AI is now on the agenda of many companies with an R&D centre or planning a funded R&D project. But writing the technical part of a project involving AI brings challenges different from conventional R&D: "Is using a ready-made model R&D?", "Where is the uncertainty?", "How will we measure success?"
This article walks through the technical design of an R&D project involving AI step by step. The application form and financial rules vary by funding programme; here we focus on the technical framework that works for any programme.
The first question: is this really R&D?
The OECD Frascati Manual (2015), the international reference, expects five criteria to be met together for an activity to count as R&D:
| Criterion | Meaning | Question for an AI project |
|---|---|---|
| Novel | Aimed at new findings | Does this solution to this problem already exist in research or on the market? |
| Creative | Original, non-obvious concepts and hypotheses | Is our approach just a straightforward application of existing methods? |
| Uncertain | Outcome not known in advance | Are we sure we'll succeed? If not, why not? |
| Systematic | Planned and budgeted | Are work packages, timeline and resources defined? |
| Transferable / reproducible | Results can be transferred or reproduced | Are experiments documented and reproducible? |
A practical warning: Building an interface around a ready-made language model or cloud service usually isn't R&D; it's software development or integration. R&D starts where existing methods fall short.
Step 1: Define the problem and the limits of existing solutions
The first paragraph of a good technical plan isn't "we'll use AI" but an unsolved problem:
- What is the problem, and who does it affect, how?
- How is it solved today (manually, rule-based, existing products)?
- Why do existing solutions fall short? (accuracy, speed, cost, data privacy, language support, industry-specific data structures…)
This section should rest on a review of research and existing solutions. A claim that "no work exists on this" that a quick search can refute costs the reviewer's trust.
Step 2: State the technical novelty and uncertainty clearly
Write the novelty as a concrete technical claim, not in abstract terms:
- Weak: "We will revolutionise quality control with AI."
- Strong: "On production lines with very few defect examples, we aim to detect defect types the current rule-based system misses by developing a computer vision approach that works with few labelled examples."
Then state the uncertainty: which technical question is still unanswered? For example, "It is unknown whether the target detection rate can be reached with few examples", or "It is uncertain whether a small on-premise language model can reach sufficient accuracy on industry-specific documents."
Uncertainty isn't a weakness of R&D; it's its defining feature.
Step 3: Describe the method and data requirements
- Data: What data will be used, where will it come from, how ready is it? Is labelling needed? If personal data is involved, how will it be processed lawfully?
- Approach: Which methods will be tried, and why? Naming alternatives to compare, rather than committing to a single method, shows the uncertainty is being managed.
- Infrastructure: Where will training and inference run (cloud, on premises)? Are GPUs needed?
- Baseline: What will the new approach be compared with? The current system, a simple model or a known method.
Step 4: Set measurable success criteria
"Better results" isn't a goal. For each goal define:
- Metric: Accuracy, precision, recall, response time, cost…
- Comparison: The target level relative to the baseline
- Test conditions: On which dataset and under what conditions it will be measured
Realistic targets matter; impossible numbers cause problems in the evaluation and in final reports.
Step 5: Use TRL to set the starting point and target
The Technology Readiness Level (TRL) is a 1–9 maturity scale also used in EU research programmes. Stating the starting and target levels clarifies scope:
| TRL | Definition (summary) |
|---|---|
| 1–2 | Basic principles observed, technology concept formulated |
| 3 | Experimental proof of concept |
| 4 | Technology validated in the lab |
| 5–6 | Validated and demonstrated in a relevant environment |
| 7 | System prototype demonstrated in an operational environment |
| 8 | System complete and qualified |
| 9 | Actual system proven in an operational environment |
A definition like "from TRL 3 (proof of concept) to TRL 6 (demonstration in an environment close to a real production line)" tells both reviewers and the team what will be delivered.
Step 6: Build the work packages
A typical work package structure that works for AI projects:
- Data preparation and analysis: Collection, cleaning, labelling, exploratory analysis
- Method development and experiments: Trying and comparing alternative approaches
- Model optimisation: Improving the chosen approach under resource and speed constraints
- Integration and prototype: A prototype connected to the real system
- Testing and validation: Measurement under real or realistic conditions
- Documentation and dissemination: Technical reports, reproducibility
For each work package, write the output, duration, owner and risk with a plan B. For example: "If labelled data proves insufficient, synthetic data generation or semi-supervised learning will be tried."
Common mistakes
- Mistaking a tool name for novelty: "We'll use a large language model" isn't novelty on its own.
- Hiding uncertainty: A project written as if everything is certain looks like an implementation project, not R&D.
- Underestimating data: Most projects slip when it turns out the data isn't ready.
- No success criteria: A goal that can't be measured can't be defended in the final report.
- Bringing in the technical team last: The technical design should be written together at the very start of the project idea.
Conclusion
The strength of an R&D project involving AI comes from a clear problem, explicitly stated technical novelty and uncertainty, measurable targets and a work plan that knows its risks.
In our R&D Technical AI Partnership module, we write the AI design of the project idea, the method and work packages with you, and take part in the project as technical adviser. Our founder Çağrı Aksu has prepared 5 R&D projects funded by TÜBİTAK. Applications, financial reporting and monitoring are handled by your R&D unit or funding adviser.
Sources