The Baromètre France Num 2026 : le numérique et l’intelligence artificielle dans les TPE et PME, updated on 12 October 2026, reports that 40% of French small businesses use at least one AI solution. However, only 11% use AI to automate tasks and 13% use it to analyse or classify documents. The survey covered 9,655 businesses.
France Num also listed email handling and answers to recurring customer questions among the practical use cases presented at the Salon SME on 13 and 14 October 2026. The practical opportunity is no longer to ask an assistant to write an isolated email. It is to embed AI in a controlled and measurable workflow connected to the company’s information systems.
The starting point: turn a shared inbox into a work queue
Addresses such as sales@, support@ or orders@ often create the same sequence of manual operations: read the message, identify the customer, classify the request, retrieve the relevant file, draft an answer, assign follow-up work and update the CRM or ERP.
The target AI agent should not replace this chain with uncontrolled automatic replies. It should:
- identify the request category and urgency;
- extract references without inventing missing data;
- consult authorised internal sources only;
- prepare a reply or operational action;
- submit its proposal to an employee;
- execute the approved action and retain an audit trail.
The expected result must be quantified from an observed baseline rather than a generic efficiency promise:
Current monthly cost = eligible email volume × active minutes per email ÷ 60 × loaded hourly cost.
An email is eligible when its handling follows a documented procedure and the required information is available. Complex disputes, HR requests, bank detail changes and contractual commitments should remain outside the initial scope.
If the process is still unclear, use the method in our guide to validating a first profitable AI workflow before selecting a platform.
Prerequisites: what must be available before development
The executive sponsor appoints a business owner for the inbox. That person must be able to decide which requests are included, which actions require approval and which messages must immediately be escalated.
The project team then collects:
- a representative set of historical emails, final replies and assignments;
- the request categories genuinely used by the team;
- read access to email, CRM, ERP and knowledge sources;
- approved reply templates;
- loaded labour costs for the employees involved;
- expected licence, model, hosting and maintenance costs;
- an explicit list of prohibited data, actions and recipients.
Historical data should be pseudonymised when full identities are unnecessary for testing. The DPO or security manager must confirm what data is sent to the model, how long it is retained, who can access it and which logs are available. The CNIL’s Guide d’auto-évaluation pour les systèmes d’intelligence artificielle calls for a clear purpose, qualified data and assigned responsibilities before production.
Choose the architecture before choosing the vendor: a Power Automate workflow in Microsoft 365, a self-hosted n8n orchestration, a Make scenario or custom development. Our guide to choosing between an AI suite, automation platform and custom development provides the corresponding decision criteria.
The step-by-step process
The following plan represents a D1 envelope of eight expert-days for a straightforward pilot, excluding custom connectors and major data remediation. This is aligned with the eight-day expert engagement used for the French Diag Data IA programme, as described in the renforcement du plan national « Osez l’IA ».
1. Define what the agent is allowed to process
Who does what. The sponsor sets the economic objective. The inbox owner lists the request categories. The project manager separates eligible messages, exceptions and sensitive decisions.
Duration. One expert-day within the envelope based on the renforcement du plan national « Osez l’IA ».
Deliverable. A scope sheet documenting inputs, outputs, escalation rules, authorised actions and baseline metrics. An unknown request must be escalated rather than forced into the closest category.
2. Map data and access rights
Who does what. The business owner identifies the authoritative sources. The integrator documents connectors and technical accounts. The DPO or security manager reviews personal data, confidential information and external service conditions.
Duration. One expert-day, using the benchmark provided by the renforcement du plan national « Osez l’IA ».
Deliverable. A matrix showing, for each source, what the agent may read or write, which technical identity it uses, how logs are retained and who approves the proposed action.
3. Build the human-approved workflow
Who does what. The integrator configures the inbox trigger, initial filtering, AI classification, reference extraction, authorised information retrieval and draft creation. The business owner writes deterministic rules that the model must not interpret freely.
Duration. Two expert-days within the envelope derived from the renforcement du plan national « Osez l’IA ».
Deliverable. A testable workflow that places each proposal in an approval queue. At this stage, the agent must neither send external replies nor modify business data autonomously.
4. Test normal, ambiguous and hostile messages
Who does what. Business users replay historical emails and record the expected category, extracted fields and changes made to the draft. The security lead tests spoofed senders, unexpected attachments, hidden instructions and prompt-injection attempts.
Duration. One and a half expert-days within the framework of the renforcement du plan national « Osez l’IA ».
Deliverable. An acceptance report listing successes, failures, causes and required corrections. The CNIL recommends controlling generative AI outputs because they may contain personal data or misleading information in its guidance on securing the development of an AI system.
5. Run the agent in shadow mode
Who does what. The agent processes messages alongside the existing team without affecting recipients. Employees record whether they would accept, edit or reject each proposal. The project manager analyses differences by request category.
Duration. One and a half expert-days spread across the observation period, within the benchmark from the renforcement du plan national « Osez l’IA ».
Deliverable. A discrepancy log and an explicit decision for every category: release, continue observation or remove from scope.
6. Enable production actions progressively
Who does what. The sponsor approves validated categories. The integrator first enables draft creation and assignment, followed by writes to business applications after human approval. Support receives a shutdown and manual fallback procedure.
Duration. One expert-day for controlled production release, within the envelope based on the renforcement du plan national « Osez l’IA ».
Deliverable. The published workflow, access matrix, monitoring dashboard, alerts, version log and manual fallback runbook.
How to measure whether it works
The number of drafts generated is not a business KPI. It measures tool activity, not value. Track indicators that can be calculated from operational logs:
- Average active handling time = total active minutes ÷ eligible emails processed.
- First-pass approval rate = drafts approved without substantial editing ÷ drafts reviewed.
- Misrouting rate = emails reassigned after AI classification ÷ emails classified by the agent.
- Monthly net value = value of labour hours avoided − licences − model calls − hosting − maintenance.
Review these metrics by request category. A satisfactory overall average may conceal poor handling of complaints or urgent orders. Include design, maintenance and replacement costs by applying the method in our guide to the full lifecycle budget of an automation project.
Mistakes that cause the project to fail
Allowing automatic replies in the first version. A plausible answer is not necessarily accurate. Begin with human-approved drafts and release only the categories that pass acceptance testing.
Combining every request type in one prompt. A quotation request, dispute and address change do not follow the same rules. Each category needs its own mandatory data, approved sources and escalation conditions.
Giving the technical account excessive permissions. An agent that drafts replies does not need permission to delete emails, change commercial terms or export the entire CRM.
Testing only clean messages. Real emails contain long signatures, forwarding chains, spelling errors, attachments, incomplete references and several requests in the same message.
Recording a heavily rewritten draft as time saved. Reading, correction and additional research remain part of the handling cost. Excluding them creates a misleading ROI.
Ignoring post-pilot costs. Model consumption, connector licences, monitoring, rule maintenance and regression testing continue after the initial build.
Getting implementation support
D1 Consulting can scope the eligible messages, measure the baseline, integrate email and business applications, build the workflow with n8n, Make or Power Automate, and manage testing and release through our Process Automation & Optimisation service.
When several teams, vendors or applications are involved, our IT Project Management service handles responsibilities, planning, acceptance testing, decisions and KPI tracking. The intended deliverable is not an AI demonstration: it is an operational, reversible and measurable business process.
👉 Book a free 30-minute diagnostic to select your pilot inbox and define the data, approvals and indicators required for its automation.
