R&D Tax Credits and AI: What CPAs Can Delegate and What They Can't

A client calls. Their software team spent the last year building a new module — they want to know if it qualifies for the R&D tax credit. You've been working this file type for years. You also have three AI platforms running in your firm. Where exactly does the handoff happen?

This is the question that's not getting answered clearly in most conversations about AI and tax practice. It's not "should we use AI?" — at this point, most CPA firms are already experimenting. The question is: which specific tasks in R&D credit work can move to AI, and which ones can't move at all?

A detailed framework appeared in Accounting Today on June 11, 2026, authored by Richard Bernstein, and it draws a precise line. Understanding that line isn't just operationally useful — it's a Circular 230 issue.

What AI Platforms Can Do on R&D Credits

The task list where AI adds real value is specific and substantial.

Modern AI platforms designed for accounting work — including tools that integrate with payroll systems, project management software, and engineering documentation — can handle:

Documentation collection and structure. AI can gather records across payroll data, time tracking systems, engineering logs, and project management tools, then organize them into the evidence file an R&D credit claim requires. What used to take a staff accountant a week can now take hours.

Qualifying business component identification. From payroll and engineering data, AI can flag activities that map to qualifying business components under the four-part test — identifying candidates for review rather than requiring a CPA to read through every line item manually.

Duplicate entry flagging. One of the most common errors in R&D credit preparation is double-counting across departments or time periods. AI catches these automatically.

Initial Form 6765 preparation. AI platforms can draft the form using structured data from the documentation phase, producing a starting point for CPA review rather than a blank form.

This is meaningful work. For a 10-person firm handling R&D credits for technology or manufacturing clients, moving these four tasks to AI changes the economics of the engagement significantly. Documentation that consumed 30–40 hours of staff time can compress substantially.

What Cannot Be Delegated

Former IRS Commissioner Charles Rettig put it plainly: "Technology cannot entirely replace the human effort in getting to the appropriate result. Tax administration will forever remain a team approach — people and technology."

The "people" side of that equation is non-negotiable in three areas:

The four-part qualified research test. Whether a project constitutes qualified research under IRC §41 requires judgment that AI platforms cannot reliably exercise. The four-part test — business component, elimination of uncertainty, process of experimentation, technological in nature — involves factual determinations that vary based on client context, industry, and IRS interpretation trends. Getting this wrong costs clients tens of thousands of dollars and exposes the firm to penalties.

Circular 230 due diligence. This is the point most firms miss. Circular 230 applies to tax practitioners regardless of whether work is produced by the practitioner, a junior associate, an offshore team, or an AI platform. The practitioner's obligation to ensure accuracy and avoid unreasonable positions attaches to the output, not the input method. Using AI to generate documentation does not transfer or reduce that obligation.

IRS defense. If the credit is challenged, the CPA defends it. No AI platform appears at an Appeals conference or responds to IRS correspondence on behalf of the taxpayer.

The Governance Problem

Here's where most firms are exposed right now.

They're using AI tools — legitimately, productively — but without a written governance policy that documents where human review happens, what the CPA's independent analysis looks like before any qualified research determination is made, and how AI-generated work product is logged and retained.

A firm with no governance policy is in the same position as a firm that sends R&D credit work offshore with no review protocol: the work gets done, most of it is probably fine, and then one engagement goes sideways and the CPA discovers that their documentation of independent professional judgment is thin.

Deloitte's 2026 State of AI survey found that only 1 in 5 organizations has a mature AI governance framework. That number is almost certainly lower in small and mid-size CPA firms, where AI adoption is happening at the staff level faster than policy is being written at the firm level.

This matters specifically for R&D credits because the IRS is paying attention. The IRS now employs AI in 129 documented use cases — including audit selection, documentation quality scoring, and claim pattern recognition. Your R&D credit files are being evaluated by systems that weight documentation quality as a signal. A well-documented claim with clear evidence of professional judgment holds up differently than a claim where the paper trail suggests the AI made every determination.

What to Do Before Your Next R&D Engagement

Before your firm's AI platform touches another R&D credit file, put one page in writing:

1. Define the handoff point. Which tasks AI platforms are authorized to perform (documentation collection, component identification, form prep) — and where they stop.

2. Document the CPA review checkpoint. What the practitioner's independent analysis of the four-part qualified research test looks like, in writing, for each engagement. Not "CPA reviewed" — the actual analysis.

3. Identify your audit log. Where AI-generated work product is stored and how long it's retained in case of an IRS inquiry.

4. Apply it consistently. A policy that applies to some clients but not others is harder to defend than a firm-wide standard.

This isn't about slowing down the AI adoption that's already making your firm more efficient. It's about making sure the efficiency gains are defensible when it counts.

The division of labor between AI and the CPA on R&D credits is genuinely useful — AI does the data work, you provide the professional judgment. That only holds up if you can show it.