Where to start with AI in an SME: first workflow

To start using AI in an SME, choose a first workflow. Not a general “AI adoption” programme. Take a real process: meeting notes to structure, a visit to record in a project file, or a case file to prepare. Describe its input, output, the person who validates it and the data it uses. Install this workflow with the team that uses it. The rest comes later.

This starting point follows the three principles of the Operational Method. AI is a means, not the subject. A person reviews, corrects and decides. The system is documented, handed over and maintained. If a spreadsheet or a better habit is enough, no AI system is needed.

What is a first AI workflow?

A first AI workflow is a bounded sequence of work. Information comes in. The system transcribes it, classifies it or prepares a first draft. A person checks the result. The validated working document then goes into the tool already used for the work: a database, project file, report or dashboard.

The workflow is not “putting ChatGPT into the company”. That phrase names neither the input, the output nor the decision. “Turn raw meeting notes into decisions, actions, deadlines and points requiring attention, then have the report reviewed” describes a workflow. Everyone can see where it starts, what it produces and who is accountable for the result.

• The input: dictation, notes, a document or a form already in use.
• The transformation: transcribing, extracting, classifying or preparing a first draft.
• The output: a named working document, stored in the right place.
• The check: a person reviews, corrects and decides before the result is treated as authoritative.

How do you choose the first workflow?

The first workflow is chosen from existing work. Look at where hours are spent, where information is entered again, where the same questions recur and where a working document is missing when a decision has to be made. The choice concerns an observable situation, the people who experience it and the documents they use.

A process is suitable when its beginning and end can be written down. The expected quality must also be visible. A report can be checked: are the decisions present? Are the responsible people named and are the deadlines correct? An automated response sent directly to a client commits the company before it is reviewed. That is not the same first workflow.

The choice of tool comes later. Existing subscriptions may be enough. Another plan may be needed. The scoping phase sets this out. The system built around the engine then connects the documents, rules, validation and business tool.

Why are dictation and meetings good places to start?

The Operational Method page shows two situations. In the first, someone dictates the outcome of a visit. The workflow structures the speech, updates a file and prepares the next action. In the second, meeting notes become a report that separates decisions, actions, deadlines and points requiring attention.

These situations have the same architecture. The material already exists within the company. It arrives in a quick form, as speech or raw notes. AI prepares a structured form. The database and dashboard receive only what belongs there. A person reviews and decides.

The subject is therefore not dictation or the meeting in themselves. It is the transition from scattered information to a retrievable working document. The same examination can be applied to reporting, preparing a case file or a standard response. The input, output and validation change. The method remains clear.

How is the first workflow installed?

The process begins with a discussion about the situation. Next comes observation of how the work is done: the hours spent, the documents used, repeated actions and the people involved. The first workflow is then scoped around these real working documents.

1. Set out the situation. What is feasible and what is not are identified.
2. Look at the work. The process to equip with a tool first is chosen from existing practices.
3. Install the workflow. Real documents and real meetings are used to configure it with the team.
4. Hand it over. The team is trained, reviews, corrects and adjusts the workflow until it can run it day to day.
5. Observe. The workflow is observed in use. The time freed up, corrections and maintenance needs inform what comes next.

Each step is documented, dated and reviewed. The scoping phase sets a scope and an end date. The assignment installs the system in the selected setting. Ongoing support then covers the infrastructure built: supervision, fixes, monitoring and alerts for interfaces and models. These services and their prices are described on the Operational Method page.

Where does the human remain in control?

A person remains in control where the result becomes a decision, a commitment or reference information. AI can transcribe, prepare and write a first draft. The team reviews, corrects and decides. A deadline extracted from a meeting does not enter the management process on its own. A file prepared after dictation is still checked before it is treated as correct.

This check is not limited to a “validate” button. What the person checks, the errors they must be able to correct and the route to follow when the result is incomplete must be documented. Training takes place on the installed workflow. The documentation describes the rules, connected tools and actions that remain human.

What data leaves the company?

The installation is self-hosted in the client’s environment by default. The client’s database, rules and history remain there. When a workflow calls an external engine such as ChatGPT, Claude or Gemini, what is sent to it is set out in the written quotation.

The setup is documented and reversible. The written quotation also sets out the handover inventory: data, rules and documentation. This description distinguishes the engine subscription, paid to its provider, from the system built around it and installed within the company.

When should you decide on the next workflow?

The next workflow is decided when the first is used day to day and its results can be reviewed. Look at the corrections still required, the points where the team takes over, the incidents encountered and the time actually freed up. These observations show whether the workflow needs adjustment, maintenance or extension.

This step may also conclude that nothing should be added. A stable first workflow does not require a general programme to be opened. Another lack of method may be corrected with an instruction, a document template or a spreadsheet. AI remains the means. The work to be done remains the subject.

Which sources inform this approach?

Two public sources document starting with a specific use, its context and limits, training and ongoing risk management.

• CNIL, “Comment déployer une IA générative? La CNIL apporte de premières précisions”, 18 July 2024. The CNIL recommends starting with a concrete need, governing uses, taking the limitations into account, choosing a secure deployment method and training users.
• NIST, “Artificial Intelligence Risk Management Framework (AI RMF 1.0)” (NIST AI 100-1), 26 January 2023. The framework connects governance, description of the context, measurement and risk management throughout the system’s life cycle.