AI adoption is accelerating. The July 30, 2026 update of the France Num Digital and Artificial Intelligence Barometer reports that 26% of French micro-enterprises and SMEs now use AI solutions. The practical question is no longer whether to experiment, but which workflow deserves an actual investment.
The EU AI Act also became applicable on August 2, 2026, subject to specific exceptions and deferred deadlines. Relevant transparency requirements therefore need to be considered when designing an AI-enabled process, as explained on the European Commission’s official AI Act page. This playbook is designed to produce an evidence-based decision before production deployment.
The starting point
Consider an illustrative scenario: a shared inbox receives customer requests. An AI workflow could identify the request category, extract relevant information, prepare a draft reply, create a CRM task and route the case to the appropriate employee. A human would remain accountable for approving customer-facing messages and consequential actions.
The target is not to “have an AI agent.” It is to produce a decision file containing:
- the workflow’s full cost;
- current processing time and measured test time;
- the rate of usable outputs;
- recorded errors and exceptions;
- a decision to stop, revise or industrialise the workflow.
Set a maximum budget before development starts. According to the Bpifrance Conseil white paper presented by France Num, an AI “quick win” automating an existing process has an estimated development cost of €10,000 to €50,000. More critical projects are allocated average budgets of €50,000 to €100,000. These public benchmarks are not D1 Consulting price quotations.
| Cost item | What it covers |
|---|---|
| Scoping | Current process, data, risks, metrics and target scenario |
| Implementation | Workflow, AI model, connectors, rules and interfaces |
| Acceptance testing | Test sample, business review and error handling |
| Operations | Licences, API usage, hosting, monitoring and maintenance |
| Internal effort | Workshops, approval, data preparation and support |
Prerequisites
Before the first workshop, the executive sponsor should assemble:
- A named process owner who understands the rules, exceptions and consequences of an error.
- A baseline built from real data: workload, processing time, rework, delays, external expenditure and incidents.
- A representative sample of real cases, anonymised where required and including incomplete, ambiguous and incorrect inputs.
- Technical access, preferably read-only at first, to the relevant mailbox, document store, CRM, ERP or business application.
- Control decisions: actions requiring human approval, prohibited data and the shutdown procedure.
- The loaded hourly cost of affected employees, supplied or approved by finance.
Access rights, logging and output controls should not be postponed until after the prototype. The French data protection authority’s AI security guidance recommends risk analysis, appropriate logging and controls over AI outputs when the system is connected to an information system. The ANSSI security recommendations for generative AI systems provide additional technical guidance.
The step-by-step process
The plan below is an operational allocation based on a current public benchmark. Bpifrance’s Data AI Diagnostic consists of eight person-days delivered over no more than three months. Its cost is €10,000 excluding VAT, with 40% public funding subject to eligibility, leaving €6,000 excluding VAT for the company.
1. Measure the current work
Who does what: the process owner describes how cases are actually handled, while the project manager extracts workload, waiting time, rework and manual intervention data. Measure operational reality rather than the official procedure. Proposed duration: one person-day within the eight-day public framework described by Bpifrance and France Num. Deliverable: a baseline sheet showing the current cost per case and the main operational bottlenecks.
2. Select a task, not an AI promise
Who does what: the business sponsor, project manager and integrator select a workflow with identifiable inputs, outputs and accountability. A suitable first target is repetitive, reviewable and reversible. An agent expected to “help every department” or act freely across several systems is too broad. Proposed duration: one person-day within the same eight-day Bpifrance framework. Deliverable: a use-case card defining the trigger, data, expected result, exceptions and human approver.
If several architectures appear viable, review our comparison of an integrated suite, agent platform and custom development.
3. Set data and risk boundaries
Who does what: the business team identifies essential information; the CIO, security lead or IT provider checks permissions and data flows; the data protection officer participates when personal data is involved. Classify every capability as reading, proposing or executing. Proposed duration: one person-day within the eight-day benchmark presented by France Num. Deliverable: a data-risk matrix listing authorised sources, retained logs, prohibited actions and mandatory approvals.
4. Design the target workflow and its exceptions
Who does what: the process owner describes the normal route and exceptions; the integrator translates those rules into n8n, Make or Power Automate; the project manager assigns each control. AI should only be used where deterministic rules are insufficient—for example, interpreting text, classifying a request, extracting information or drafting content. Proposed duration: one and a half person-days within the eight-day Bpifrance benchmark. Deliverable: a workflow diagram, business-rule list and exception register.
5. Test real cases in an isolated environment
Who does what: the integrator configures a limited version; business users run the test set without being shown the expected AI outputs; the project manager records every correction, rejection and workaround. Proposed duration: two and a half person-days within the eight-day Bpifrance cycle. Deliverable: an acceptance report comparing the workflow with the baseline, including an error log and manual recovery procedures.
The NIST AI Risk Management Framework recommends defining the task, expected benefits, costs, human oversight and measurement methods, then testing AI systems before deployment and throughout operation.
6. Make a go, no-go or retest decision
Who does what: the business sponsor decides with finance, IT and the project manager. They compare the calculated value with the full cost, then review quality and risk separately. Financial benefits do not offset a potentially severe business error. Proposed duration: one person-day, bringing the total to the eight person-days defined for the Data AI Diagnostic. Deliverable: a signed decision, budget, accountable owner, acceptance plan and, if approved, a production roadmap.
How to measure whether it works
Set the thresholds before testing so that results cannot be reinterpreted afterwards.
- Net time saved = previous human processing time − approval time − exception-handling time. Multiply the result by actual workload.
- Accepted-output rate = outputs usable without substantial correction ÷ outputs reviewed. Record the reason for every rejection.
- Exception rate = cases escalated to a person or blocked ÷ cases processed. This often reveals an excessively broad scope.
- Net value and payback = valued time savings + avoided expenditure − recurring costs; then initial investment ÷ periodic net value.
The dashboard should keep productivity, quality and risk separate. The method described in our article on measuring automation ROI can be reused for an AI workflow by adding approval, supervision and model-usage costs.
Mistakes that cause the project to fail
- Starting with the tool. Purchasing a licence before defining the workflow forces the business requirement to fit available features.
- Automating an inconsistent process. AI then reproduces duplicate work, unclear ownership and unmanaged exceptions. Begin with the principles of business process optimisation.
- Testing only clean cases. Acceptance testing must include the missing documents, ambiguous wording, duplicates and unexpected formats found in normal operations.
- Confusing model cost with project cost. API usage is only one component. Integration, permissions, testing, monitoring, maintenance and internal effort belong in the full cost.
- Hiding human review time. If every output requires extensive correction, the expected productivity benefit disappears. Measure that time explicitly.
- Giving AI write access too early. Start with reading and proposing. Only allow changes to a CRM, ERP or contractual document once controls have passed acceptance testing.
- Leaving ownership undefined. A workflow without a business owner deteriorates after the first process or system changes. Our article on managing an IT project without a CIO explains which responsibilities must be formalised.
Getting support
D1 Consulting can handle use-case scoping, baseline measurement, workflow design, the selection of n8n, Make or Power Automate, AI-agent integration, access controls, human approvals, acceptance testing and performance indicators through our Process Automation and Optimisation service.
Where several departments, applications or suppliers are involved, our IT Project Management service covers project scoping, planning, arbitration, acceptance and production deployment. If public funding is being considered, D1 Consulting is also a France Num Activator and can help prepare the project’s technical documentation without prejudging eligibility.
👉 Book a free 30-minute diagnostic to select an AI workflow, calculate its full cost and define the evidence required before investing.
