The problem
A small or mid-sized CPA firm charges roughly $3,000 per SMB client for year-end books and a corporate return. Behind that invoice sits 80 to 100 hours of work on shoebox PDFs, bank statements and journal cleanup, most of it done by a junior who is learning on the client's dime. At that ratio the firm barely breaks even, so the smallest clients are the least loved and the most error-prone.
QuickBooks and Xero were built for an in-house finance team that keeps books all year. They were not built for the accountant who receives a year of documents in March and has to produce GIFI-mapped journals and a T2-ready file by June. I had sat in those discovery calls. The workflow broke in the same three places every time: intake, categorisation, and the review that nobody had time to do properly.
The bet
JournalRelay is the CPA's virtual junior accountant. It ingests raw documents, produces clean books with GIFI mapping from day one, and leaves the accountant with a review queue instead of a data-entry job. The CPA still reviews, signs and owns the client relationship. We never tried to replace the signature. We tried to make the signature cheap.
The model works and the practitioner will sign off on it are two different bars, and closing the gap between them is mostly a product problem.
That framing drove every decision below. A 95% correct set of books is worth less than nothing to an accountant, because checking it costs more than redoing it. So the product is the checking.
Rules accountants trust
The matching and categorisation engine runs on a strict order of precedence. I wrote these rules with three design-partner firms, not for them, and they have barely changed since.
Balancing and control-account rules are enforced on the server, so neither a person nor a model can post an unbalanced entry. Posted entries are immutable. Corrections are new entries with a reference to what they correct, which is how an auditor expects a ledger to behave.
Letting agents in through the front door
Once the ledger was stable the obvious next step was to let AI agents do the routine queue. The tempting way to do that is a privileged back door. I shipped an MCP server instead, so that an agent works inside JournalRelay exactly like a member of staff: the same permissions, the same balancing rules enforced on the server, the same audit trail with the agent named as the actor.
The practical consequence is a clean division of labour. Agents take the routine queue. People take what the agent is unsure about. A firm can see, for any entry, whether a person or an agent posted it, on what evidence, and with what confidence. That visibility is what let partners accept the uplift in price.
Results
Pricing moved from $299 per file at proof of concept to partners accepting $1,500 to $2,000 per file with errors-and-omissions cover built in. The product graduated from AMII, the Alberta Machine Intelligence Institute, and is funded by Alberta Innovates micro-vouchers and SR&ED. We were selected for the fall cohort of the Alberta Revenue Accelerator and are in an enterprise conversation with a mid-sized Canadian firm.
Each workflow has a golden test set built from real historical client books rather than synthetic cases, so a change to a prompt or a rule can be checked against human-labelled ground truth before it reaches a firm.
What I would do differently
- Instrument override rate from day one. We added it after the first design partner. It is the single best signal of whether the product is trusted, and I would now ship it before the first proposal is ever shown.
- Price on outcome earlier. The $299 proof-of-concept price anchored the first conversations too low. Firms were willing to pay for a reviewed, insured file, and the earlier we said so the better the partners we attracted.
- Treat the review UX as the product. The first version spent its effort on extraction accuracy. The version that firms adopted spent it on how fast a senior could clear a queue.