An R&D tax credit study you can defend, every dollar tied to evidence written at the time
A research credit you can defend, every dollar tied to evidence written at the time.
See one case, screen by screen ↓The work product that supports a research credit claim: each project tested against the four-part test, the qualified research expenses behind it, the exclusions and the computation. R&D Tax Credit Study builds it from project, time and payroll records — seven agents quote evidence written at the time, tie wages to payroll, argue the IRS's side and compute the credit, and the tax director approves and signs.
How to choose AI solutions for back-office and enterprise work →
Updated
The research credit is won or lost on the four-part test, project by project — a permitted purpose, technological in nature, uncertainty about capability, method or design, and a process of experimentation for substantially all of the work — and on proving whose wages went into each project. Many engineering teams keep no project time records, so wages are estimated, and the study is usually written six to eighteen months after the work, from interviews with engineers reconstructing what they did.
That reconstruction is what fails in an exam. Courts have denied credits where the experimentation was not recorded at the time, and exclusions — research after commercial production, funded research, routine testing, internal-use software below the high threshold — are where claims come apart. Since 2025 the same records also drive the Section 174A deduction, and Form 6765 Section G asks for the numbers per business component.
Estimated split for one year's study across two R&D centres, by hand and with the solution.
Seven agents find the projects, quote the evidence, ask the engineers, split the wages, argue the IRS's side and compute the credit; the tax director signs.
Kestrel & Vane Group's US group files a consolidated return. Its FY2026 study (1 Jul 2025 – 30 Jun 2026) covers two R&D centres, Energy Systems in Austin and Distribution Technology in Dayton, with the return due 15 Apr 2027 on extension. The agents finished their run on Monday 12 October. It is Thursday 15 October. Here is what happens next, screen by screen, in the working solution.
Nadia's home screen opens on four numbers: a research credit of $411,936 ready for her approval, $4,862,400 of qualified expenses in 11 projects, $1,186,300 left out in 6 weak projects, and $21.0M of §174A deductions. The agents read 61,240 tickets, 38,912 commits and a year of payroll. Six things need her; the study sign-off is the main one, and beside it the agents report in their own words.
Interview agent: “39 of 46 engineers answered, in a median of 1 min 52 s. Two are past the 5-day clock; I told their managers.”
Nadia asks the agents in plain words. The Calculation agent answers with sources: decide the pump seal challenge first — the only project still disputed, $121,500 — then choose the §174A catch-up, then approve the study. The interviews still open do not change any project's verdict. Under the answer, one button opens the pump seal project.
Every answer carries numbered citations that open the passage it rests on.
Each of the 17 projects is a row. The four columns are the four tests, each cell a quote from Jira, Git or a design document with the date it was written; on the right, the expenses and a kept or excluded verdict. 1,846 quotes, 97% written at the time. Filters show what is kept, excluded, challenged or waiting on an engineer.
Skeptic agent: “I scored every project the way an examiner would read it. Six fail a test and are left out; two are kept but challenged, and their cells are marked weak.”
Hydra battery thermal model, owned by Mateo Ruiz, is the largest project at $1,920,400. Its experimentation cell opens the repository kv-hydra-thermal: tags v0.1 to v0.14, from baseline air cooling at ΔT 6.8 °C in September 2025, through “iter 7: channel 4 mm → ΔT 4.1 °C, fail”, to “iter 14: graded channels + pads → ΔT 2.6 °C, pass; pump 0.94 kW” on 20 Feb 2026. 412 commits by 9 authors sit between the tags.
Evidence agent limit: it never paraphrases a quote, and anything written after the year ended is marked “written later” and cannot carry a test alone.
Instead of an hour-long interview months later, the Interview agent sent Mateo three questions in Teams, each answer drafted from his own records: Hydra 75%, field installation support 20%, other work 5%, from Workday time records and 412 of his commits; the unknown he was chasing, quoted from Jira HYD-212; and the 14 designs he tried. He confirmed all three in 1 min 48 s and added that iterations 10–12 were run with the thermal lab.
The same three questions go to everyone, from a set Owen Castell approved on 2 Oct. Five working days to answer, a reminder at day 3, and the manager is told at day 8.
Pump seal lifetime test, $121,500, scores 73 and is challenged. Its uncertainty is recorded — “Unknown if silicon-carbide faces survive repeated dry starts” — but only two seal variants were run and the test plan says “confirm the supplier rating”. The Skeptic proposes exclusion as routine testing unless the team shows the dry-start work went beyond the supplier's test. Nadia can keep it with changes, accept the exclusion or ask the engineer again.
“The Skeptic proposes; only a person keeps or excludes.”
Billing UI refresh is out for want of uncertainty: its own design document says “Reuse the existing design system components; no new technical risk expected” — $214,600. The Finance ERP release upgrade fails the internal-use software test, Lumen sensor co-development is funded by the customer, and Hydra's work after the first shipment on 21 Mar 2026 is split out as field installation support. Owen agreed all six on Mon 12 Oct; reinstating one needs a reason and goes to the VP Tax.
Skeptic agent: “Leaving these out avoids a 20% penalty on the tax they would have saved.”
The Defence agent asked itself 14 information requests an examiner typically sends and drafted each answer from the evidence. Twelve are strong. Two are weak, with $420,400 of expenses behind them: the route optimiser's claim that nothing on the market fitted rests on an interview given after the year, and the pump seal testing may look routine. The evidence binder holds the 1,846 quoted passages, 17 project files and the 39 interview records.
Defence agent limit: it answers only from cited evidence; weak answers are flagged, never padded.
The alternative simplified credit, $411,936, beats the regular credit of $284,610, and the base years tie to the filed returns. The same records give §174A: $9,428,064 of domestic research costs deducted, the Brno team's foreign costs capitalised over 15 years, and $11,620,000 of 2022–2024 costs to catch up over one year or two. One year gives $2,440,200 of tax effect now; FY2026's effect comes to $4,832,029.
Nadia proposes the catch-up; it goes to Graham Ellery, VP Tax, to co-approve.
Nadia chooses A1 — “Approve the study as prepared: projects, evidence, wages and the credit are right” — adds a note for the file if she wants, and ticks the e-signature: “I approve the FY2026 research credit study as Nadia Brooks, Tax Director.” Until she ticks it, Approve and sign stays locked. The study is sealed, and the evidence binder with it.
The agents file nothing. Credits or elections over $1M also need the VP Tax.
The same study, seen by five of the people behind it — what the year-end scramble looked like, and what it looks like now.
Seven agents find the projects, quote the evidence, ask the engineers, split the wages, argue the IRS's side and compute the credit; the tax director signs.
Finds candidate projects — business components — in Jira epics, repositories and design documents, groups the work and names the people. On the sample year it found 17 projects in 2,418 epics.
Quotes the evidence for each of the four tests with the date it was written, and builds the iteration timeline from the commits.
Asks engineers three one-click questions in Teams or e-mail, pre-filled from their own tickets and commits, reads the answers and reminds the ones who have not answered.
Splits each person's wages across projects from time records, commits and answers, applies the 80% rule, ties every dollar to box 1, and reads supplies, contract research and cloud costs from the ledger and contracts.
Argues the IRS's side: tests each project against the audit guide, the regulations and case law, scores how well it would stand up, and challenges it or proposes to leave it out.
Computes the alternative simplified and regular credits, the §280C choice, the §174A deduction and catch-up, and Form 6765 with Section G.
Writes the study report from the evidence, rehearses the IRS's information requests and builds the evidence binder.
Approves the study; the VP Tax co-signs elections. The agents propose; a named person decides.
Every screen has the agent that did the work beside it. Ask about a project, a test, a person's wages or the credit — the answer cites its source, and it can open or filter what you need. Nothing changes until a person decides.
What should I do first?
Decide the pump seal challenge first: it is the only project still disputed, $121,500. Then choose the §174A catch-up, and approve the study. The interviews still open do not change any project's verdict.
Why is Billing UI refresh excluded?
Its design document says “reuse the existing design system components; no new technical risk expected”, and none of its tickets records an unknown. Without uncertainty it fails the test; $214,600 is left out. Owen Castell agreed (X1).
Why is Mateo Ruiz at 75%, not 100%?
His own answer splits his year: Hydra 75%, Hydra field installation support 20%, other work 5%. Field support came after the first shipment on 21 Mar 2026, so it is left out, and below 80% the rule does not apply.
What if we drop the pump seal project?
Expenses fall by $121,500 and the credit by $17,010, to $394,926. That decision is yours on the project screen; I change nothing until you record it.
13 screens from the working solution, on its sample data. Pick one to see it large.
The rules every project is tested against are settings with their sources beside them. Changes are drafts until the VP Tax approves them, and the Skeptic agent will tell you what a change would do to this year's study before you send it.
| Setting | Default | Choose from |
|---|---|---|
| Four-part test | All four tests need a quote written in the tax year or within 90 days after | IRC §41(d) · Treas. Reg. §1.41-4(a) |
| Written later | Interviews and notes after that date support a test but cannot carry it | IRS Audit Techniques Guide |
| Internal-use software | High threshold of innovation: innovative, significant risk, not commercially available | Treas. Reg. §1.41-4(c)(6) |
| Commercial production | Research ends at the first customer delivery in SAP | IRC §41(d)(4)(A) |
| Contract research | 65% of amounts paid; we must bear the risk and keep the rights | IRC §41(b)(3)(A) |
| Confidence below which an agent asks a person | 90% | Threshold · a draft until the VP Tax approves |
| Skeptic score that opens a challenge | 75 | Threshold · a draft until the VP Tax approves |
| Interview clock | 5 working days; reminder at day 3; manager at day 8 | Threshold · a draft until the VP Tax approves |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published rules (Treas. Reg. §1.41-4(a) and Little Sandy Coal Co. v. Commissioner, 62 F.4th 287 (7th Cir. 2023) on “substantially all”; IRC §6662 on the 20% accuracy-related penalty; Form 6765 instructions, rev. Dec 2025, on Section G, optional for tax years beginning before 2026). Other context: IRC §41(c)(4), (d); Section 174A (OBBBA, 4 Jul 2025); Siemer Milling Co. v. Commissioner, T.C. Memo. 2019-37. People, companies, projects and figures named on this page are fictional — characters and sample data in the working solution.
The work product that supports a research credit claim: each project tested against the four-part test, the qualified research expenses behind it, the exclusions and the computation. R&D Tax Credit Study builds it from project, time and payroll records — seven agents quote evidence written at the time, tie wages to payroll, argue the IRS's side and compute the credit, and the tax director approves and signs.
The Evidence agent quotes a passage for each test — permitted purpose, technological in nature, uncertainty, experimentation — from Jira, Git or a design document, with its author and the date it was written. Each cell opens to its source. Notes and interviews written after the year can support a test but cannot carry it alone.
No. The Interview agent sends three questions in Teams or e-mail with the answers already drafted from each engineer's tickets and commits; they confirm or correct them. In the working solution, 39 of 46 engineers answered in a median of 1 min 52 s, and a correction flows straight into the wage allocation.
The Allocation agent splits each person's wages across projects from time records, commits and interview answers, applies the 80% rule, and ties every dollar to box 1 in the payroll register. Any difference over $1,000 goes to the Payroll Lead to explain.
No. The Skeptic agent reads each project against the audit guide, the regulations and case law, scores it and proposes exclusions with the rule and the quote — research after commercial production, funded research, routine testing, internal-use software below the high threshold. A person agrees or reinstates, with a reason.
Yes. The Calculation agent models the domestic deduction, foreign amortisation over 15 years, the one-year or two-year catch-up for 2022–2024 costs and the §280C choice beside the credit, and fills Form 6765 with Section G per business component.
Yes. The agents read every system of record read-only and file nothing. A person decides every verdict, exclusion and election with a reason code, the tax director approves the study with an e-signature, and credits or elections over $1M also need the VP Tax. Every step is on a sealed audit trail.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your study method, prior-year studies and returns, and the systems your engineers and payroll use — and it goes live once every quality gate has passed. We will walk you through it on your own projects first.
Every country tested, every field traced, the return ready months before the deadline.
1-min demoE-Invoicing ComplianceEvery invoice checked before it leaves, and every rejection fixed the same day.
1-min demoTransfer Pricing DocumentationEvery margin tested before the year closes, and every file ready inside 30 days.
1-min demoWe'll run R&D Tax Credit Study on a sample of your own projects, tickets, commits and payroll.