SR&ED for AI Startups: What Can Qualify
‘We built an AI product’ is not an SR&ED claim. The claim is the hard technical problem, why existing methods didn’t solve it, and the systematic work your team did. AI startups often have strong SR&ED work, and often describe it badly.
Does AI development qualify for SR&ED?
AI and machine learning work can qualify when the team faces a technological uncertainty that known techniques don’t resolve, and investigates it systematically. Using an off-the-shelf model or API in a conventional way usually doesn’t. The same CRA tests apply as for any software (CRA’s eligibility guidelines).
Which AI work is more likely to qualify?
| More likely to qualify | Usually hard to support |
|---|---|
| Getting a model to meet accuracy, latency or cost targets that standard fine-tuning or retrieval couldn’t reach | Calling a hosted model API with prompts |
| New techniques for privacy, such as training or inference without exposing customer data | Building a chat interface around an existing model |
| Making unreliable model output dependable enough for a regulated workflow, with tests and failed approaches | Prompt tweaking without a technical hypothesis |
| Data pipelines for messy or scarce data where known methods failed | Standard labeling, data cleaning and dashboards |
How should you describe AI work?
Write the uncertainty as a technical question with a baseline: what known approaches you tried or considered, why they didn’t meet the requirement, and what you didn’t know. For example: ‘Retrieval-augmented generation with standard chunking returned the wrong policy clause 18% of the time on our benchmark; it was unknown whether structure-aware retrieval could get below 2% within a 300 ms budget.’
| Field | Entry |
|---|---|
| Uncertainty | Can structure-aware retrieval cut wrong-clause answers below 2% within 300 ms? |
| Baseline | Standard chunking and embeddings: 18% wrong-clause rate |
| Hypotheses | H1 clause-level chunks · H2 section graph · H3 reranker |
| Results | H1 failed (11%) · H2 2.9% at 420 ms · H3 1.6% at 280 ms |
| People and time | 2 engineers, 140 hours this month |
Which AI costs can count?
Salaries of people directly doing the eligible work are usually the biggest part. Cloud and compute costs, contractors and materials need care: they must relate to the eligible work, and contractor rules differ. See SR&ED eligible expenses.
Failed experiments are evidence, not waste.
The approaches that didn’t work are what show uncertainty. Keep them.
What should AI startups keep?
- Benchmarks and evaluation results, including failures.
- Experiment logs and model versions.
- Code branches for approaches you tried and dropped.
- Notes on why known methods didn’t work.
- Time spent by each person on each project.
A simple monthly routine is on SR&ED documentation. For positioning AI features to customers, see AI Features in SaaS, and for the general software test, does software qualify.
Frequently asked questions
Does AI development qualify for SR&ED?
It can, when it resolves technological uncertainty through systematic investigation, not when it uses AI in a standard way.
Does using the OpenAI or another model API qualify for SR&ED?
Using a hosted model API conventionally usually doesn’t. Solving a technical problem the API couldn’t might.
Does prompt engineering qualify for SR&ED?
Trial and error without technical hypotheses usually doesn’t. Systematic work on an unresolved technical problem may.
Can cloud and GPU costs count for SR&ED?
Possibly, if they relate directly to eligible work and follow the expenditure rules. Check with your advisor.
How do I describe AI work in an SR&ED claim?
State the technical uncertainty with a baseline, the hypotheses tested, results including failures, and what was learned.
What evidence should an AI startup keep?
Benchmarks, experiment logs, model versions, abandoned branches, notes on failed methods and time records.
Sources & further reading
Standards and platform rules change. These sources let you verify the current requirements directly. All screens shown are mock-ups of fictional products.