Process automation: the complete 2026 guide for an SMB
Process automation means having software carry out, according to rules you set, tasks that were until then done by hand: when an event occurs, one or more actions follow without anyone stepping in. For an SMB, it is decided one process at a time, by first quantifying the time lost (frequency × duration × people), then choosing the level of tool: an improved spreadsheet, a no-code platform or custom software.
In most SMBs, a process lives in a spreadsheet, three mailboxes and the memory of the person who runs it. It holds, until the day that person goes on leave. Automating is not buying a robot: it is deciding which event triggers which action, then handing execution to software.
This guide to process automation is written for the head of an SMB with no technical team. It does not compare tools. It gives a method for choosing what to automate and quantifying the time lost. Then it says at which point an improved spreadsheet, a no-code platform or custom software is the right answer.
🧭 What is process automation?
Process automation means having digital tools carry out automatically tasks that were until then done by hand. That is the definition given by France Num, the French government service dedicated to the digitalisation of small and medium-sized businesses (SMBs). It speaks of processes “that are triggered automatically, and carry out actions according to conditions you define” (our translation). The software executes; the rule remains the company’s decision.
A process, in the sense of the ISO 9001 standard as cited by AFNOR, is a set of interrelated activities that turn inputs into an expected result. Chasing an unpaid invoice, onboarding an employee, handling a job application: each of these sequences is a process, whether it is written down somewhere or lives only in the head of the person who does it. This guide sets out the method.
Trigger, rule, action: the mechanism in one sentence
Every automation fits in one sentence: when this happens, if that condition is met, then do this, and notify that person. France Num boils it down to “if this happens, then do that”. The tools on the market all use this vocabulary. At Zapier, a scenario “consists of a trigger, which starts it, and one or more actions”. At Microsoft Power Automate, “a trigger is an event that starts a flow”, and a condition decides whether the tasks run.
The fourth block is the one that gets forgotten. An automation that fails silently is worse than a manual task. The person who used to do it no longer does, and nobody knows the software no longer does it either.
BPA, RPA, BPM, workflow, no-code, custom: who does what
Vendor pages pile up the acronyms. They designate different levels of tool, and that level is what matters for an SMB. The definitions below are those of the vendors who use them, since no standard sets them; only the BPMN notation, used to draw a process, is an ISO standard (ISO/IEC 19510).
The useful distinction is between RPA and no-code. RPA mimics a human “at the user interface level”, writes SAP. No-code platforms, for their part, “can access APIs directly”, writes IBM. The second route is more robust, because it does not depend on the screen.
What has changed: AI in automation
For two years, vendors have been adding artificial intelligence building blocks to their platforms. Two things need to be told apart. A workflow that calls a language model at one step, to classify an email for instance, is still a workflow: its path is defined in advance. An agent, by contrast, “dynamically directs its own processes and tool usage”, according to the distinction drawn by Anthropic in December 2024. OpenAI takes it up: an application that integrates a model without handing it control of the flow is not an agent.
For an SMB, the consequence is simple. The vast majority of management processes have a predictable path: they belong to the workflow, with or without an AI step. The agent is reserved for open-ended problems, where the number of steps cannot be foreseen, and it “often trades latency and cost for better task performance”, again according to Anthropic. The CNIL, France’s data protection authority, notes in a July 2026 paper that agentic AI “renews and amplifies the risks to personal data”: decision-making autonomy, persistent memory, responsibilities that are harder to pin down. As of 17 September 2026, no primary source measures the adoption of AI agents by SMBs.
🧩 Who is concerned, and when should you start?
Process automation concerns any company where a repetitive task is done by hand several times a week, by several people, with errors. In France, the smaller the company, the less equipped it is with management software: the gap between companies with 10 to 49 persons employed and those with 50 to 249 is clear, as the Eurostat chart shows. The starting point, in an SMB, is therefore rarely software: it is a spreadsheet.
The SMB whose processes live in Excel and emails
A spreadsheet is an excellent tool to start with. France Num says so plainly: a small business comfortable with its spreadsheet can build a working dashboard in it. The problem comes with time. The same public service lists the limits that appear “very quickly”: manual entries, tables piled up to consolidate, lost files, multiplied versions, scattered data. On inventory, France Num observes that a majority of small business owners still manage it in Excel, or even in a notebook. Standalone tools “quickly become insufficient as soon as activity picks up”.
The limit is not the software’s capacity. An Excel worksheet accepts 1,048,576 rows and 16,384 columns according to Microsoft. The limit is human. An academic review presented at the EuSpRIG conference in 2015 found errors in 94% of the 85 operational spreadsheets inspected, for a per-cell error rate of around 3.9% in the laboratory. Rare per cell, almost certain at the scale of a workbook, and detected in only about 60% of cases. The studies date from 1995 to 2001; the mechanism has not changed.
The signs that it is time to leave the spreadsheet, we see them come back from project to project. The same data entered in two places, a file named “v3-final-bis”, a single person who knows how it works, a Friday spent consolidating. The article “Why Excel always ends up becoming a problem” details these signs; the guide “Migrating from Excel to business software” describes the way out.
Eurostat and INSEE count companies with 10 or more persons employed: 53.9% had an ERP in 2025. The 2025 France Num Barometer surveys small and medium-sized businesses with 0 to 249 employees, overwhelmingly very small ones: 23% have an ERP, but 88% have at least one management solution and 69% invoicing software. None of these sources measures the share of companies that run a process in a spreadsheet, nor the share that automate their processes.
When an off-the-shelf tool or no-code is still enough
A no-code platform is enough as long as the process consists of moving a piece of information from one piece of software to another, according to a simple rule, between tools that already exist. An order received on the website creates a row in the tracking table and notifies the workshop. An invoice marked as paid in the invoicing tool closes the file in the CRM. Processes of this kind can be automated on your own. According to the France Num guide, written by two no-code practitioners, simple automations can be built after half a day of training on Zapier, one day on Make, two days on n8n.
No-code reaches its limit when the rule becomes your own. A quote whose price depends on twenty business parameters, an approval that follows a workflow specific to your company, a calculation that nobody else does the way you do. The platform can move the result; it does not know how to produce it. That is the threshold described in our comparison of no-code and custom development.
When not to automate
Some processes should not be automated, or not yet. France Num proposes three criteria for prioritising: the time spent, the estimated complexity of the automation, and the impact in the event of an error. The guide recommends favouring “processes with high time savings, low complexity and a medium impact in the event of an error”. Four situations fall outside the frame:
- The process is not defined. If it changes depending on who does it, automating it freezes an arbitrary version. You clean up before you tool up; automating disorder makes it faster
- Exceptions dominate. A rule with more special cases than normal cases cannot be written; it is handled by hand, with a tool that helps
- The volume is too low. A monthly ten-minute task does not justify two days of set-up. The formula in the next section settles it
- Judgement is required. RPA targets processes that are “rule-based, not subjective”, Microsoft points out; embedded AI is “designed to support execution and decision-making, not to replace accountability”, writes SAP. A decision remains a decision
France Num adds a useful reminder: poor-quality data, incomplete, duplicated or out of date, distorts decision-making and exposes the company to compliance risks. Automating a process fed with wrong data produces errors faster.
🧰 What to gather before you start
Automating a process starts with describing it as it really is, with volumes, exceptions and an owner. Without that description, no-code produces an automation nobody understands six months later. Custom software, for its part, produces software that matches an imaginary process. France Num notes it: “it is essential to maintain clear documentation” so that automations “can easily be taken over by other members of the team”.
What you gather in-house
- A map of the real process. Who does what, in what order, with which tool, starting from what happens and not from what should happen. One page is enough. Common mistake: describing the dream process; you find out at the first case that does not fit.
- Volumes and frequencies. How many times a day or a week, how many minutes each time, how many people. France Num warns that people often prioritise the process whose friction is most visible, while the time lost is elsewhere once you quantify it.
- The list of exceptions. The cases where the rule does not apply, and what you do then. This is the list that decides between no-code and custom: short, no-code is enough; long, you need software that carries the rules.
- An inventory of the tools in place and their exports. Invoicing, CRM, bank, email, spreadsheets: which have an API or an export, which are dead ends. An automation goes no further than the most closed tool in the chain.
- An owner and a recipient for alerts. The person who decides the rules, and the one who is notified when the automation fails. Not necessarily the same person.
What is built with the vendor
When the process calls for custom software, the first week is spent writing the rules. The method published by OTTOPILOTE starts this way: in week 1, “we take stock of your calculations, checks and workflows with the people who apply them, edge cases included. That document is the reference for what follows”. It is validated before the first line of code. The quote is free and capped: the price is set before starting. Your five items above are its raw material. The software specification is its complete form for a larger project.
🛠️ Automating a process step by step
Automating an SMB’s processes takes six steps: quantify the time lost, map, choose the level of tool, build a first process, measure, then extend. How long each step takes depends on the level chosen. Count a few days for a no-code automation between two existing tools, six to eight weeks for a first batch of custom software. The two paths share the first three steps.
Quantifying the time lost yourself
The time spent on a process is calculated with the formula France Num gives: frequency × unit duration × number of people involved. The gain is then calculated as time saved × hourly cost, minus the cost of the tools and the set-up. No need for a vendor statistic. Your own volumes are enough, and they are the only ones that count.
Take France Num’s example: 20 job applications a week, 15 minutes each to read, file, reply and enter in a tracking table. That is 5 hours a week, or 235 hours a year over working days excluding leave. After automation, less than a minute per application: more than 4 hours freed up every week.
The hourly cost of €44.7 is INSEE’s 2025 average for companies with 10 or more employees, across all market sectors; INSEE does not publish a figure by company size. Replace it with your own: it is your cost, not an average, that makes the calculation right. And note what the calculation leaves out: the errors avoided, the reminders that go out on time, the person who can take their leave.
The steps, with realistic durations
- Quantify, then prioritise. A table of candidate processes with time lost, complexity and impact in the event of an error, following France Num’s three criteria. Half a day with the people who do the work. The first process selected is the one that combines high gain, low complexity and medium impact.
- Map the real process. Trigger, rules, actions, exceptions, alert recipient, on one page. One day. It is also the moment you discover that part of the process needs no tool, only a clear rule.
- Choose the level of tool. Few exceptions and open existing tools: no-code platform. Company-specific rules, calculations, approval workflows: custom software. The two combine: the software carries the rules, the platform connects it to the rest.
- Build the first process. In no-code, the France Num guide counts 1 to 2 days of initial training, then 0.5 to 2 days for a first automation that is “well designed, simple and robust”, and half a day per subsequent automation once the tool is mastered. In custom, OTTOPILOTE ships a first batch to production in six to eight weeks, the pace of a management tool. The timeline below details the eight weeks.
- Measure. Four weeks of real use, with the indicators set at step 1: time spent, errors, delays. In no-code, this is also the period when you discover the cases the rule had not foreseen.
- Extend, or stop. The next process starts again at step 1 with what you have learned. The subsequent batches of custom software “start from real use”, not from the initial specification.
The gap between the two paths is not only the timeline. On your own, you build a no-code automation in a few days, and you handle its monitoring and updates. With a vendor, the first custom batch takes six to eight weeks, the quote is set before starting, and it is the vendor who carries the build and the testing.
A realistic timeline for a first custom batch
Integrations with your tools and data migration are what make this timeline vary the most. A first batch that connects to three pieces of software and takes over five years of history does not fit in eight weeks: it gets split up.
The processes most often automated in an SMB
No statistic measures how often each process is automated. The six below are the ones we come across most, and for each the level of tool that most often fits. The 2025 France Num Barometer gives an order of magnitude of how small and medium-sized businesses are equipped: 69% have invoicing software, 25% a purchasing or inventory solution.
💶 How much process automation costs
Process automation costs internal time first, then a subscription that rises with volume, or a custom software project priced after scoping. The prices below were read on 17 September 2026 on the vendors’ pricing pages, in dollars or euros as displayed. They change often. The custom ranges are those published by OTTOPILOTE, not market statistics: no official source publishes any.
The cost by line item
The cost that rises with volume, and reversibility
The platforms’ pricing pages all show the same mechanism: the price follows the volume. Zapier counts one task per step and per connector call, with tiers from 100 to 2 million tasks. Make bills one credit per action, in blocks of 10,000. n8n moves from 2,500 to 10,000 then 40,000 runs, from €20 to €50 then €667 per month. An automation that works is an automation that gets used more, and therefore costs more. The profitability calculation has to be redone at next year’s volume.
Reversibility is the blind spot. Each of the four platforms exports its scenarios in its own format. n8n saves its workflows as JSON, Power Automate as .zip packages, Make as “blueprints”, and Zapier reserves export for its Team and Enterprise plans. No format is shared, and none exports the connections to the applications. Your business data, for its part, is not in the platform: it stays in the software the platform connects. Changing tools means rebuilding.
Custom software reverses the logic, on one condition: that the contract provides for the assignment of the code. Article L113-9 of the French Intellectual Property Code automatically transfers rights to the company only for software created by its employees. For a vendor, a written clause is indispensable. OTTOPILOTE provides for it in the contract, with handover of the repository and the documentation on delivery, once the project is paid. On the platform side, n8n offers a middle way. Its self-hostable edition is free for internal use, under a licence that is not open source in the strict sense, and the vendor, n8n GmbH, is based in Berlin.
A platform that connects your tools sees the data it moves, customer names included. Zapier hosts its data on AWS servers in the United States; Make offers regions in Europe and the United States without publishing the country of the data centre; n8n Cloud hosts in the European Union; Power Automate follows the region of the environment, France included. A tool hosted outside the European Union is a transfer within the meaning of Chapter V of the GDPR, to be recorded according to the CNIL. As of 17 September 2026, the adequacy decision with the United States remains in force for certified companies, but an appeal is pending before the Court of Justice: check the day's status before signing.
🧭 What automation makes possible, and what it does not
Process automation removes repetition errors, not design errors. France Num puts it this way: repetitive tasks are “sources of human error: omissions, typos, information entered in the wrong place”. Automation “eliminates these risks by always carrying out the same actions in the same way”. It also carries out, always in the same way, a wrong rule.
Guarantees and limits
Time given back, provided you measured it beforehand, otherwise nobody knows where it went.
Data entry errors that disappear, as long as the rule and the input data are right.
Deadlines met without thinking about them, if someone receives the alert when the automation fails.
A process that survives absences, because it is written down somewhere, and no longer merely known.
A trace, over 7 to 60 days depending on the tool and the plan: enough to understand an incident, not enough to keep a log in the CNIL's sense.
Decide for you. A rule executes; judgement, exceptions and responsibility stay with a person.
Fix a vague process. Automated, it becomes a faster vague process.
Keep up with a process that changes every month. Rules have to be updated, France Num reminds us; every change is work.
Make up for wrong data. Duplicated or out-of-date data produces wrong results, faster and with more confidence.
⚠️ The most frequent pitfalls
Six pitfalls come back in SMB automation projects, from the first no-code scenario to custom software. No statistic measures them: they are the ones we come across, with the countermeasure we apply.
Six pitfalls, and their countermeasures
The rule is written by management, never with the person who does the work. Edge cases are discovered after go-live, when the team works around the tool.
Fix Write the rules with the people who apply them, then replay real cases before going live.
Thirty no-code scenarios written by three people, one of whom has left. Nobody knows any more which feeds which, and the bill follows the volume: at Zapier, scenarios, AI steps and code all draw on the same pool of tasks.
Fix A register of automations, an owner per scenario, and the move to custom software when the rules pile up.
The rule covers the normal case. The customer with a special VAT status, the partial order, the incomplete file go through the standard flow and come out wrong.
Fix List the exceptions before choosing the tool; provide a "handle manually" exit in every scenario.
An expired password, a renamed field, and the scenario stops. Zapier switches off a scenario when 95% of its runs fail over 7 days; n8n does not alert without a dedicated error workflow; Power Automate does not send alerts by default on every flow.
Fix A named alert recipient per scenario, and a weekly check of the runs.
The scenario moves names, addresses and amounts to a tool hosted in the United States. That is a transfer within the meaning of the GDPR, and the tool is your processor: it processes the data only on your documented instructions, transfers included (Article 28).
Fix Inventory the tools and where they are hosted, read the data processing agreement, choose a European region when one exists.
You automate the most annoying process, not the most expensive one. France Num points it out: visible friction often hides less time lost than a quiet process does.
Fix Frequency × duration × people for each candidate, and start with the heaviest.
🎯 In summary
An SMB’s path rarely goes from the spreadsheet to custom software in one step. It passes through no-code, and each stage keeps something from the previous one.
Automating processes is not a question of tools, it is a question of rules. Once the rule is written, measured and owned by someone, the tool is chosen in a single meeting. To find out whether yours fits in no-code or calls for code, describe your process in a few sentences: a human reply within 24 working hours, free quote.
📚 Sources
- France Num, “Automation: an essential solution to save time and better manage your small business” (in French): definition, mechanism, prioritisation criteria, time-lost formula, durations and subscription ranges (updated 9 July 2026)
- France Num, “Small businesses: why computerise your company’s financial management?” (in French): limits of the spreadsheet (updated 25 March 2026)
- France Num Barometer 2025 (in French): equipment of small and medium-sized businesses in management solutions and AI
- Eurostat, enterprises using ERP or CRM software (isoc_eb_iip) and digital intensity (isoc_e_dii): 2025 data, enterprises with 10 or more persons employed
- INSEE, “Hourly labour cost by activity” (in French): 2025 hourly cost
- AFNOR, ISO 9001 process mapping: definition of a process and the management, operational, support typology
- Workflow Management Coalition, glossary: definition of workflow
- IBM, “What is business process automation?” (in French), SAP, “Process automation” (in French) and Microsoft, “What is RPA?”: vendor definitions of BPA, BPM, RPA and intelligent automation
- OECD, “The agentic AI landscape and its conceptual foundations”: definition of AI agents (February 2026)
- CNIL and CIANum, exploratory paper on agentic AI (in French): risks to personal data (July 2026)
- Anthropic, “Building effective agents”: distinction between workflow and agent (December 2024)
- OECD, Employment Outlook 2019 and ILO, global index of exposure to generative AI: work automation, jobs and tasks (2019 and May 2025)
- Panko, “What we don’t know about spreadsheet errors today”, EuSpRIG 2015: spreadsheet error rates
- Microsoft, “Excel specifications and limits”: worksheet capacity
- Pricing pages read on 17 September 2026: Zapier, Make, n8n, Microsoft Power Automate, Airtable, Timetonic
- Zapier, error alerts, n8n, Error Trigger, Microsoft, flow failure notifications and n8n, export and import of workflows: monitoring and reversibility
- Zapier, data privacy, n8n, security and Microsoft Learn, Power Automate regions: hosting locations
- GDPR, Chapter V (CNIL, in French) and Chapter IV, Article 28; adequacy decision (EU) 2023/1795: transfers outside the European Union, status checked on 17 September 2026
- French Intellectual Property Code, article L113-9: assignment of rights over software
- n8n, Sustainable Use License: terms of use of the self-hosted edition
Frequently asked questions
It means handing software the sequence of a process that used to be done by hand: when a job application arrives, it is filed, an acknowledgement goes out and the tracking table fills in, with nobody stepping in. Automating an isolated task, such as a scheduled send, is not yet automating a process: we speak of a process when several steps and several people follow one another according to a rule.
The most widespread typology in quality management, used by AFNOR to map ISO 9001 processes, distinguishes management processes (steering, deciding), operational or core processes (producing the service sold) and support processes (HR, accounting, IT). The ISO 9001 standard itself defines what a process is without imposing this typology. Automation first affects support and core processes.
No neutral typology serves as a reference: IBM distinguishes task, workflow, process and intelligent automation; Automation Anywhere retains process, integration and AI; SAP speaks of RPA, workflows and intelligent automation. The most useful breakdown for deciding remains the level of tool: spreadsheet and macros, a no-code platform that connects applications, RPA that replays actions on screen, custom software that carries the rules, and an AI agent that chooses its own steps.
Work automation refers to machines or software carrying out tasks previously entrusted to people. The studies that measure it separate the task from the job: the OECD estimated in 2019 that 14% of jobs faced a high risk of automation and that a third would see their tasks change substantially; the ILO concluded in 2025 that the transformation of jobs is far more likely than their replacement. In an SMB, automating a repetitive task moves time towards what requires judgement.
No, as long as the automation connects tools you already use: no-code platforms are set up through menus, by choosing a trigger and actions. What you learn is the logic of rules and exceptions, not a language. Knowing how to code, or having code written, becomes necessary when the software must calculate or validate according to rules that exist only in your company.
Yes, provided you start from a single, measured process. No-code platforms offer free plans limited in volume, for example 100 tasks per month at Zapier or 1,000 credits at Make in September 2026, then subscriptions of €10 to €50 per month according to France Num. The question is not the size of the company but the time lost: frequency × duration × number of people, compared with the cost of the tool and its set-up.
By recording the same three indicators before go-live and then after a few weeks of real use: time spent on the process, number of errors or reworks, and the delay between the event and the response. The "before" measurement is the one most often forgotten; without it, the gain cannot be proven. Also note the time spent monitoring and fixing the automation: it comes off the gain.
A project in mind? Let’s talk.
A brief, an honest read and a free, capped quote within 24 hours. We help you pick — or build — the right tool.
OTTOPILOTE