CARDARELLI GROUP CASE STUDY
The AI That Calls You
How GrowWise Partners turned SR&ED claim preparation into a production system — with an AI agent that interviews clients in more than twenty languages.
Executive summary
GrowWise Partners prepares Canadian SR&ED tax credit claims. Its consultants serve more than 100 Canadian innovators and have filed over $20 million in claims, with 98% audit accuracy and less than five hours of the client's time per claim — against the 20 to 40 hours the traditional process demands.
Behind that service is a purpose-built platform. The platform's AI agent, Sam, conducts a structured technical interview in the client's own language — one of more than twenty — then hands the consultant an English transcript. Sam reads the client's T4 slips and invoices and drafts the CRA technical narrative under the form's own word limits.
The platform ships at unusual speed with unusual discipline: seventeen versions in the last 141 days — 1,797 commits, a release about every nine days — each one converting a manual consultant task into an automated one.
The hard part of SR&ED was never the tax law. It was getting the story out of the engineer's head. Sam listens to clients, in their own language, and gets it.
The client
GrowWise Partners is an SR&ED consulting firm serving companies across Canada — Toronto, Montreal, Vancouver and beyond. Its promise is plain: bigger refunds, a fraction of the client’s time required, senior consultants supercharged by AI, and a contingency fee under 20 percent against an industry standard of 25 to 30.
What an SR&ED claim actually demands
The CRA asks a company to prove three things in narrative form, per project, on form T661 — each under a hard word limit the platform enforces in the product:
| T661 line | The question | Word limit |
|---|---|---|
| 242 | What technological uncertainty could standard practice not resolve? | 350 |
| 244 | What systematic work did you do to resolve it? | 700 |
| 246 | What technological advancement did you achieve? | 350 |
At the same time the company must produce financial detail: every employee's salary apportioned month by month to each project; contractor invoices at 80 percent of arm's-length payments; materials; capital property, newly claimable again for property acquired after December 15, 2024; and government assistance, which reduces the claimable base.
And it must survive review. A CRA reviewer can ask, two years after filing, what evidence supports a given submission.
A claim is three exercises at once — technical writing, cost accounting, and evidence retention — across many clients, each with a filing deadline eighteen months after fiscal year end. This is the workload the platform industrializes.
The insight, and the philosophy
The insight the product is built on: SR&ED preparation is bottlenecked by information extraction, not tax expertise. The consultant's scarce hours go to chasing engineers for interviews, re-typing invoices, reconciling payroll, and translating engineer-speak into the CRA's narrative register. The platform automates the extraction. Judgement stays with the consultant.
The philosophy, written into the platform's own engineering guidelines: the AI is permissive — it casts a wide net and captures anything that might qualify, so no billable project is missed. The human expert is strict — a qualified consultant applies the SR&ED criteria and filters. The AI proposes; the expert disposes. That division is what makes automation safe to point at a tax filing, and it is defensible under CRA scrutiny in a way "the AI decided" is not.
The AI casts the net. Our consultants make the judgement calls. That division is why I can stand behind every claim we file.
Sam's voice: an AI agent that interviews in more than twenty languages
Most tax software sends a questionnaire. Sam is an AI agent: it acts. Sam telephones the client at the scheduled time — or the client clicks and Sam picks up — and conducts a structured technical interview by voice, from project identification through follow-up interviews that target identified gaps. Sam places its own outbound calls, retries on no answer — unanswered calls trigger a follow-up email — and pauses and resumes an interview across days without losing the thread. Agent behaviour, not a chat window.
The language list runs past twenty: English, French, Spanish, Mandarin, Punjabi, Arabic, Tagalog, Hindi, Portuguese, Urdu, Korean, Italian, German, Persian, Vietnamese, Tamil, Ukrainian, Russian, Japanese, Bengali, Romanian. The platform keeps both transcripts: the client is interviewed in their own language, the consultant reviews in English, and the two sit side by side. Interviews pause and resume in any language — across days, across reschedules — with the first leg's transcript intact.
Why it matters: Canada's technology sector has significant diversity and multiculturalism. Sometimes, the person best able to explain a technological uncertainty is the senior engineer whose first language is Mandarin, Punjabi, Persian or Ukrainian. Interview them in English and the narrative gets thinner — or the interview never happens. Sam interviews them in their language. The client-side barrier and the consultant-side workflow are decoupled.
Document intelligence: Sam does the typing
Financial documents stop being hand-keyed. Up to fifty files per batch: the AI agent reads each one and extracts vendor, amounts, tax, dates and currency; a review screen mirrors the manual form field for field with the original invoice alongside; vendors match automatically but never lock; duplicates get flagged; a mis-categorized invoice moves in one click with its file attached.
T4 slips are the hardest target, because the output is per-person money flowing into the credit calculation. The pipeline meets that with multi-tier name matching where fuzzy matches are suggestions only, never auto-assigned. It draws a distinction between the machine's high-confidence suggestion and the user's manual confirmation — data commits only when a human clicks Add to Payroll. Annual income prorates across the months the employee actually worked. And every populated cell carries a provenance badge — manual, payroll-actual, or T4-prorated — with variance flags where sources disagree.
Automated T4 extraction does not capture Social Insurance Numbers: Confidential information is deleted before Sam reads a slip, ensuring information never becomes part of a model’s training. GrowWise operates its solution on the trusted Google Cloud Platform.
The evidence engine: audit defence by construction
A CRA reviewer can ask, two years after filing, what evidence supports a given submission. Today that answer lives in the client’s memory and a folder. The platform is built to make it a queryable property of the claim.
Every project document and completed interview automatically becomes chunked, embedded, searchable evidence — no one triggers anything. The chunking respects structure: headings preserved, tables never split, interviews chunked by conversational turn. One deliberate rule: only the client's answers become evidence — Sam's scripted questions use the T661's own vocabulary, and embedding them would let the system present the interviewer's question as the client's evidence.
The numbers reconcile
The financial engine implements the current federal and provincial rules — the 35 percent enhanced CCPC rate, the expenditure limit now doubled to $6 million for fiscal years starting after December 15, 2024, reinstated capital eligibility, the 1.55 salary proxy, contractors at 80 percent, Ontario's 8 percent OITC and 3.5 percent ORDTC — each rate mapped in writing to its legislative source.
What the client sees is a refund waterfall: how the number was reached, not just the number. What the auditor would see is reconciliation: the labour allocation export agrees to the cent with both the filing file and the accrual report, because all three share one calculation rather than three implementations. And a claim cannot be filed incomplete — the platform blocks it.
Seventeen versions in just the last 141 days
Version 3.4.6 reached production on August 15, 2026 — the seventeenth shipped version in 141 days, on 1,797 commits, a release about every nine days, each carrying user-visible capability.
The pattern repeats: each release converts a manual consultant activity into an automated one. Voice arrived, then gained multilingual recovery and pause-resume. Invoice extraction arrived, then gained approval gating. Narrative generation arrived, then gained word-limit enforcement. Speed, without the usual debt.
Seventeen releases in five months, and the exports still reconcile to the cent. That combination is the whole story.
About GrowWise Partners
GrowWise Partners prepares SR&ED and related tax credit claims for Canadian companies — more than 100 innovators served, over $20 million filed, 98% audit accuracy, under five hours of client time per claim, at a contingency fee under 20 percent. growwise.ai
About Cardarelli Group
Cardarelli Group is a group of companies building artificial intelligence for government, financial services, and law, operating across the United States, Canada, and the United Kingdom. The group owns GovCore, a leading provider of regulatory technology to government — regulatory bodies across the United States and Canada run their licensing, enforcement, and operations on it. cardarelligroup.com