AI + R&D Experts: Why Automation-Only R&D Studies Fall Short

When a company decides to claim the R&D tax credit, it usually runs into two unappealing options. On one side are traditional consulting firms that do thorough work but charge accordingly and take months. On the other side are new automation-only tools that are fast and inexpensive but produce a study that is, too often, only as deep as a questionnaire. Neither is what most companies actually want, which is a study that is both efficient and defensible. The better answer is not to pick one side — it is to combine the strengths of AI with the judgment and accountability of experienced R&D experts.

The two-bad-options problem

Expensive consultants. The legacy model relies heavily on senior staff conducting interviews, drafting narratives, and assembling documentation by hand. The work can be excellent, but the cost and timeline reflect all that manual effort, which puts a quality study out of reach for many smaller companies and makes it painful even for larger ones.

Shallow automation-only tools. The newer model swings to the opposite extreme. Software walks a user through a form, applies some templated logic, and generates a study with minimal human involvement. It is fast and cheap. The problem is what gets lost: qualification under the R&D credit is a judgment-heavy exercise, and a tool that maps form answers to boilerplate can both over-claim routine work and under-claim genuine research — while producing documentation that does not hold up well under scrutiny.

The credit deserves better than a forced choice between thorough-but-slow and fast-but-thin. Our how it works page lays out how we structure the alternative.

Why human R&D experts finalizing the study matters

The R&D credit is not a self-executing calculation; it is a position on a tax return that you may have to defend. That is exactly why experienced R&D experts review and finalize the study before it is issued.

• Professional judgment on qualification. Deciding which activities meet the four-part test, which costs are qualified, and how much to claim requires applying tax law to messy facts. That is expert work — not something a form-to-template engine can do reliably.
• Accountability. People who understand the rules review the conclusions and stand behind the study. That accountability is precisely what an automated output cannot provide.
• Built for examination. If the IRS examines the credit, you want a study that experienced people understood and built with the exam in mind from the start — not a report nobody is responsible for.

This is the difference between a document that looks like a study and a study built to be defended. It is also why every Ricerca study comes with Audit Protection: if the IRS examines the R&D claim, we represent you and provide the documentation they request — limited to the R&D study and credits, and subject to your engagement agreement.

What AI should — and shouldn’t — do

The answer is not to reject technology. Used well, AI makes an expert-led study dramatically faster and more thorough without lowering the standard. The key is dividing the work correctly.

AI should accelerate:

• Intake — gathering and organizing information from engineering systems, payroll, and project records, and asking smart follow-up questions instead of a static form.
• Drafting — producing first-draft technical narratives and business-component descriptions from real project data for an expert to review and refine.
• Computation — handling the mechanical calculation of qualified expenses and the credit, consistently and quickly.

Humans should decide:

• Qualification — whether an activity actually meets the four-part test.
• Amounts — which expenses are qualified and how much to claim.
• Final conclusions — what position to take, and standing behind it.

In short, AI does the heavy lifting on speed and scale; experienced R&D experts make the judgment calls, finalize the study, and own the result. That division is what lets a study be both efficient and defensible rather than trading one for the other. For more on why we built the platform this way, see why Ricerca.

The documentation standard examiners expect

A defensible study is ultimately a documentation exercise, and the standard is specific:

• The four-part test, applied per business component. For each component, the study should show the permitted purpose, the technological nature, the uncertainty that existed at the outset, and the process of experimentation used to resolve it. (New to the credit? Start with our R&D tax credit guide.)
• Contemporaneous records. Examiners give the most weight to evidence created while the work was happening — tickets, design documents, test results, lab notes, engineering change orders — rather than narratives reconstructed after the fact. Strong studies tie conclusions back to these underlying records.
• Form 6765, Section G. The revised Form 6765 (updated in 2024) added Section G, calling for business-component-level detail about the research. This formalizes, on the return itself, the expectation that a claim can be broken down and substantiated component by component. A study that cannot populate that detail credibly is a study with a problem.

Automation-only tools tend to struggle here. Producing plausible-sounding narrative text is easy; producing component-level documentation that is accurate, grounded in your actual records, and consistent with the law is the part that still needs experienced people.

The takeaway

The R&D credit is worth claiming, and it is worth claiming well. The fast-but-shallow and thorough-but-expensive extremes each give up something important. The model that holds up is the one that uses AI for what software is good at — speed, organization, first drafts, and computation — while experienced R&D experts make the judgment calls and finalize the study before it is issued. That combination is how you get a credit claim that is both efficient to produce and built to withstand an examination. If you would like to talk through your situation, get in touch.