AI for accounting firms
AI for accounting firms
Return prep, month-end close, and client document chase without the busy-season grind. 5x ROI in 30 days, or we work free.
- Bookkeeping practices
- Tax practices
- Audit and assurance
- Fractional CFO shops
- Full-service CPA firms
Teams we build for
- Hoyes Michalos
- Nurse Next Door
- Fedi
- UBC Sauder
- Merchant House Capital
- Picton Investments
- Campbell Froh May & Rice LLP
- Barnakl
- Hungerford
- Breez
Teams we build for
- Hoyes Michalos
- Nurse Next Door
- Fedi
- UBC Sauder
- Merchant House Capital
- Picton Investments
- Campbell Froh May & Rice LLP
- Barnakl
- Hungerford
- Breez
Teams we build for
- Hoyes Michalos
- Nurse Next Door
- Fedi
- UBC Sauder
- Merchant House Capital
- Picton Investments
- Campbell Froh May & Rice LLP
- Barnakl
- Hungerford
- Breez
Proof, honestly
We have not published an accounting case study yet, and we say so. Founding clients get priority scheduling, direct founder involvement, and a co-published case study once the numbers are real. The 5x ROI guarantee applies from day one.
We guarantee 5x ROI inside 30 days of deployment, in writing, measured against a baseline you sign before we build. If the system misses the bar, we keep working for free until it clears.
Busy season runs on overtime.
The same returns, the same checklists, the same April. Capacity is capped by hours nobody has, and the ceiling is set by what the team can survive.
Clients send documents late and wrong.
Half the engagement is chasing slips, statements and signatures. None of it is accounting.
The close drags into the next month.
Reconciliation and categorization by hand, every month, for every client. The last clients in the queue get their numbers weeks after the month they describe.
Every February, the firm re-decides last February.
Salary or dividend for this owner-manager. How the firm treats a client's crypto gains. Decided before, recorded in a prior-year working paper, and re-researched from scratch because asking the file is not possible.
The people who survive April quit in May.
The season is staffed with evenings and weekends, and the invoice for those hours arrives as spring resignation letters. New US CPA exam candidates fell from 42,626 in 2023 to 28,082 in 2024 before enrollment turned up 12% (AICPA, 2025, US pipeline). The replacement senior comes out of that pool, and the next April starts shorter-handed.
Advisory is always next year's plan.
Every partner retreat ends on the same slide about moving up to advisory. Then the slips start arriving and everyone goes back to keying box numbers. The constraint was never intent. It is capacity.
Accounting firms, before and after
The manual path is dashed: Slips chased by email, Returns prepped from scratch, Close drags into next month. The system path replaces it, and a person approves before anything ships: Documents pursued automatically, First-pass prep for review, Close runs on schedule.
Before: by hand
- 01Slips chased by emailhuman
- 02Returns prepped from scratchhuman
- 03Close drags into next monthhuman
After: the system
- 01Documents pursued automatically
- 02First-pass prep for review
- 03Your approvalhuman
- 04Close runs on schedule
The research
Accounting firm AI adoption jumped from 9% to 41% in 2025 (CPA.com, 2025).
CPA.com · 2025
The economics
Before and after economics
Line
- Hours in scope, modelled
Before
At the page defaults, 6 people spending 10 hours a week on the chase, categorization and assembly is 60 hours, $5,100 a week at $85 loaded cost
After
The calculator below prices your own baseline, and the 5x ROI guarantee is measured against a baseline you sign before we build
- The document chase
Before
A person emails, waits, and emails again, client by client, slip by slip
After
The list pursues itself until complete, and staff open a file when it is ready to work
- Return prep
Before
Keyed by hand from whatever the client sent, in whatever shape it arrived
After
Extraction and first-pass assembly land before a person opens the file, and a CPA reviews and signs
- The close queue
Before
Each client's close starts when someone frees up, and the last clients wait longest
After
Closes run on schedule against firm rules, and exceptions are the only thing a person touches
No row on this table is a client outcome, because we have not published an accounting case study yet and this page does not pretend otherwise. The single modelled row multiplies this page's calculator defaults, a team of 6 at 10 manual hours a week each at an $85 loaded hourly cost, which is 60 hours or $5,100 a week in scope. Every other row is a qualitative state change. The first founding engagement replaces this table with measured numbers, co-published.
Run your numbers.
Your operations
Savings use the low end of our hours-reclaimed range. The math is conservative on purpose.
The math
Calculated at the low end of every range.
Every first build is covered in writing: 5x ROI in 30 days. Or we work for free.
The last Friday of March, as the tax manager runs it
She starts in the review queue, because it is full.
humanNine returns sit ready for review, each one complete, slips attached, prior-year comparison run. She picks the first and reads. Nobody stayed late last night assembling any of them, which is the point of the whole build.
A shoebox arrives by phone camera and becomes a file.
A client photographs two T5s and an RRSP receipt in a parking lot. Extraction reads issuer, box numbers and amounts into the TaxCycle file for his return, and flags one blurry amount for a person instead of guessing it.
She decides what a machine must not.
humanThe flag on another file is not a bad scan. It is a support payment that is deductible or not depending on what the separation agreement says. She reads the agreement and decides, and the decision is recorded against the file with her reasoning.
The chase asks for the missing slip by name.
The prior-year comparison shows a T3 from the same fund trust last March and none this year. The follow-up email names the fund and the slip. It stops the moment the document lands, and it never asks a client for something the file already holds.
The April board recounts itself.
Partners see returns by state: waiting on documents, ready for prep, in review, out for signature. The two files stuck on one document each say which document. Nobody walks the office asking.
The only path to EFILE runs through her review.
humanShe works the queue through the morning. On each return her sign-off releases the T183 request to the client, and transmission under the firm's EFILE number happens only after the signed form is back. The system prepared everything. It filed nothing.
The decision is filed where next February will look.
The support-payment call and the reasoning behind it land in the working papers, indexed. When the question comes back next season, the preparer gets the firm's answer and the engagement it was decided in, not a blank research session.
Who this is built for
The same headcount clears a bigger April
The season caps the year. Every return the firm cannot absorb in those ten weeks is revenue that walks, and hiring for the peak means carrying the cost through August. When the chase, the keying and the assembly move off people, the ceiling moves without the payroll moving. We publish no capacity number yet because no accounting case exists yet. The founding engagement sets that number, measured against a baseline you sign, with your firm's name on the case if you want it there.
You review returns you did not assemble
The file reaches you complete. Slips extracted, prior-year comparison run, the low-confidence reads flagged for a person instead of guessed. Your day is review notes, judgment calls and the client conversation about the balance owing. The T183 request goes out after your sign-off, never before it.
Client 14 closes the same week as client 1
The monthly queue used to be a race the last clients always lost. Categorization runs against each client's own rules the night the bank feed lands, exceptions queue for a person, and the close checklist advances on its own state. You spend the month on the books that need a bookkeeper, not the ones that need a typist.
Five clients, five ledgers, one briefed Monday
You sell judgment, and the raw material of judgment is current books. When each client's ledger reconciles on schedule and the meeting brief assembles itself from the closed numbers, the hour a client buys is an hour of thinking. The export-and-format hour disappears from your week without ever appearing on an invoice.
What stays human
- She starts in the review queue, because it is full
- She decides what a machine must not
- The only path to EFILE runs through her review
The steps the day below leaves to a person, by design.
What we build for accounting firms
Get the documents in
Client document chase.
PBC lists pursued automatically until complete: slips, statements, signatures. On assurance files the list lands before fieldwork starts and items file into the CaseWare binder.
Engagement letters and onboarding.
New clients gathered, papered and filed without partner chasing. The letter waits for a partner, not the other way around.
Run the season
First-pass return prep.
Slips and statements extracted into the return workflow, prior year compared, low-confidence reads flagged. A CPA reviews and signs. The assembly work disappears.
Bookkeeping categorization with firm rules.
Your chart of accounts and your judgment calls, encoded once, graded against closed engagements, applied consistently.
Month-end close automation.
Transactions categorized and reconciled first-pass against each client's rules. Exceptions queue for a person. The checklist advances on its own state.
Sell the judgment
Advisory prep briefs.
Client financials summarized before every meeting, whether the charts come from Fathom or the ledger itself, so the conversation starts at the insight.
The numbers in accounting firms
Accounting firm AI adoption jumped from 9% to 41% in 2025 (CPA.com, 2025).
CPA.com · 2025
AI-using accountants closed monthly statements 7.5 days faster (Stanford GSB, 2025).
Stanford GSB · 2025
Where the data comes from
Ledger and GL
The client books themselves, QuickBooks Online, Xero or Sage 50, one file per client and dozens per firm. Bank feeds, each chart of accounts and the firm's categorization rules live here. First-pass categorization is proposed into the review lane the ledger already has, so a bookkeeper confirms lines instead of keying them.
Where it stops. Proposals only. Adjusting entries, closing entries and anything that touches a filed period stay with a person, and the ledger's own permissions decide which person.
Practice mgmt
Karbon, TaxDome or CCH iFirm holds the engagements, the deadlines and the April board. It is where the PBC list lives, where a return's state runs from documents-in to out-for-signature, and where partners look when they want the true picture of the season.
Where it stops. The system updates checklist and job state as work completes. Review steps and sign-offs can only be cleared by the named reviewer, so the board shows what a person approved, never what a machine claims.
Client portals
Where the engagement's raw material arrives: slips and statements through the portal, receipts through Dext or Hubdoc, the rest by email attachment and phone photo. Everything is extracted where it lands, filed against the right client and the right year, and the chase asks for what is still missing by name.
Where it stops. Outbound messages to clients are reminders from templates the firm approved. Anything that resembles advice, a fee or a scope conversation routes to a person before it leaves. Client documents stay in the firm's store and never train a model.
How the system is built for accounting firms
See the full capability mapRetrieval
The firm's own treatment decisions, working papers and client history, indexed so last year's answer is findable this year. A preparer asks how the firm handled a position before, and gets the file it was decided in.
- pgvector
- Full-text BM25
- Reciprocal rank fusion
Agents and orchestration
Agents run the close checklist and the document chase on a schedule, categorize against firm rules, and escalate what does not reconcile. A CPA reviews and signs. The agent never files.
- agent-worker
- Scheduled runs
- Proposal queue
Evaluation
Categorization and extraction are graded against engagements the firm has already closed. Accuracy is reported per field, so nobody has to trust a vendor's word for how well it reads a statement.
- Eval graders
- Field-level accuracy
- quality-worker
Models
Document models for slips, statements and receipts. Frontier models for the judgement questions. Rules and classifiers for the high-volume categorization, where deterministic beats clever.
- Document layout extraction
- Frontier models, one gateway, routed per task
- Rules and classifiers
Data boundary
Client data is scoped per engagement at the database layer, so staff see the files they are on and nothing else. Nothing is used to train a model.
- Supabase row-level security
- Per-engagement scoping
- No-training API terms
Ledger and GL, Practice mgmt, Client portals feed a hybrid index. The agent runtime works from that index, and every consequential action passes a human approval before it reaches Workflow builder, Task board, Ask your brain.
Your systems
- 01Ledger and GL
- 02Practice mgmt
- 03Client portals
The system
- 01Hybrid index
- 02Agent runtime
- 03Your approvalhuman
Where your team works
- 01Workflow builder
- 02Task board
- 03Ask your brain
How an answer is found in the firm's working papers
A question runs against the firm's working papers two ways at once: vector search for meaning and full-text search for exact wording. Reciprocal rank fusion merges both result sets, and the answer carries the source it came from.
- 01A question
- 02Vector search (meaning)
- 03Full-text search (exact wording)
- 04Rank fusion
- 05Answer, with its source
Built around your rules
| Regime | What it demands here | How the system complies |
|---|---|---|
| CPA Code of Professional Conduct | Client confidentiality survives every tool choice, and the member answers for all work issued under their name, including work a system prepared. | Every filing-bound output is a draft with a named CPA reviewer, and confidentiality is enforced per engagement at the database layer, not by memo. |
| CRA records and the Income Tax Act | Books, records and workpapers kept and producible for the years the Act requires, with a trail a CRA reviewer can follow. | Workpapers, source documents and treatment decisions file into the firm's own store, attached to the engagement they belong to. The system prepares. Your people file. |
| CRA EFILE and the T183 | A signed T183 in hand before a return is transmitted, and the electronic filer answerable for everything filed under its number. | The system never transmits. A return's signature request is released by the reviewing CPA's sign-off, and filing happens under the firm's EFILE number after the signed form is back. |
| FINTRAC obligations for accountants | Where the firm performs triggering activities, client identification, record keeping and reporting duties follow. | Client-identification records are prepared and gaps are flagged. Whether anything is reported stays with your compliance lead. |
| BC PIPA and PIPEDA | The slips and statements an engagement collects carry SINs, account numbers and home addresses. Both statutes govern how they are gathered, held and retained. | Client documents live in the firm's store with residency scoped per engagement, retention follows the firm's own file schedule, and nothing trains a model. |
The regimes that govern accounting firms, what each demands, and how the system complies
Regime
- CPA Code of Professional Conduct
What it demands here
Client confidentiality survives every tool choice, and the member answers for all work issued under their name, including work a system prepared.
How the system complies
Every filing-bound output is a draft with a named CPA reviewer, and confidentiality is enforced per engagement at the database layer, not by memo.
- CRA records and the Income Tax Act
What it demands here
Books, records and workpapers kept and producible for the years the Act requires, with a trail a CRA reviewer can follow.
How the system complies
Workpapers, source documents and treatment decisions file into the firm's own store, attached to the engagement they belong to. The system prepares. Your people file.
- CRA EFILE and the T183
What it demands here
A signed T183 in hand before a return is transmitted, and the electronic filer answerable for everything filed under its number.
How the system complies
The system never transmits. A return's signature request is released by the reviewing CPA's sign-off, and filing happens under the firm's EFILE number after the signed form is back.
- FINTRAC obligations for accountants
What it demands here
Where the firm performs triggering activities, client identification, record keeping and reporting duties follow.
How the system complies
Client-identification records are prepared and gaps are flagged. Whether anything is reported stays with your compliance lead.
- BC PIPA and PIPEDA
What it demands here
The slips and statements an engagement collects carry SINs, account numbers and home addresses. Both statutes govern how they are gathered, held and retained.
How the system complies
Client documents live in the firm's store with residency scoped per engagement, retention follows the firm's own file schedule, and nothing trains a model.
Client financial data stays inside firm-controlled accounts under no-training API terms, scoped per engagement with audit logs.
The objections
Faster prep means smaller invoices. We bill this work by the hour.
Most bookkeeping and compilation work already bills flat, and on a flat fee every manual hour is margin, not revenue. Tax runs the other way. An April practice is capacity-limited, the firm turns work away in the exact weeks demand peaks, and a return you decline bills at zero. Recovered hours do not shrink invoices. They change how many invoices April can hold.
Owner-manager tax is judgment. Salary versus dividend, the capital dividend account, when to trigger a gain. Software does not call that.
Correct, and nothing here calls it. The system does the extraction, the assembly and the chase, the work underneath the judgment. What it adds above that line is memory. When the salary-or-dividend question comes back for a client, the working papers where the firm decided it last time surface first, with the engagement they came from. The CPA decides. The deciding starts from the firm's own history.
Our clients barely use the portal we already pay for.
The chase does not ask clients to change. It reads what arrives however it arrives, portal upload, email attachment, a phone photo of a T5, and then asks for what is missing by name. A named ask for a named slip gets answered. A reminder about a checklist gets ignored. The portal stays the firm's filing surface, not a habit your clients must adopt.
We tried receipt-capture tools. The categorization was wrong and the partners stopped trusting the file.
Generic capture tools guess from a general model of what businesses look like. This build encodes your chart of accounts and your firm's own calls, then grades itself against engagements you already closed before it touches a live file. Accuracy is reported per field, and anything under the bar queues for a person instead of posting. The trust problem was never automation. It was automation nobody measured.
My seniors will read this as the firm automating them out.
Show them what moves and what stays. What moves is retyping box numbers, chasing a missing T3 and reconciling a bank feed on a Sunday. What stays is review, client calls and every judgment call in the file, which is the work people stay at firms to do. Firms do not lose seniors to automation. They lose them to April.
How we build it for accounting firms
Codify the firm's rules before automating them.
Categorization rules live in the partners' heads. We write them down and test them against engagements already closed, which is usually where the firm discovers two partners disagree.
Automate the chase, then the close.
Document chase is the highest-volume, lowest-judgement work and the safest place to start. Close automation follows once extraction accuracy is measured per field.
Move up to advisory once capacity is real.
The point of the build is redirecting CPA hours to advisory work. That only counts if the reclaimed hours are measured against the baseline we recorded in step one.
What we will not automate
Sign-off, filings and any position that carries professional liability. A preparer reviews everything, and the system never files.
What that means in practice
Vision, documents and speech
AI for eyes and ears: reading documents, watching camera feeds, transcribing calls.
Where we stop. Bespoke computer vision is justified by volume and latency, not by novelty. Below that bar, a vision-language model on demand is cheaper to run and easier to maintain, and we will tell you which side of the bar you are on before anything is built.
Retrieval and RAG
Connecting AI to your actual documents so it answers from your knowledge, accurately and with citations, instead of making things up.
Where we stop. A dedicated vector database is justified by scale, not by default. Most of RAG quality is won or lost in chunking and indexing strategy, not in the model choice, and we have walked clients back from RAG to plain search when that was the honest answer.
Evaluation and observability
How we prove the AI actually works: measured, monitored and regression-tested like real software, not vibes.
Where we stop. There is no engagement where we skip this. The honest variable is depth: a document pipeline gets faithfulness and extraction suites, an outbound agent gets human review sampling, a classifier gets a held-out test set. We size the harness to the risk, never to zero.
Where your team works
Tour the platformWorkflow builder
Describe the workflow. Watch it assemble.
Say what should happen in plain language and the builder assembles the automation on a canvas you can read, run, and change.
Task board
People and agents, working the same board.
Every piece of work lives on one shared board, whether a person or an agent owns it. Handoffs between the two are explicit, so nothing falls in the gap.
Ask your brain
Ask a question. Get the answer and where it came from.
Every answer cites the document behind it, so you can check the source yourself. When the brain does not have the answer, it says so instead of guessing.
Approval queue
The AI proposes. You approve. Nothing sends itself.
Every consequential action arrives as a proposal with the risk, reversibility, and expiry spelled out before you say yes. Control stays in the room.
By team
The same system, seen from the desk that runs it.
The hard questions
Free · 3-5 days
Know your number in five days.
We map your operations, find the highest-ROI automations, and hand you a ranked plan with the payback math attached. Yours to keep, whoever builds it.
Prefer to talk first? Book 15 minutes with James. No pitch deck.