narrowly. Since 11 August 2026, cold-calling a private individual in France also requires their prior consent, valid for a year at most, except within an existing contract, and the Bloctel opt-out register closed that same day.
For an order of magnitude, a turnkey AI phone receptionist aimed at small businesses listed, on 14 November 2026, €79 a month before VAT for 150 minutes, then €0.39 a minute. The ways to connect it and the other offers are compared in our guide to AI phone answering. Balt makes calls, answers them and transcribes them, introduces itself as your company’s AI employee at the start of every call, and you decide which calls it handles alone, as the phone page explains.
How much does an AI agent cost, from custom-built to ready-made?
The price of an AI agent runs from tens of thousands of euros, when an agency builds it for you, to under a hundred euros a month for a ready-made AI employee. The first is paid up front and then every year in maintenance, whereas the second is paid monthly, in proportion to the work you hand it.
At the custom end, the grid La Fabrique du Net publishes from the prices French agencies charged in 2026 puts a customer service agent connected to the CRM at €25,000 to €50,000 for a small business, and an autonomous agent that acts inside an ERP or other software at €50,000 to €90,000. The same guide adds annual maintenance of 15 to 20% of the initial cost. The lower end of the market is more accessible: one agency that publishes its rates listed, on 14 November 2026, a first process from €2,000 before VAT for one or two standard tasks such as invoicing or reminders, then €8,000 to €25,000 before VAT for three to five workflows across several departments.
At the ready-made end, you pay for usage, month by month. Balt starts at €99 a month for 40,000 credits, billed per organisation and shared by the whole team, with unlimited users and 20,000 credits free on sign-up, no card required. The Managed plan, where our team sets it up in your tools, costs €3,000 a month with an initial three-month commitment. Why per-seat pricing does not suit an agent, and what to ask before you sign, is covered in our article on what an AI agent costs.
Which AI agent example should you start with?
Start with the example whose trigger already lives in one of your tools and whose result can be checked at a glance, even if another one takes up more of your time. A due date in the accounts or a meeting in the calendar sets the agent to work without anyone having to think about it, and a reminder or a set of minutes can be judged in seconds, which lets you see several times in a row that the work is done well.
Two more criteria separate the candidates. A mistake must be able to stop at a review, which is true of anything that waits for your approval before it leaves, and the time you get back needs a written destination before you start, otherwise it dissolves into the week and its value can never be shown. Applied to the examples in this guide for a small service business, these criteria give the following order, which your own volumes may change.
| Example | Trigger already in a tool | Result checkable at a glance | When to start it |
|---|---|---|---|
| Chasing late payment | the invoice due date | yes | first |
| Meeting minutes | the meeting in the calendar | yes | first |
| Documents missing from a file | a file being opened | yes | first |
| Timesheets in IT services | the end of the month | yes | first, for an IT services firm |
| Following up quotes | the date the quote went out | yes | next, before it prices anything |
| Searching the talent pool | a role opening | partly | next, without ranking candidates |
| Prospecting | the Monday review | no, a list is judged over time | later |
| Phone | an incoming call | only afterwards | later, once the scope is written down |
Prospecting and the phone come last for different reasons. A prospect list can only be judged after several weeks of replies, so a targeting mistake costs a month before it shows, whereas a call takes place before anyone can read it. Allow three to four weeks after that before the team hands the agent work without being reminded to, and move to the second example once the first needs no more than a quick review. The order that works, week by week sets out that progression, and our grid for scoring your tasks lets you redo this ranking on your own work.
Balt takes on these examples from the tools you already use
Balt, the AI employee for small businesses, takes on these roles as a single colleague, where seven separate agents would each keep their own copy of your company. You write to it in Slack, Teams or WhatsApp, it works in the tools the company has opened to it, and it remembers what it learns, from the client who always pays on the 10th to the tone of your best messages. Work that comes back every week becomes a routine, asked for in one sentence and run at the right moment, and a routine that fails five times in a row stops instead of retrying quietly.
Anything that leaves the company waits for your approval, whether a reminder or a skills dossier, and that rule is built into the product with no setting to switch it off. It is what lets you hand Balt one example after another without watching every step, starting with the one your own files point to.
Frequently asked questions
What is the simplest example of an AI agent in a business?
Chasing overdue invoices. Every morning the agent picks up the invoices past their due date, checks the bank account to make sure no payment has come in, sets aside the disputed ones and drafts a reminder for each client that fits how late they are, which you read before it goes out. The result takes seconds to check and a mistake stops at the review, which makes it a good first example.
Can an AI agent write to a client without approval?
It should not. Anything that leaves the company, whether a reminder, a quote or a campaign, goes past a person before it is sent, and that limit has to be built into the product rather than written into an instruction, because an instruction can be talked around by another sentence. Whatever stays inside, such as reading a file or preparing a draft, needs no approval.
What is the difference between an AI agent and an automation in these examples?
An automation sends the same reminder fifteen days after the due date to every client who is late. An agent first reads each client’s file, notices that one paid yesterday or is disputing a line, and adjusts what it prepares. When every branch of a task can be written down in advance, an automation stays cheaper and more predictable.
How long before an AI agent is useful in a small business?
The agent can do the work as soon as its access is open and its rules are set, which takes a few days. Allow three to four weeks after that before the team hands it work without being reminded to, which is why it pays to start with an example whose result is quick to check.
How much does an AI agent cost a small business?
Built to order by an agency, an agent that acts inside an ERP or other software costs a small business €50,000 to €90,000 according to La Fabrique du Net’s grid, and one agency that publishes its prices lists a first automation of one or two tasks from €2,000 before VAT. A ready-made AI employee is paid by subscription or by usage: Balt starts at €99 a month for 40,000 credits shared by the whole team, with 20,000 credits free on sign-up.
Sources
- France Num, 2026 France Num barometer: digital tools and artificial intelligence in French SMEs (Crédoc for the DGE, 9,655 businesses surveyed from 23 March to 18 April 2026, published September 2026)francenum.gouv.fr
- Insee Première no. 2120, Information and communication technologies in businesses in 2025 (21 July 2026)insee.fr
- CNIL, Commercial prospecting by email, SMS-MMS and automated callingcnil.fr
- Banque de France, The 2025 annual report of the Observatory of Payment Terms (24 September 2026)banque-france.fr
- Service Public Entreprendre, Payment terms between businesses and late payment penalties (checked on 7 August 2026)entreprendre.service-public.gouv.fr
- Wisembly and Ifop, barometer on meetings and employee engagement (press release, 24 September 2018)ifop.com
- CNIL, GDPR, chapter III, article 22: automated individual decision-makingcnil.fr
- European Commission, Transparency obligations under Article 50 of the AI Act (FAQ)digital-strategy.ec.europa.eu
- Service Public, Telephone canvassing banned: what are the new rules? (published 21 July 2026, updated 11 August 2026)service-public.gouv.fr
- Elio, AI phone receptionist pricing, before VAT (checked on 14 November 2026)eliocall.com
- La Fabrique du Net, AI agency rates 2026, detailed price grid of AI agencies (checked on 14 November 2026)lafabriquedunet.fr
- Jaikin, AI automation for SMEs, price grid (checked on 14 November 2026)jaikin.eu

