Blog›AI & Innovation ›Leveraging AI for innovation (and when your project can qualify as R&D)

Artificial intelligence has become the default tool for innovation. Many companies in Chile are building models, automating processes and launching AI-based products. A question we hear often is: can that qualify under the R&D Law?

The short answer: sometimes yes, sometimes no. It depends on whether your project uses AI or creates new knowledge with AI.

Using AI is not doing R&D

The R&D Law doesn’t reward adopting technology; it rewards generating knowledge. So it helps to separate two kinds of project:

The law expressly excludes “routine, repetitive or minor improvements, adaptations and analyses, even if they use technology”. Implementing an off-the-shelf AI solution, even one that is new to your company, usually falls under that exclusion.

When an AI project can qualify

The law expressly includes software development, as long as it:

• systematically resolves a scientific or technological uncertainty, or
• produces a substantial, innovative improvement in a process, product or service.

In an AI project, technological uncertainty usually shows up in questions like:

• Can we reach a given accuracy with the data we have, which is scarce, noisy or of a little-studied kind?
• Can the model run within real latency, cost or hardware constraints?
• Which approach works when published methods don’t transfer to this domain?

If your team doesn’t know in advance whether the solution is technically possible, and pursues it with a systematic method (hypotheses, experiments, measurement), you are probably looking at R&D.

A documented example

A University of Chile thesis (Olivares, 2014) analyses the case of a telecom company that applied with a project to detect large-scale network failures from user behaviour patterns, instead of the equipment’s traditional alarms. The problem had no known solution and required developing new algorithms: a good example of software that resolves a technological uncertainty.

Which AI project costs can be included

Within Corfo’s accepted expense categories, an AI project typically includes:

• Salaries of the data science and engineering team, according to their time on the project.
• IT services: cloud compute, storage and training services.
• Subcontracts with universities, centres or companies for specific R&D tasks.
• Equipment: servers or GPUs bought for the project, as capital expenditure.
• Intellectual property: the cost of registering software that comes out of a certified project.

With the benefit of 35% as a credit and 65% as an expense, the government can effectively fund about half of the project. We explain how in What is a tax credit?.

How to document an AI project for Corfo

Software and AI projects are often rejected because they are described as products, not as research. A few practices help:

1. State the uncertainty explicitly. “We don’t know whether X is possible with Y” is stronger than “we will build an AI platform”.
2. Review the state of the art. Show what exists today and why it doesn’t solve your problem.
3. Define hypotheses and metrics. For example: “a model trained on this data will reach at least 90% accuracy in this case”.
4. Keep an experiment log. What was tried, what failed and what was learned. Failures are evidence of R&D too.
5. Separate R&D from product development. The interface, integration and commercial rollout usually aren’t R&D; the algorithmic core can be.
6. Make spending traceable. Use a dedicated cost centre for the project and record the team’s time allocation.

AI can also help you apply

Beyond the project itself, AI is useful for managing the process: organizing the state-of-the-art review, structuring the experiment log or drafting the technical write-up. Two cautions:

• Confidentiality: don’t upload sensitive technical or financial information to tools that don’t control how your data is used.
• Expert review: the application must accurately reflect the real project. Generic text produced without judgment is exactly what Corfo tends to flag as a “routine improvement”.

In short

• Adopting AI does not qualify on its own; creating new knowledge with AI can.
• The key is showing a technological uncertainty and a systematic method to resolve it.
• Team, compute, subcontracts and equipment can be included as project costs.
• Document from day one: hypotheses, experiments, results and costs.

This article is for information only and is not tax advice for your specific case. Sources: Law 20,241, Certification Regulation (Supreme Decree 102/2012), Corfo Exempt Resolution No. 17 and Olivares (2014), “Análisis del incentivo tributario a la I+D”, University of Chile.