The President’s Council of Advisors on Science and Technology predicts that U.S. companies will spend upward of $100 billion on AI R&D per year by 2025. Much of this spending today is done by six tech companies — Microsoft, Google, Amazon, IBM, Facebook and Apple, according to a recent study from CSET at Georgetown University. But what if you’re a startup whose product relies on AI at its core?

Can early-stage companies support a research-based workflow? At a startup or scaleup, the focus is often more on concrete product development than research. For obvious reasons, companies want to make things that matter to their customers, investors and stakeholders. Ideally, there’s a way to do both.

Before investing in staffing an AI research lab, consider this advice to determine whether you’re ready to get started.

Assuming it’s your organization’s priority to do innovative AI research, the first step is to hire one or two researchers. At Unbabel, we did this early by hiring Ph.D.s and getting started quickly with research for a product that hadn’t been developed yet. Some researchers will build from scratch and others will take your data and try to find a pre-existing model that fits your needs.

While Google’s X division may have the capital to focus on moonshots, most startups can only invest in innovation that provides them a competitive advantage or improves their product.

From there, you’ll need to hire research engineers or machine learning operations professionals. Research is only a small part of using AI in production. Research engineers will then release your research into production, monitor your model’s results and refine the model if it stops predicting well (or otherwise is not operating as planned). Often they’ll use automation to simplify monitoring and deployment procedures as opposed to doing everything manually.

None of this falls within the scope of a research scientist — they’re most used to working with the data sets and models in training. That said, researchers and engineers will need to work together in a continuous feedback loop to refine and retrain models based on actual performance in inference.

The CSET research cited above shows that 85% of AI labs in North America and Europe do some form of basic AI research, and less than 15% focus on development. The rest of the world is different: A majority of labs in other countries, such as India and Israel, focus on development.