A research and development project is not managed like an integration or deployment project. Technical uncertainty is structural, not accidental. Every design decision commits resources based on assumptions that may be invalidated along the way. Successfully completing an R&D project therefore requires a methodological framework adapted to this reality, not a transposition of classic project management methods.
AI Governance and AI Act: the regulatory constraint that redefines R&D framing
An R&D project that involves artificial intelligence models is no longer solely the responsibility of the technical management. Since February 2, 2025, the AI Act imposes governance obligations on organizations that develop or deploy AI systems. Specifically, the initial framing of the project must include identifying the company’s role (provider or deployer), training the relevant teams, and ensuring traceability of uses.
When the project targets a system classified as high-risk, compliance must be considered from the design stage. The requirements focus on the quality of datasets, technical documentation, human oversight, and security. We recommend formalizing these constraints in the initial R&D brief, rather than at the end of the cycle, when technical decisions are already locked in.
This logic of compliance by design applied to AI projects also modifies the choice of models and the conditions for reusing data produced during the exploratory phase. Meanwhile, the presentation from L’Essentiel Pro details the milestones to establish in order to structure an R&D project from its early phases.
Structuring hypotheses before the experimental protocol
We regularly observe teams that embark on constructing a protocol without stabilizing their hypotheses. The result: costly iterations, unusable data, and difficulty demonstrating the coherence of the work in case of a tax audit (CIR).
The sequence to follow is strict. First, the literature review validates or invalidates the existence of prior work. Next, the problem statement is finalized, and the hypotheses are formulated in a testable manner. Only at this stage does the experimental design take shape.
What the CIR file really expects
Each project must be linked to a documented scientific or technical uncertainty. The administration is not looking for a synthetic report written afterward. It checks for the existence of contemporary traces: dated hypotheses, test results, pivot decisions, involved individuals. Granular time tracking, associated with each research task, significantly strengthens the robustness of the file.
R&D documentation is not delegated to an external writer at the end of the year. It is produced throughout the project, by the researchers and engineers themselves. We recommend appointing a documentation manager from the kick-off, with direct access to lab notebooks or code repositories.
R&D project methodology: choosing between iterative and phase-gate
The methodological choice depends on the level of uncertainty and the size of the team. Two approaches dominate R&D project management, and they are not interchangeable.
- Iterative approach (Agile/spiral): suitable for projects with many technical unknowns. Each iteration produces learning that redefines the objective of the next. Small, co-located teams benefit the most from this.
- Phase-gate approach (Stage-Gate): relevant when the project follows an identifiable linear progression, with validation milestones before each commitment of additional resources. Distributed teams or projects subject to strong regulatory constraints find a more readable framework here.
- Hybridization: some projects combine both, with exploratory phases in iterative mode and formal gate transitions before industrialization or compliance phases.
A common pitfall is to impose an Agile method on a project whose regulatory milestones require sequential deliverables. The methodology must serve the nature of the project, not the other way around.
Traceability of R&D work and anticipating tax audits
The preparation of a CIR file is becoming increasingly granular. The required justifications are no longer limited to a general description of the work. The administration demands the ability to link each project to a documented state of the art, a traced experimental approach, and results, including negative ones.
Elements to produce continuously
- Dated hypothesis sheets, with validation or invalidation criteria defined in advance
- Timestamped test logs (lab notebooks, Git commits, automated test reports)
- Time tracking by collaborator and by task, ideally with weekly granularity
- Decision reports during progress meetings, mentioning pivots and their technical justifications
A well-documented project defends itself during an audit. Teams that document continuously spend less time reconstructing their file and significantly reduce the risk of reassessment.
The most common mistake remains considering CIR documentation as a separate administrative task from the project. In reality, a properly managed R&D project naturally produces most of the necessary evidence. Project management and tax compliance converge when research objectives, test results, and decisions are formalized as they occur.
The last point not to be overlooked concerns the preservation of raw data. Source files, datasets, and intermediate versions must be archived with a clear versioning system. In the event of an audit several years after the declaration, only contemporary and non-reconstructed evidence holds probative value.
