If your company is leaning into the exploding opportunities created by artificial intelligence (AI) and machine learning (ML), you’re probably excited about what it could do for your business. Lots of people and companies are champing at the bit for AI/ML products and services, especially user-friendly, trustworthy ones. If you can find a way to sate that appetite, there’s absolutely money to be made.
And because so much in the world of ML and AI falls under the umbrella of research and development, you might be able to shave some money off your taxes along the way, too.
The federal R&D tax credit allows you to claim a credit for R&D activities and expenses, from the salaries of people with certain job titles to the raw materials you use in your research. All told you might be able to slash your tax bill by up to $500,000.
Don’t be too hasty here, though. In the last year, we’ve seen some companies run into issues because they assume all of their ML or AI work is qualifying R&D — and it isn’t.
Where you started is key
Newness is a huge component of qualifying for the R&D tax credit. In fact, two of the four tests accountants often use to determine if something counts toward the tax credit focus on creating something new.
One of those tests is the “elimination of uncertainty” test, which essentially means you’re undertaking R&D to discover new information and clarify something you didn’t know before.
Another test is the “new or improved business component” test, which evaluates whether you’re using the R&D to create a new business component or to better an existing one. (As a quick aside, you can’t be adapting an existing business component to a specific customer’s request — qualifying R&D has to be aimed at improving the component in general.)
Long story short, qualifying research expenses have to be geared toward discovering or creating something new. Great, you might be thinking since so much in the world of AI or ML is new.
But the issue we’re seeing companies come up against is their starting point. If you’re creating your own large language model, for example, you can likely count the salaries of the people involved toward your credit. But if you’re using an existing LLM, things get trickier here. Because the work you’re doing is predicated on someone else’s work, the newness tests — particularly the elimination of uncertainty one — can be difficult to pass.
This can get pretty nuanced. If, for example, you’re creating a new user interface that links with an existing LLM, the part of the work dedicated to the new UI can likely qualify for the tax credit. Talking with your accountant can help you get a good idea of what you can and can’t claim.
Don’t wait to have that conversation, either. Claiming the tax credit requires careful recordkeeping — and it’s a whole lot easier to gather up the documentation you need as you do the work rather than during filing season.
Claimable employee wages
Most companies who hit the $500,000 maximum on the R&D tax credit do so largely through the salaries they pay. Because of that, we wanted to call out a few things all AI and ML companies should know about which wages they can claim.
For starters, job titles matter a lot here. The salaries for computer programmers, engineers, technicians, CTOs, and software developers should all count relatively easily toward your credit.
But if you’re using an existing deep learning algorithm, your team might have just a few of those job titles, backed by a larger team of marketing and sales personnel. Even if the salesperson or marketing pro engages with your AI or ML work, their salary likely won’t count toward your credit. For wages to qualify, that employee needs to be actively engaged in work that’s technological in nature and that uses the process of experimentation.
Where your team is located also matters. Salaries only count as a qualified research expense if they’re going to a person working in the U.S. If you hire developers, engineers, etc. who live and work internationally, their wages won’t count toward your credit total.
Claiming compute
In some cases, computer rental expenses can count as qualified research expenses. That means you might benefit from renting servers and data center equipment rather than purchasing it outright. Once you buy it, the IRS considers it a depreciable asset, disqualifying it from the R&D tax credit bucket.
If you’re renting, the expense could be a qualifying R&D one — in certain circumstances. Per the Internal Revenue Code, if you’re paying someone else to use an off-site computer, owned by another entity, and of which your company isn’t the primary user, it qualifies. This regulation went into effect in 1985 when most companies didn’t have on-site computers. It sat dormant for a long time, but with cloud computing on the rise, it’s relevant again.
That said, your ability to claim compute expenses as part of your credit total depends on quite a few nuanced factors. For example, proving that you’re not the primary user of that specific rented server might be key in applying that cost toward your credit.
We advise ML and AI companies to not make any assumptions here. Have a chat with your accountant about what compute costs you might be able to count toward your credit.
Obviously, there’s a lot to consider for AI and ML companies that want to maximize their R&D tax credit. We’re here to help. Our team of R&D specialists has worked with tech companies in leading-edge spaces for years, so navigating artificial intelligence and machine learning is nothing new for us. We can evaluate your company’s R&D activities to figure out how much you can claim — and to help you maximize this tax savings opportunity.
To get started with a fixed-fee R&D tax credit study or to simply have a conversation with our team of specialists, contact us.