How a Canadian R&D tax credit firm replaced the human expert at the front of every SR&ED claim
A Canadian R&D tax credit firm wanted to replace the human expert on the front end of every SR&ED claim. We built multiple AI agents that ran the technical interviews live, pulled in supporting documents from prior tax returns, and handled the structured intake. A process capped by billable experts became scalable.
The bottleneck this firm lived with
A SR&ED claim, in this firm's process, started with a human expert. That expert had to interview the company's technical leads, identify the R&D projects that qualified for the credit, and write the supporting narrative the CRA would actually accept.
It was expensive work. Slow work. And the only way to scale it was to hire more experts, which the market for expert SR&ED talent in Canada makes hard.
The firm's question wasn't whether AI could help. It was whether AI could replace the first hour of the expert's time on every claim.
The architecture
We didn't build one giant agent. We built three smaller ones, each scoped to one workflow:
- Eligibility interview. Structured conversation with the company's technical lead, drawing out which projects might qualify and which won't.
- Project narrative generation. Drafts the supporting write-up from the interview, in the shape the CRA expects.
- Lead qualification. A lighter version of the eligibility flow for the long tail of "is this even a SR&ED-eligible project?"
All three sat on top of a Qdrant vector store containing the firm's prior tax returns and supporting docs. That retrieval layer was the part that mattered. Without grounding, the agents drifted into generic SR&ED language. With it, the questions and the narratives anchored to the company's actual R&D history.
Delivery was a web-based app the firm's internal team operates without us.
The thing we missed at scoping
We scoped the voice and agent layer. We didn't scope the existing backend the agents had to plug into.
That backend needed a redesign of its own. We didn't catch it until we were deep into the build. The agents worked. Wiring them into the broader system took longer than it should have because the system underneath wasn't where it needed to be.
Lesson: on any build that sits on top of someone else's stack, scope a Phase 0 discovery on the existing infrastructure before committing to the voice work. Standard practice for us now.
Where it landed
Architecture, prompt design, and working agent flows shipped to the firm. They moved into deployment with their internal team after our engagement closed. Outcome metrics belong to them; we don't claim them.
Where this case applies
Professional services firms where the expert interview is the bottleneck. Tax, legal, accounting, claims work, technical due diligence, anywhere the gating step is "an expert spends an hour with the client". That hour is the thing voice AI is good at replacing, or at least pre-staging so the expert spends less of it.
The interview is the work. If you can structure it, you can scale it.