Where to start with AI automation when you run an SME in 2026?#

By Laurent Perello, founder of Perello Consulting -- web pioneer for over 25 years, AI operator in production since 2024. Last updated: 13 April 2026.

The question comes up in every first meeting: where do I start? Lists of use cases abound and answer nothing. Choosing an automation is not choosing a tool. It is isolating, within an existing business, the process whose cost is measurable, whose data is clean, whose gain is provable and whose risk is contained. This article offers a seven-step method, each step grounded in a public French or European source. It is applied to a fifteen-person SME in business services. The preliminary costing (how much an SME loses each month on automatable tasks) is covered in the previous article; we start from its results to build the decision tree here.

Why most executives get the order wrong#

The dominant reflex in 2026 is to choose a tool before mapping the hours. You open ChatGPT, you test Copilot, you hear a peer talk about a specialised agent, then you look for where to plug it in. The order is reversed. France Num, in its AI Self-Diagnostic published by the DGE and Bpifrance, requires starting with the assessment of integration capacity before any technology choice¹. This is not an editorial recommendation. It is the official French framework.

Why is this reversed order so costly? Because AI tools are interchangeable, whereas the structure of hours is specific to each business. Two SMEs of the same size in the same sector never have the same distribution of monthly hours between admin, sales, marketing, support. Choosing a tool first means buying a solution without having identified the problem. France Strategie reminds us in Les metiers en 2030: technology-driven transformation primarily affects high-volume support functions, and that volume is measured business by business².

The second bias is visibility bias. A website chatbot demonstrates well in a meeting. An internal meeting-notes automation is invisible. The first attracts the budget, the second recovers the hours. The OECD, in the Employment Outlook 2024, documents this displacement: real AI gains concentrate on invisible, repetitive, high-monthly-volume tasks³. The market, however, primarily sells what demonstrates well. You therefore inherit an offering calibrated for visibility, not impact.

[UNIQUE INSIGHT] The third bias is the absence of prior measurement. Without costing the hours spent on automatable tasks in your own organisation, no arbitrage is possible. That is the purpose of article 1: to produce that measurement before any methodological choice. If you have not yet laid that foundation, start there. The method that follows relies on that costing and cannot do without it.

The seven-step decision tree#

We now enter the method. Each step poses a question, provides scoring criteria, references a public source, then receives a concrete application on a fifteen-person SME in business services. The goal is to arrive, in half a day, at a single process candidate for the pilot, defensible in committee.

Step 1 -- Map the three most time-consuming functions#

Question: in which three functions do your employees collectively spend more than sixty percent of their monthly hours? You are not trying to be exhaustive. You are looking for the three blocks that weigh. The scoring criteria, zero to five per function, combine cumulative declared monthly hours, number of people involved, and weight in loaded payroll. A function scores zero to fifteen.

The reference framework is France Num's AI Self-Diagnostic¹. The hours-to-cost conversion relies on the INSEE ESANE database⁴ and the ICHTrev-TS index⁵. The functional weighting follows France Strategie's Les metiers en 2030². These four sources suffice to build a costed map.

For the fifteen-person SME in business services, the ranking obtained is: administration and back-office, one hundred and twenty cumulative monthly hours across three people, score eleven out of fifteen; sales (prospection, CRM, meeting notes), ninety-five hours across four people, score twelve out of fifteen; marketing and content, seventy hours across two people, score eight out of fifteen. The three retained functions are admin, sales, marketing.

Step 2 -- Identify high-frequency repetitive tasks#

Within each retained function, list the three most frequently repeated tasks with predictable content. Three criteria, yes or no: weekly or higher frequency, reproducible pattern (identifiable input, processing, output), variability below twenty percent of cases. A task that validates all three enters the funnel.

The empirical basis is documented by Goldman Sachs (task automation exposure)⁶ and by the OECD (Employment Outlook 2024)³. DARES working-conditions surveys provide the function-level breakdown for France⁷.

Applied to our SME, this step produces nine candidate tasks. Admin side: supplier invoice entry, payment reminders, bank reconciliations. Sales side: meeting notes, inbound lead qualification, CRM updates. Marketing side: social reformatting, weekly monitoring, performance reporting. No elimination at this stage. The funnel stays wide.

Step 3 -- Score each task on volume, repetitiveness, sensitivity#

Which combines volume, repetitiveness and low data sensitivity? Score each candidate task on three axes, zero to five. Axis A, cumulative monthly volume in hours: zero for under five hours, five for over forty. Axis B, repetitiveness and automatability: zero for under twenty percent of treatable cases, five for over sixty percent, reference Goldman Sachs ("between 25% and 50% of workload could be replaced"⁶) and Anthropic Economic Index (thirty-six percent of occupations with Claude usage on more than a quarter of their tasks⁸). Axis C, data sensitivity, inverted: zero for sensitive personal data (health, HR), five for public or non-personal internal data, per CNIL framework⁹ and AI Act, article 6¹⁰.

Maximum score is fifteen. For our SME, the top three after scoring are clear: client meeting notes (A=4, B=5, C=4, total thirteen out of fifteen); supplier invoice entry (A=5, B=4, C=3, total twelve); inbound lead qualification (A=3, B=4, C=4, total eleven). The priority process is identified, but not yet validated. Three steps remain before the pilot.

Step 4 -- Validate upstream data quality#

Is the necessary data accessible, clean, structured, compliant? Five criteria, yes or no: accessibility via API or export (not only on paper or in someone's head); cleanliness above eighty percent (no duplicates, stable formats, filled fields); structured or semi-structured (spreadsheets, JSON, timestamped transcripts); clear GDPR legal basis (legitimate interest, contract, consent); AI Act assessment completed (minimal, limited, high, unacceptable risk).

Sources are CNIL's Recommendations on AI system development⁹, the CNIL AI Sandbox¹¹, the AI Act (articles 10 and 50)¹⁰ and ANSSI's Security recommendations¹². This is not an optional framework. It is the condition of lawful processing.

For client meeting notes, the verdict is favourable with reservation. Audio recordings are already produced on Google Meet; transcription is possible; cleanliness exceeds ninety percent; the GDPR basis is achievable via participant consent and a ninety-day retention policy; the AI Act classification places the system at limited risk, with an information obligation under article 50. Validation: yes, with consent and retention policy to formalise before the pilot.

Step 5 -- Estimate risk-weighted monthly ROI#

What is the net monthly gain, once costs and risk are removed? The public formula extends the framework from article 1 and adapts it to a single process:

net_ROI = (hours x loaded_hourly_cost x automatable_rate)
        - (licences + amortised_deployment + human_supervision)
        - (risk_provision)

Three decision criteria. Gross monthly gain above three times the monthly cost of licences and supervision: green light. Risk provision between ten and twenty-five percent of gross gain depending on sensitivity documented at step 4. Deployment payback under six months.

Calculation sources are public: OECD Employment Outlook 2024 for the exposure rate³; INSEE ESANE and ICHTrev-TS for the loaded hourly cost⁴⁵; URSSAF for the employer loading coefficient of approximately 1.42¹³; Bpifrance Diag Data IA as an order of magnitude for deployment¹⁴.

[ORIGINAL DATA] Applied to client meeting notes on our SME. Hours per month: twenty-eight (meetings plus manual drafting). Loaded hourly cost, sales: eighty-two euros (INSEE services, coefficient 1.42). Automatable rate: fifty-five percent (transcription, structuring, draft writing). Gross monthly gain: 28 x 82 x 0.55, approximately one thousand two hundred and sixty-three euros. Monthly costs: transcription and LLM licence approximately ninety euros; human supervision two hours per month at eighty-two euros, i.e. one hundred and sixty-four euros; deployment amortisation three thousand euros over twelve months, i.e. two hundred and fifty euros; total costs five hundred and four euros. Risk provision at fifteen percent of gross: one hundred and ninety euros. Net ROI approximately five hundred and sixty-nine euros per month, payback under three months.

Step 6 -- Test on a 30-to-60-day pilot#

On what restricted scope is the gain provable before generalisation? Four criteria. A single sub-process, a single team, a single data type. Minimum duration thirty days for raw measurement, sixty days recommended for measurement corrected for learning effects. Baseline metrics set before the pilot (see step 7). Named human supervisor, output sampling rate at or above twenty percent.

The institutional framework exists: ANSSI describes the isolated experimentation phase in its Security recommendations¹²; the CNIL, through the AI Sandbox, provides an example of a supervised experimental framework¹¹; MIT Sloan documents the typical enterprise learning curve¹⁵.

For our SME, the pilot lasts forty-five days. It covers two salespeople (one senior, one junior), covers one hundred percent of client meetings, applies a thirty percent supervision sampling rate on generated notes, and relies on a baseline measured over the preceding thirty days. Everything is written before launch. Nothing is adjusted mid-course without traceability.

Step 7 -- Measure before/after to validate the actual gain#

Does the observed gain confirm the step 5 estimate, and to what percentage? Decision criteria are explicit. Gap between estimated and observed gain below thirty percent: generalise. Gap between thirty and sixty percent: revise scope before generalising. Gap above sixty percent: stop or redesign. Quality perceived by internal users, minimum score four out of five across ten users. Compliance or security incidents: zero.

The before/after measurement method is set out by France Strategie in Artificial intelligence and work¹⁶ and by the OECD in Employment Outlook 2024³. BCG, in AI at Work, gives the empirical order of magnitude (approximately one hour per day per user)¹⁷, an observed figure, not a declared one.

Applied to our SME, after forty-five days: twenty-three monthly hours recovered versus twenty-eight estimated, an eighteen percent gap. Measured net gain five hundred and twelve euros per month versus five hundred and sixty-nine estimated, a ten percent gap. Quality score four point three out of five. Zero GDPR incidents. Decision: generalise to all four salespeople, then open the next step on supplier invoice entry (second in scoring).

The case of a 15-person SME, B2B services#

We consolidate here the seven steps applied to the same reference SME. Fifteen full-time equivalents, business services sector, split into one executive, one sales director, two salespeople, two marketing, three delivery, two admin-finance, one HR, one support, two production operations. The method produces, in half a day of scoping, one pilot candidate process and an estimated net ROI.

The synthesis master table#

Reading the result#

The priority process identified is automating client meeting note generation. The estimated net gain for a two-person pilot is approximately five hundred and seventy euros per month. After generalisation to all four salespeople, extrapolation yields a range of one thousand one hundred to one thousand four hundred euros per month, subject to supervision maintained for the first two months and zero GDPR incidents. The measurement framework is published; it is reproducible by any external auditor.

Why this result is defensible in committee#

Each step produces a written deliverable: the map, the scored task list, the data validation grid, the ROI calculation, the pilot protocol, the before/after report. You can present six documents in a management committee, before an auditor, or to the CNIL in case of inspection. Sources are public, the formula is explicit, the prudential adjustment is documented. No line rests on a private benchmark that you could not produce. That is the difference between publishable costing and a sales slide.

The four main mistakes to avoid#

The seven steps hold as long as you skip no rung. Four mistakes, in our field experience, collapse the method before the pilot even starts. Each is concrete, each has a public counter-example.

Starting with visibility rather than impact#

[PERSONAL EXPERIENCE] Choosing a task that demonstrates well in a meeting (a website chatbot, marketing image generation) instead of the task that actually weighs in the payroll. That flatters a committee's ego, not the business. The France Num AI Self-Diagnostic framework specifically requires starting with capacity -- the actual state of processes -- before choosing a use case¹. France Strategie's Les metiers en 2030 reminds us that transformation primarily affects high-volume support functions, not showcases². Concrete example: a services SME spends eight thousand euros on a website chatbot that recovers two hours of support per week. The same amount, invested in meeting note automation, recovers five to eight times more hours with clean internal data.

Skipping the data quality step#

Automating a process whose input data is dirty, fragmented or non-compliant is amplifying error at scale. Noise becomes signal, gaps become invisible, and an external auditor can no longer reconstruct the decision chain. The CNIL writes it explicitly in its Recommendations on AI system development: the quality of training and inference data conditions the lawfulness of processing⁹. ANSSI reminds us in its Security recommendations that the security of an AI system rests on the integrity of its datasets¹². The AI Act, article 10, requires data governance practices for high-risk systems¹⁰. Concrete example: an SME that automates commercial scoring on a CRM with thirty-five percent duplicates and twenty percent invalid emails does not generate leads. It produces AI-signed noise, challengeable by a prospect in case of dispute.

Automating without human supervision at the start of deployment#

An AI system, even one that performs well in pilot, drifts. Input data evolves, edge cases accumulate, users adapt their behaviour. At the start of deployment, a human sampling rate of twenty to thirty percent of outputs is the minimum norm documented by ANSSI¹² and implicitly recommended by the CNIL⁹. The AI Act, article 14, requires effective human control over high-risk systems¹⁰. Cutting supervision to save supervisor hours is the primary source of post-pilot incidents. Concrete example: a firm that generalises meeting note generation without sampling review for three months discovers, via a client complaint, that six percent of notes contain a name or amount confusion. The reputational incident exceeds twelve months of automated gain.

Omitting the before/after measurement#

Without a baseline measured before the pilot, the gain cannot be proven. It is narrated, not demonstrated. That is the difference between a return on investment and a sales argument. France Strategie, in Artificial intelligence and work¹⁶, and the OECD, in Employment Outlook 2024³, insist on empirical before/after measurement as a condition of any credible assessment of AI gains. BCG, in AI at Work, quantifies the average gain at approximately one hour per day per user, but specifies that this figure is observed, not declared¹⁷. A pilot without a baseline allows neither generalisation nor a stop decision. Concrete example: an SME deploys a commercial AI tool for a year, declares a "significant improvement", cannot produce either recovered hours or quantified additional leads. At the budget review, the CFO cuts the budget with no possible counter-argument.

What this method changes#

The method does not guarantee that a first AI project succeeds. It guarantees that a first AI project is decided on defensible grounds, with a bounded budget, a framed risk, a measurable gain. These are four properties that are missing, today, from most deployments in French SMEs.

First change: prioritisation becomes explicit. You no longer choose between three tools you barely know; you choose between nine candidate tasks scored on public criteria. The arbitrage rests on a grid, not a conviction. A disagreement in committee addresses a scoring line, not a preference.

Second change: the budget becomes proportional to the gain. Step 5 produces an estimated net monthly ROI that you can compare term by term to the quotes received. If a provider proposes a forty-thousand-euro deployment for an estimated net gain of five hundred euros per month, you know immediately. The information asymmetry with the AI consulting market shrinks, without prior technical expertise.

Third change: risk becomes bounded. The risk provision from step 5, the GDPR and AI Act validation from step 4, the supervision from step 6 are not formalities. They absorb foreseeable incidents, and they allow you to respond to a CNIL inspection or external audit without panic. Compliance is no longer a hidden post-deployment cost; it is costed into the initial calculation.

Fourth change: the gain becomes measurable. The step 7 baseline, set before the pilot, transforms declared ROI into observed ROI. You no longer depend on a provider's narrative. You have two comparable measurements, produced by your team, reproducible at the next budget review.

Fifth change, deeper: the automation decision enters the normal scope of SME financial management. It ceases to be a technology bet. It becomes an investment, with a ticket, a payback, a documented risk, a written deliverable. That is the condition for a CFO, an accountant, an auditor to accept responsibility for it.

Frequently asked questions#

Where do I start if I have never deployed AI?#

With measuring the hours, not choosing a tool. The minimum sequence is four steps: list the three functions consuming the most monthly hours, isolate within each the three most repetitive tasks, score these nine candidates on volume, repetitiveness and data sensitivity, retain the highest-ranked. This sequence takes half a day and requires no technical tool. It provides the scope for a first pilot. The /free-ai-audit diagnostic executes this sequence in ten minutes with a structured questionnaire and delivers a written report.

Do I need a data team to get started?#

No, not for the first process. The cases that score highest (meeting notes, social reformatting, inbound lead qualification) rely on simple textual data, with no complex pipeline. A data team becomes useful for the second or third process, when volume crosses multiple internal sources. To get started, the critical skill is not technical. It is the ability to scope a pilot, set a baseline, maintain a review supervision at twenty to thirty percent of outputs for one to two months. This skill is what the Perello team provides.

What budget should I plan for the first process?#

Order of magnitude, for an SME of ten to thirty people: between three thousand and twelve thousand euros all-in for the first year, excluding internal time. This budget covers scoping, technical component licences, integration, user training and supervision for the first three months. As a public benchmark, Bpifrance's Diag Data IA is positioned around thirteen thousand euros excluding tax with partial subsidy, for a full audit scope, not a deployment¹⁴. The expected net ROI must cover this budget in under six months, otherwise the project changes target.

How long from decision to operational pilot?#

Four to eight weeks for a well-scoped first pilot. Weeks 1-2: hours mapping, task scoring, process selection, data validation, GDPR and AI Act scoping¹⁰. Weeks 3-4: baseline measurement on the current state, technical component selection, light integration, supervision protocol drafting. Weeks 5-8: actual pilot on a sub-process. Going faster means skipping the baseline or data validation, two shortcuts that always cost more in rework than they save in time.

How do I measure real ROI and not a marketing figure?#

By setting a quantified baseline before the pilot and measuring the same indicators after. Minimum indicators: actual hours consumed by the process over a reference month, number of quality incidents, human rework rate, internal satisfaction scored out of five. After the pilot, you compare term by term. A credible ROI is a documented gap between two measurements, not a percentage announced by a provider. Article 1 publishes the associated costing formula, reproducible by any auditor from INSEE⁴, France Strategie² and OECD³ sources.

Which process should I not automate first?#

Any process that combines two criteria: sensitive personal data (health, HR, data protected by professional secrecy) and absence of a reproducible pattern (fine contextual decisions, human arbitrage). This is the red zone of the AI Act classification, high-risk systems or borderline cases¹⁰, which requires heavy documented governance, incompatible with a first project. Also to avoid: tasks with high direct relational content (negotiation, HR interviews, complex client relations) where automation degrades perceived quality. These processes come fourth or fifth, not first.

What happens if the pilot fails?#

It is a frequent and anticipated outcome. Three cases. Estimated-to-observed gain gap above sixty percent: you stop, you document the gap, you capitalise on the baseline. Gap between thirty and sixty percent: you narrow the scope, you swap a component, you extend by thirty days. Gap below thirty percent: you generalise. A pilot that fails cleanly costs between five and fifteen thousand euros and produces learning usable for subsequent processes. A pilot that succeeds without a baseline proves nothing, which is worse.

When do I move from pilot to generalisation?#

Four cumulative conditions. Estimated-to-observed gain gap below thirty percent. Quality perceived by internal users above four out of five across a sample of at least ten people. Zero compliance or security incidents during the pilot period⁹¹². Human supervision at twenty-five percent maintained for the first two months of generalisation, then reduced to ten percent in documented fashion. If a single condition is missing, you do not generalise. You extend, or you redefine the scope.

Take action#

The method fits on one page. The seven steps, the four mistakes to avoid, the case applied to a fifteen-person SME constitute a reproducible, auditable, committee-defensible framework. You can apply it alone, from your own figures, with the public sources cited at the end of this article.

If you want an outside perspective for the first iteration, our methodology presents the full approach. The team executes the scoping in four weeks and delivers a written report. The entry point is the AI audit, free and with no commitment, covering steps 1 to 3 on your organisation. The preliminary costing (monthly loss on automatable tasks) is detailed in article 1 of this series.

Request your AI audit ->

Read also#

• How much does a French SME lose on automatable tasks in 2026?
• What are the real risks of poor AI use in a French SME?

Sources and methodology#

This article is dated 13 April 2026. Sources were consulted on that date. A revision is scheduled every six months to integrate the most recent editions of the cited reports. All URLs below are clickable and public. Specific figures used in the body refer by superscript to the corresponding entry. Sources marked as reused from the previous article retain their initial verification.

Main French sources

International sources

About the author#

Laurent Perello runs Perello Consulting, an independent AI automation firm for French SMEs. After 25 years building products for the web, he now orchestrates ten AI agents that he pilots alone, with a production log published daily at perfectaiagent.xyz. He publishes his methodologies and pricing online so that every executive can decide with full information.

Orchestrator: Alpha -- Perello Consulting | 2026-04-17