Skip to content

GuidesAI at work

AI automation: which tasks a small business should automate

AI automation for small businesses: a grid to score your tasks, how to measure the gain before and after, and what no-code, an agency or an AI employee costs.

21 min read

AI automation pays off first on the tasks that come back every week, take several hours a month, cost little when they go wrong and can be checked at a glance: sorting the inbox, chasing late invoices, matching incoming transfers to invoices, asking for the documents missing from a file, or the Monday update. The method has three steps: score each of your tasks on those four criteria, measure for two weeks how long it takes before you touch it, then choose between running it yourself in a no-code tool, handing it to an agency or giving it to an AI employee, knowing what each of those routes still leaves you to do.

Small businesses are only at the start. According to the 2026 France Num barometer, carried out by the research centre Crédoc across 9,655 companies for the French government, 40% of small and mid-sized businesses say they use at least one artificial intelligence solution, against 13% in 2024, and the share reaches 53% among companies with ten employees or more; yet only 11% use it to automate tasks. Insee, the French statistics office, finds for its part that among companies with ten or more employees that did not use AI in 2025, 71% said they did not see the point. What is missing is therefore not the tool, since there are dozens of them, but the answer to the question that comes before it: which task, in your business, is worth automating, and how will you know that it worked?

This article starts from your list of tasks rather than from a technology. Choosing between an agent and a classic automation for a given task is covered in our rule for choosing between an AI agent and an automation, and the inventory of an assistant’s tasks in what an administrative assistant costs. The question here is the one facing a manager who wants to automate with AI and has to decide where to start, what to expect and who should do the work.

What does AI automation add to classic automation?

A classic automation runs a scenario written in advance, such as “when an invoice is issued, add it to the tracking sheet”, and stops as soon as a case falls outside what was planned. AI automation adds two abilities: reading what is not structured, an email, a PDF or a scanned attachment, and handling cases that vary without every branch having been written beforehand. A sheet filled in from a form belongs to the first kind, while an order that arrives by email, in a different format for every client, belongs to the second.

The companies that have started are looking for two things above all. Among those with ten or more employees that used AI in 2025, Insee reports that 73% do so to cut costs, through time saved or the automation of repetitive tasks, and 64% to improve the quality and reliability of their processes. Those two motives match the two halves of the grid below: what the task costs today in time, and what it risks costing tomorrow in mistakes.

That second half is the one people underestimate. A classic scenario fails loudly, with an error message the same day, whereas a model that gets it wrong produces a plausible and false result that nothing flags. That is why the grid gives checkability as much weight as time saved: a result that can only be checked by redoing the work saves nothing, however well it is written.

Which tasks should you automate with AI?

A task is worth automating when it scores at least 9 out of 12 on the grid below, with no 1 on either of the last two rows: frequency, time spent, the cost of a mistake and checkability are each rated from 1 to 3.

Criterion 1 point 2 points 3 points
Frequency less than once a week every week every day
Time spent a month under an hour one to five hours more than five hours
Cost of a mistake it commits the business and cannot be undone: a payment, a contract, a decision about a person it would reach a client or a supplier, but a review stops it it stays internal and is fixed in a minute
Checkability you have to redo the work to know whether it is right a review of a few minutes a glance is enough

The first two rows measure what the task costs today, the last two what it will cost once automated, and the most common mistake is to stop at the first two. A long and frequent task whose result cannot be checked, such as ranking prospects or sorting job applications, moves the work into review instead of removing it. We have explained elsewhere why you delegate what can be checked first rather than what takes the longest, and the grid only puts that principle into numbers.

The cost of a mistake is scored with approval in mind. A payment reminder leaves the business, so a mistake would reach a client, but a review before it is sent is enough to stop it, which earns it a 2. A bank transfer, signing a contract or turning someone down score 1 whatever the quality of the tool, because the mistake has already had its effect by the time it is found: those tasks stay with a person, even when software prepares the material for them.

Applied to the tasks found in most small businesses, the grid gives the following ranking. The scores are those of a services business of about ten people, and yours will change with your volumes, which is the whole point of the exercise.

Task Frequency Time Mistake Checkability Total Verdict
Sort the inbox and prepare replies 3 3 3 3 12 launch
Enter orders received by email into the system 3 2 2 3 10 launch
Match incoming transfers to invoices 2 2 3 2 9 launch
Chase late invoices 2 2 2 3 9 launch, with approval
Ask for the documents missing from a file 2 2 2 3 9 launch, with approval
Prepare the weekly update 2 2 3 2 9 launch
Prepare a quote from past quotes 2 2 2 2 8 waits its turn
Pay suppliers 2 1 1 2 6 ruled out by mistake cost
Answer a sensitive complaint 1 1 1 2 5 stays with a person

Two lessons come back almost every time. Internal tasks, the ones whose result stays inside the business until a person releases it, come first because a mistake can be fixed there before it reaches anyone; and tasks that commit money or a delicate relationship leave the list through the mistake-cost column, even though they take time. The grid does not say those tasks cannot be tooled, only that they should not come first.

How do you measure the gain before and after automating?

The gain from an automation is measured in net hours: the time the task took before, minus the time spent afterwards reviewing, approving and correcting what the tool produces, minus the time spent monitoring and maintaining it.

Before launching anything, measure the task for two weeks. Count how often it comes back, time a dozen occurrences from start to finish, looking things up included, and note the misses: the reminder that went out ten days late, the document requested twice, the order entered with the wrong quantity. That record is the baseline without which no result can be claimed, and it often has an unexpected effect, since a task you thought was huge sometimes turns out to take twenty minutes a week and drops off the list before you have spent a cent.

After launch, measure the same things for four weeks, adding what the automation creates: review time, corrections, the cases the tool sends back to you because it cannot handle them. That side of the calculation is the one the announcements leave out, and it decides the result, since a drafting gain remains a workload handed to someone else as long as review is not counted.

Finally, write down before you start what the freed hours will become: ten more clients called back a month, quotes sent within forty-eight hours, an extra half-day on site. Time saved with no destination dissolves into the other tasks and shows up in no figure, which is the first reason why so many automations return nothing even though they work technically.

What does automating three tasks give back to a ten-person business?

In the example below, three automated tasks give a ten-person services business back 17.8 hours a month, worth about €324 of an assistant’s time. These figures are assumptions chosen to show the calculation, not a measurement, and the two-week record exists precisely to replace them with yours.

The assumptions are as follows. The business issues 80 invoices a month, about twenty of them paid late and needing two reminders on average, so 40 reminders of eight minutes each including the check of the bank account. Its shared inbox receives 25 messages a working day that need a reply, at about two minutes each, and it matches 60 transfers a month to its invoices, at four minutes each. After automation, we count one minute of review per reminder, forty seconds to review and send each draft reply, one minute per match to check, and half an hour a month per task to monitor the runs.

Task Before Review after Monitoring Net hours given back
Invoice reminders (40 a month) 5.3 h 0.7 h 0.5 h 4.1 h
Shared inbox replies (525 a month) 17.5 h 5.8 h 0.5 h 11.2 h
Matching transfers (60 a month) 4 h 1 h 0.5 h 2.5 h
Total 26.8 h 7.5 h 1.5 h 17.8 h

At the cost of an administrative assistant paid €2,100 gross a month, about €18.20 per hour worked for a French business with fewer than eleven employees, those 17.8 hours are worth about €324 a month. That amount is the ceiling on what automating these three tasks can cost each month and still pay for itself in time, and it only holds if the freed hours have a written destination. If they are the manager’s hours, the rate changes but not the logic: an hour of a manager’s time is only worth its price if it goes back to a client, a site or a sale.

The calculation also gives an order of magnitude that quotes do not. At this volume, an agency project at the median price of €12,000 would take about three years to pay for itself, and even a first automation at €2,000, which would only cover one or two of these tasks, would need more than six months before maintenance, whereas a subscription of a few tens or a few hundred euros a month can pay for itself in the first month if setting it up takes only a few hours. A single task rarely pays for a tool, and it is grouping several tasks of the same family that makes the numbers work.

Chasing invoices, finally, is the task that gives back the least time and perhaps the most money. Its value lies in invoices paid sooner and reminders no longer forgotten, two effects that a calculation in hours does not count, and it is also the task where the writing matters most, since each file calls for a different reminder letter depending on what it contains.

Do it yourself, pay an agency or take on an AI employee?

Three routes lead to the same automation, and they differ less in price than in what they leave you with: the no-code tool leaves you the design and the upkeep, the agency leaves you the ownership of the process once it has delivered, the AI employee leaves you the approval of whatever goes out.

The no-code tool, if you can write the scenario

On 1 November 2026, Make, Zapier and n8n listed entry prices of $12, $19.99 and €20 a month for their first paid plans, n8n stating annual billing, and each offers a way in without a subscription: 1,000 credits a month on Make’s free plan, 100 tasks a month on Zapier’s, and an n8n community edition, published on GitHub, that you host yourself. All three now let you call an AI model inside a scenario. The subscription is nonetheless the smallest part of the cost, because the real price is the design time, every branch to anticipate, the testing, then the upkeep each time a piece of software changes its interface or an unforeseen case turns up.

This route suits a stable flow whose every branch you can write, in a business where someone enjoys building these scenarios and has the time to keep them up to date. It suits the manager who discovers the subject on a Sunday evening much less well, because a scenario built alone on a weekend has no owner but the person who built it.

The agency, if the volume justifies a project

Across 400 budgets for artificial intelligence projects collected by La Fabrique du Net, the median was €12,000 and the average €23,518 as calculated on 1 November 2026, all projects combined, and the same guide puts annual maintenance at 15 to 20% of the initial cost. The bottom of the market is more affordable: an agency that publishes its prices listed on the same day a first automation from €2,000 for one or two standard tasks, delivered in three to five days, then €8,000 to €25,000 for three to five flows.

The scoping can be subsidised. Bpifrance’s Diag Data IA, from the French public investment bank, pays for eight days of an approved expert over three months at most, for a quote capped at €10,000 excluding VAT and €6,000 left to pay after a 40% subsidy, but it is reserved for companies of 10 to 2,000 full-time equivalent employees with at least €1 million in revenue, and it pays for the diagnostic and the roadmap rather than the build.

What the agency leaves you is what people most often forget to price: the description of the process, which it cannot invent, access to your tools, a person to approve its deliverables, then the ownership of what has been built once the engagement is over. An automation delivered without an internal owner ages at the pace of the software it connects, and every change becomes a new quote. Small businesses seem to have made up their minds: across all French small and mid-sized businesses, France Num finds that 17% have spent on AI subscriptions or licences, against 3% on custom development and 2% on consulting and services, and Insee that 80% of companies using AI acquire it as ready-to-use software.

The AI employee, if you want to write neither a scenario nor a specification

An AI employee works in your software from a request written in one sentence, with no scenario to draw, and is billed by subscription or by usage: we have put the recurring cost of an agent in a company at €200 to €3,000 a month, depending on the volume of work handed over. What it leaves you is shorter than with the other two routes without disappearing: opening access to the tools it needs to read, saying what is expected, naming who approves, and reviewing each message before it goes, which is the review time counted in the example above.

Its limits are those of what it has been given. It only knows what is written in your tools or what someone has explained to it, a muddled process stays muddled once delegated, and you should allow three to four weeks before the team thinks of handing it the work without being reminded.

Route What you pay What you supply What stays with you afterwards It fits when
No-code tool $12 to €20 a month at entry level, plus your building time the complete scenario, the build, the testing upkeep, monitoring, unforeseen cases the flow is stable and every branch can be written
Agency €12,000 median for an AI project, from €2,000 for one or two tasks, plus maintenance the process description, access, an approver ownership of the system, every change to be re-quoted the volume justifies a project and several systems are involved
AI employee a usage-based subscription, €200 to €3,000 a month for an agent in a company access, a written request, a person who approves approval of what goes out, the decisions tasks vary and nobody has time to draw scenarios

Nothing forces you to pick a single route. A no-code scenario keeps all its value for moving a piece of data from one system to another without judgement, and an agency for heavy work on a line-of-business system; an AI employee then takes what requires reading, writing or handling a case that varies, on top of that plumbing.

In what order should you automate your tasks?

One automation at a time, starting with the highest-scoring task that sends nothing out of the business, then moving to outgoing tasks under approval, and keeping for last the ones that require judgement.

  1. The record. For two weeks, note every task that comes back, with its frequency, its duration and its misses, without looking for a solution yet.
  2. The grid. Score each task from 1 to 3 on the four criteria and keep those that reach 9 points with no 1 on mistake cost or checkability.
  3. An internal task. Start with sorting the inbox or matching transfers, where nothing goes out and a mistake can be fixed, which quickly teaches you what the tool understands of the way you work.
  4. An outgoing task, approved. Then move on to reminders or missing documents, reviewing each message before it goes, until you know what the tool gets right and what it misses.
  5. The review. After four to six weeks, compare with the baseline, then keep, fix or stop before launching the next task.

One task at a time feels slow, yet it is the pace that lets you know what worked: three automations launched in the same month give you a quarter in which something improved without anyone being able to say what, or which one cost more than it gave back. If you are hesitating between two tools, have them work on thirty cases drawn from your own history, whose right answer you wrote down before the first attempt, as we describe in our method for testing an agent before you sign.

Which traps cancel the gain from an automation?

Four traps come back in almost every project that disappoints: automating a process that is already broken, launching an automation without an owner, letting what goes out leave without approval, and not monitoring the runs.

Automating a broken process

An automation reproduces the process it is given, at its own speed. If invoices go out without a purchase order number, if payment terms appear nowhere or if half the client records name a contact who has left the company, the automated reminder will simply send messages that lead nowhere, faster. The remedy is to fix what the task touches at the moment you automate it, rather than to overhaul everything before starting, which postpones the project indefinitely, and it is the rule we apply to the data an agent is about to read.

An automation without an owner

A scheduled run has no requester, so nobody watches it run and nobody notices when it stops. Name for each one a person who answers for it, who knows what it does and who takes it over if its author leaves, because an automation tied to the access of whoever created it falls with them the day that access is closed. That handover is prepared during the notice period, while the author is still there to explain what they built.

What goes out without approval

A reminder, a request for a document or an email to a supplier commits the business’s relationship with someone outside, and a mistake there takes effect before it is found. The rule we apply is mechanical: everything that goes out passes through a person who releases it, and the gain is counted in drafting time and turnaround, never in removing the check. That rule still has to be built into the tool, as a permission never granted or a mandatory approval step, rather than written into an instruction, which is the purpose of the list of what an agent should not be able to do.

No monitoring

An automation that fails without saying so is indistinguishable from an automation that found nothing. Insist that every run leaves a trace, including when it had nothing to do, that a repeated failure stops the work instead of retrying in silence, and that someone reads those traces every week, otherwise the breakdown is discovered the day a client complains: that is the problem of the 3 a.m. run nobody watches.

What Balt takes on

Balt, the AI employee of small and mid-sized businesses, is the third route in this article. He lives in Teams, Slack, WhatsApp and his web app, where the team writes to him as it would to a colleague, and he works in the software the business gives him access to, among more than 3,200 integrations. Recurring work becomes a routine requested in one sentence, with no workflow to draw: he runs it every Monday at 9 a.m., every hour or when something happens, a message in a channel or a change in a tool, and answers you wherever you asked for it.

After a run, he keeps the steps he learned as scripts, so the following runs are faster and cost less. He chases late invoices while leaving out the disputed accounts, as shown on the cash collection page, and asks for the documents missing from each file while showing you every request before it is sent, which is part of what he does for operations. He remembers what he learns about the way you work, and that knowledge serves the whole team rather than only the person who taught it to him.

Whatever leaves the business waits for your approval, and a routine that fails five times in a row stops instead of retrying endlessly. He keeps your data in Europe and is billed to the organisation rather than per user, from €99 a month for 40,000 shared credits, with 20,000 credits offered at sign-up and no card required, or in a managed plan at €3,000 a month, with an initial three-month commitment, in which our team sets him up in your tools and installs the routines, as the pricing page sets out. He does not fix a broken process, pays nothing and sends nothing without your approval, and the two-week record and the grid are still yours to do.

Where do you start this week?

With a sheet of five columns, the task, its frequency, its duration, what a mistake would cost and how long it takes to check the result, which you fill in over two ordinary weeks without looking for a tool. The task that comes out on top is your first automation, and the question that follows, who approves it and who answers for it once it runs on its own, is settled before its first run rather than after its first mistake.

Frequently asked questions

  • How do you automate a task?

    Measure it first for two weeks: how often it comes back, how long it takes, what a mistake would cost and how long it takes to check the result. If it scores at least 9 out of 12 when you rate those four criteria from 1 to 3, choose the route by what you can supply: a complete scenario for a no-code tool, a described process and a budget for an agency, a written request and a person who approves for an AI employee. Launch one task at a time and compare the result with your baseline after four weeks.

  • What are the automation tools?

    There are three families. No-code tools such as Make, Zapier or n8n chain actions between your software according to a scenario you write, and now let you call an AI model inside it; the automation features built into your software, such as the scheduled reminders of an invoicing tool, handle one task inside one tool; AI agents and AI employees work across several pieces of software from a written request, with no scenario to draw. The right choice depends mostly on whether you can write every branch of the task in advance.

  • How much does an AI automation agency cost a small business?

    Across 400 budgets for artificial intelligence projects collected by La Fabrique du Net, a French agency directory, the median was €12,000 and the average €23,518 as calculated on 1 November 2026, and some agencies list a first automation from €2,000 for one or two standard tasks. Maintenance comes on top, which the same guide puts at 15 to 20% of the initial cost a year, along with the tool subscriptions. Bpifrance’s Diag Data IA covers 40% of a diagnostic capped at €10,000 excluding VAT for companies with at least ten employees and €1 million in revenue, but it does not pay for the build.

  • Can you automate with AI for free?

    To try it out, yes: on 1 November 2026, Make’s free plan offered 1,000 credits a month and Zapier’s 100 tasks a month, while n8n publishes a community edition you host yourself. These plans are enough to test a simple scenario, rarely to run a business task every day, and the time spent building and maintaining it remains the real cost anyway. For a task that comes back every week, compare the price of the paid plan with the net hours it gives you back.

Sources

  1. France Num, 2026 barometer of digital and AI use in French small businesses (Crédoc for the DGE, 9,655 companies, published 17 September 2026, updated 12 October 2026, in French)francenum.gouv.fr
  2. Insee Première no. 2120, Information and communication technologies in companies in 2025 (21 July 2026, in French)insee.fr
  3. Make, pricing page, read on 1 November 2026make.com
  4. Zapier, pricing page, read on 1 November 2026zapier.com
  5. n8n, pricing page, read on 1 November 2026n8n.io
  6. La Fabrique du Net, AI agency rates in 2026 (updated 16 October 2026, 400 budgets, calculated 1 November 2026, in French)lafabriquedunet.fr
  7. Jaikin, AI automation for small businesses, price grid read on 1 November 2026 (in French)jaikin.eu
  8. Bpifrance, Diag Data IA, consulted 1 November 2026 (in French)bpifrance.fr

Balt does this work

  • Cash collectionInvoicing, collections, cash flow
  • EmailHis own address, and your inbox sorted.
  • RoutinesThe work that comes back, handled every week.

Read next

Give Balt its first job.

Thirty seconds to add it to Teams and get to work.

Try for free