SR&ED for Machine Learning: Algorithms vs. Applications

Artificial Intelligence (AI) and Machine Learning (ML) are currently the hottest sectors in tech. Unsurprisingly, the CRA sees a massive volume of SR&ED claims involving neural networks, LLMs, and predictive models.

However, there is a fundamental disconnect between what startups consider "AI" and what the CRA considers "experimental AI."

The "Application" Trap

If your company builds a wrapper around an existing model (like an OpenAI GPT API, Anthropic, or Hugging Face model) and fine-tunes it on specific customer data to generate business value, you have built an application.

While this application might be highly lucrative, the underlying technological uncertainties regarding how the model works have already been solved by the creators of the foundation models. Feeding prompt instructions or proprietary data into a commercially available neural network is considered routine usage of an existing tool.

True Machine Learning SR&ED

To qualify for SR&ED in the machine learning space, you must typically be engaging with the algorithms themselves, attempting to push past their known theoretical or practical limitations.

Examples of Eligible ML Projects:

1. Algorithmic Efficiency: Modifying the mathematical structure of a deep learning algorithm to reduce GPU training time by 90% without losing inference accuracy, because existing architectures cannot scale to your required processing window.
2. Novel Topologies: Inventing a completely new neural network topology because existing Convolutional Neural Networks (CNNs) fail to identify specific anomalies in highly noisy, degraded satellite imagery.
3. Hardware-Constrained Inference: Developing custom quantization and pruning algorithms to force a massive language model to run on edge devices (like isolated mobile phones) with severe memory constraints, where existing compression techniques cause catastrophic failure.

The Data Pre-Processing Hurdle

Often, the actual "training" of the model is routine, but the preparation of the data is experimental. If you are dealing with unstructured, disparate data formats and must invent novel algorithms to align, clean, and vectorize that data before it can even be fed into a model, that pre-processing phase often constitutes robust SR&ED.

Conclusion

When claiming AI, ignore the final product's output (e.g., "The model writes perfect legal contracts"). The CRA doesn't care about the contracts. Focus exclusively on the structural and algorithmic hurdles your math and engineering teams faced when building or modifying the engine itself.