Now running free back-tests with UK R&D tax advisers
Every statement in the claim, traced to the record behind it.
ClaimEvidence reads a client's engineering records, such as ticket history, specifications and design documents, and builds a project-by-project evidence pack. Each statement about the baseline, the uncertainty and the work done carries a dated, verbatim source. Where the record is silent, you get a question for the engineer, not invented text.
Built for advisers who sign the claim. You stay the author; we do the digging.
Technological uncertainty
The team could not establish whether ledger reconciliation could stay within one second at peak volume using the existing event store[1], and the first two approaches failed load testing[2].
No known way to keep reconciliation under 1s once we pass 4k tx/s on the current store.
Sharded read model also misses target, p99 2.7s. Abandoning, trying CRDT merge next.
The claim is quick to write. The evidence behind it is not.
Most of an adviser's time on a software claim goes into finding and translating evidence, not applying expertise.
- Engineers who won't reply
Technical interviews get postponed, run short, and depend on people remembering work they finished a year or more ago.
- Records nobody wrote for tax
The evidence exists, scattered across tickets, comments and specifications, but it reads "fix sync bug", not "uncertainty about whether the approach could work".
- Narratives rebuilt after the fact
Tribunals keep giving more weight to records made at the time than to a well-written retrospective report. A narrative that can't point to its record is a weak one.
- The Additional Information Form up front
Every claim now needs project-level detail on the baseline, the advance and the uncertainties before it is accepted, so thin evidence shows up earlier.
- Enquiries that never end cheaply
HMRC checks a large share of claims. Defending one means rebuilding the evidence trail under deadline, and some genuine claimants give up rather than pay to defend.
Why "AI for R&D claims" has a bad name
Advisers have good reasons to be wary. Tools that write the technical narrative produce fluent, generic text, and fill gaps with plausible experiments nobody ran. When an inspector asks where a sentence came from, a generated paragraph has no answer.
The profession has noticed. In January 2026 a tax tribunal described an accountant's unverified AI-generated case summaries as professional incompetence. The same month, the seven PCRT bodies issued guidance that advisers remain responsible for AI-assisted work and must be able to explain how a conclusion was reached. Some firms now advise clients to keep AI out of claims altogether.
ClaimEvidence works the other way round
It doesn't write your narrative. It finds, quotes and organises the evidence, so the narrative you write can stand on it. Three rules are built into the design:
- Quotes are checked mechanicallyEvery quote is matched against the source text before it reaches you. A quote that doesn't match is removed, not reworded.
- Gaps stay gapsWhere the record doesn't support a statement, the pack says "insufficient evidence" and asks the engineer a specific question.
- You can show your workingEvery statement links to the ticket, comment or document it rests on, with its date. That is the explanation PCRT asks you to be able to give.
HMRC is checking more, with better data
Scrutiny of R&D relief isn't easing. The question an enquiry asks is no longer whether the narrative reads well, but whether each statement can be traced to evidence from the time.
HMRC has said it now checks over 20% of R&D claims, and has more than doubled its R&D compliance staff in three years.
Tax Adviser magazine; GOV.UK
HMRC's ten-year contract with Quantexa, signed May 2026, to connect its data and help identify tax at risk.
Bloomberg Tax; UKTN
Every claim needs an Additional Information Form with project-level detail. Without it, the claim is invalid.
GOV.UK
What you get for each claim
An editable evidence pack, organised the way you write the technical report and the Additional Information Form.
Candidate projects
Related work is grouped into projects across the claim period. Every ticket is kept: repeated attempts at the same problem are evidence of experimentation, not noise.
Baseline and uncertainty
What the team knew at the start and what they couldn't work out, each statement quoted from the record with its date.
What was tried, and what happened
The sequence of approaches, failures and changes of direction, in date order, so the systematic work is visible.
Routine work, set aside
Integration, configuration, interface work and ordinary bug fixes are listed as excluded, with a reason, so the claim isn't inflated.
Questions for the engineers
Instead of a long interview, a short list of targeted questions where the record is incomplete. Their answers go back into the pack as new evidence.
A source index
Every ticket, comment and document cited, with dates, ready to hand over if HMRC asks for the records behind the claim.
Against the ways claims are evidenced today
| Engineer interviews | AI narrative writers | ClaimEvidence | |
|---|---|---|---|
| Where the content comes from | Memory, often a year later | The model, prompted with notes | The client's dated project records |
| Each statement traceable to a record | Rarely | No | Yes, with date and quote |
| What happens to gaps | Follow-up calls | Often filled in plausibly | Flagged and turned into questions |
| Routine work separated | Depends on the adviser | Rarely shown | Listed, with reasons |
| Engineer time | Interviews per project | Low, but output unverified | Targeted questions only |
| Who writes and signs the claim | You | Unclear | You |
This compares approaches, not named products. Many advisers combine interviews with their own review of records; ClaimEvidence is designed to replace the digging, not the review.
Built on an engine that already runs in production
ClaimEvidence is built by Sprint Ready Ltd on the same engine as SprintReady, which turns specifications into traced development tickets for regulated software teams. Document reading, source-locked generation that refuses to invent, traceability records, EU hosting and tenant isolation all run in production there today.
The evidence-pack layer is new, and we're building it with advisers rather than for them. That's why we're starting with back-tests on claims you've already filed: you can judge the output against something you know.
What it isn't
- Not tax adviceIt doesn't decide whether a project qualifies. That judgement is yours.
- Not a claim filerIt doesn't complete or submit the AIF or CT600, or contact HMRC.
- Not a cost calculatorIt doesn't apportion staff time or qualifying expenditure.
- Not a narrative generatorIt organises evidence for the narrative you write.
Pricing
Back-tests are free during the pilot. Pricing for live claims will be set with our first adviser partners, based on the time a pack actually saves per claim, and published on this site in full, as it is for SprintReady.
We won't price on a percentage of your client's relief. You set your fees; we're a tool you use, not a competitor for your client.
What a back-test involves
- You choose a project from a claim you've already submitted, ideally one HMRC asked questions about.
- The client exports the ticket history for that period and any design documents, anonymised if you prefer.
- We return an evidence pack within an agreed time.
- You compare it with what you filed and tell us what was useful, wrong or missing.
Questions advisers ask
Will it make things up?
It's designed not to, and it checks itself. The model can only use the records you provide. Every quote it relies on is then matched mechanically against the source text; anything that doesn't match is removed. Where the evidence isn't there, the pack says so and asks the engineer, rather than filling the gap.
No AI system is perfect, which is why every pack is a draft for your review and every statement shows its source, so checking it is quick.
Does it decide whether a project qualifies?
No. It organises the evidence around the questions you need to answer, such as the baseline, the uncertainty and how it was resolved, and it flags work that looks routine with a reason. Whether a project meets the definition of R&D is your professional judgement.
How does this fit with PCRT and my professional obligations?
PCRT's guidance on AI says advisers remain responsible for their work, should keep appropriate oversight, and should be able to explain how a conclusion was reached. ClaimEvidence is built for exactly that: you review a draft, every statement is traceable to a record, and nothing goes to HMRC unless you put it there. You should still decide, under your own firm's policy, what to tell clients about the tools you use.
What records does it need?
For software claims, an export of the client's ticket history for the claim period from Jira or Azure DevOps, plus any specifications or design documents. Word and Excel are fully supported; tell us in advance if most of the material is PDF. It doesn't need access to source code or live access to the client's systems.
Where is client data processed, and is it used to train AI?
Processing runs on Amazon Web Services in the EU (Stockholm) region, using Anthropic's Claude model through Amazon Bedrock. Under AWS's and Anthropic's published terms, content sent this way isn't used to train their models. Each client's data is kept separate. Full detail, with links to those policies, is on the Trust & data page.
Does HMRC's use of AI change what my clients need?
It raises the value of claims that hold up when looked at closely. HMRC's data programme is about connecting information and identifying risk; it doesn't change the legal tests. Contemporaneous, specific evidence is what has always settled enquiries, and it is what ClaimEvidence is built to surface. More on what HMRC has and hasn't said.
Who is behind ClaimEvidence?
Sprint Ready Ltd, a UK company registered in England and Wales (no. 17375471). ClaimEvidence uses the same production engine as our SprintReady product for regulated software teams.
Test it on a claim you've already filed
Pick one project, ideally one HMRC questioned. We'll build the evidence pack, and you judge it against what you submitted. There's no charge and no commitment.