Accounting and tax firms: where AI saves time, and where it creates risk
AI is already at work in professional services, the sector accounting firms belong to. Task by task, here is where it saves time, which technique fits, what must stay in the professional’s hands, and what the CPA Code of ethics, the CRA and Revenu Québec already require.
By Kantara Fofana
In short
- In an accounting or tax firm, AI saves time on work that can be checked: reading documents, drafting requests to clients, summarizing correspondence, classifying transactions, finding a rule.
- It creates risk when it is asked to calculate a tax, reach a conclusion without review, process files in an unapproved tool or cite a reference nobody checks.
- The CPA Code of ethics requires reasonable steps to protect confidential information at every stage (s. 40) and keeping in the file the reasons supporting each document (s. 21).
- The CRA can suspend an electronic filer who fails to protect clients’ information.
- The rule that follows: delegate what can be checked, keep what you attest to.
Contents
- One evening in March, twelve statements and an AI assistant
- Three AI techniques, and rules that are not AI
- A file’s journey, task by task
- Five uses that save time, each with its safeguard
- Four uses that expose the firm
- What professional ethics and the authorities already require
- Before the 2027 season: a plan for a 10- to 30-person firm
One evening in March, twelve statements and an AI assistant
One evening in March, in a 15-person firm, an accounting technician has to break down a self-employed client’s expenses from twelve credit card statements. She uploads them to a consumer AI assistant and asks for totals by category. The answer comes back in seconds, neat and plausible. But a client’s statements have just left the firm for a tool nobody has assessed; the totals are an estimate from a language model, the engine behind these assistants, not a calculation; and nothing in the file explains the figures.
The gain was real, and so was the risk. With an approved tool that reads the statements, flags uncertain lines and leaves the totals to the spreadsheet, she would have saved the same time within a controlled framework.
According to Statistics Canada, in the second quarter of 2026, 32.4% of businesses in professional, scientific and technical services, a sector that includes accounting firms, had used AI to deliver their services in the previous 12 months. In the third quarter, 54.8% planned to do so over the next 12 months, up from 26.3% a year earlier. In March 2026, 65.6% of workers in that sector had used generative AI tools at work over the previous 12 months. Your competitors are equipping themselves and your staff are not waiting: the question is no longer whether AI belongs in the firm, but where.
The rule to remember
Hand AI what can be checked: reading, sorting, drafting, summarizing, finding. Keep what you attest to: the conclusion, the signature, the authorization to file. Calculations belong to the tax software and to formulas.
Three AI techniques, and rules that are not AI
A firm’s work calls on four distinct tools, all lumped together under the word “AI”:
- Rules and formulas: a calculation written in advance, with the same result every time it runs. This is not AI, and it is often what you need: a fee schedule, bank matching, the list of expected documents.
- Optical character recognition (OCR): it turns the image of a document into text and fields, often with a confidence score for each value read.
- Machine learning: a model trained on examples recognizes and classifies a document or an expense. It proposes, and is sometimes wrong.
- The language model, the engine of generative AI: it produces text by predicting the most likely continuation. It drafts, summarizes and reads irregular documents. On its own, it does not calculate: it estimates.
The costliest mistake is AI for everything: asking the language model for what a formula calculates better, and paying each time for a plausible value that has to be rechecked. See our article on reading documents.
A file’s journey, task by task
For each stage: the technique, the gain, the risk and the control. The last column is the one that matters: a gain without control is a risk deferred to the notice of assessment.
| Task | Technique | Gain | Risk | Control |
|---|---|---|---|---|
| Collecting documents | Rules, no AI: expected documents based on last year; a portal | Knowing what is missing without searching | A forgotten document; a social insurance number sent by email | A list per file; uploads through the portal only |
| Sorting and filing documents | Machine learning | Every document in the right file | A T4A taken for a T4 | Review of uncertain classifications |
| Reading and entering amounts | OCR; a language model for irregular documents; a data-entry robot | Reviewing instead of retyping | A misread figure, an amount entered twice | Confidence score, review below the threshold, comparison with CRA and Revenu Québec data |
| Classifying transactions (self-employed client, rental property) | Machine learning and the firm’s rules | A proposal for each line | A personal expense classified as a business expense | Review of uncertain lines and large amounts |
| Bank reconciliations | Matching rules; machine learning for the hard cases | Only the discrepancies are left | A forced match that hides a discrepancy | Any unexplained discrepancy goes to the professional |
| Working papers and calculations | A spreadsheet with checked formulas; a language model for the note | A file note drafted faster | An amount “calculated” by a language model, plausible and wrong | Calculations redone; what comes from AI is flagged in the file |
| Individual and corporate returns (T1, TP-1, T2, CO-17) | The firm’s certified tax software calculates; AI prepares the data | Data entry without retyping | An amount no document supports | Every amount traces to its document; review before the client’s authorization |
| Review | A language model, guided by the firm’s checklist | Inconsistencies spotted before the review | A rubber-stamp review | The reviewer redoes the key points and endorses nothing they do not understand |
| Communicating with the client | A language model to draft and summarize | Clear requests, summarized exchanges | Information sent to the wrong client | Approved tool; the professional reads and sends it personally |
| Billing | Rules, no AI: the firm’s fee schedule | Fees announced in advance | An account the client does not understand | Foreseeable cost announced (s. 62 of the Code of ethics); services provided clearly identified (s. 67) |
Five uses that save time, each with its safeguard
Read documents, with a confidence score
Statements, invoices, receipts: in Microsoft’s Azure Document Intelligence service, a value read often carries a score from 0 to 1; at 0.95, a value is likely correct 19 times out of 20, and the documentation suggests using the score to decide on human review. Set the firm’s threshold, have everything below it reviewed, then compare slips with official data: Auto-fill My Return at the CRA (with the client’s authorization), the client’s download code at Revenu Québec.
Draft requests to the client
The language model drafts the request from the list of what is missing, in the firm’s tone; the professional reads it and sends it personally.
Summarize correspondence
Thirty emails become one page before the call: what was asked, received, promised. The emails themselves stay in the file.
Classify transactions, with review
Machine learning proposes a category for each transaction; the professional reviews uncertain cases and large amounts, and a correction that keeps coming back becomes a firm rule.
Find a tax rule
A language model connected to sources finds the provision and summarizes it; the professional reads the official text, then records in the file the rule applied and why (s. 21 of the Code).
Four uses that expose the firm
1. Having a language model calculate a tax or an amount
A language model predicts text. The Ordre des CPA du Québec’s guide says AI is not designed as a reliable calculation tool unless you have coded the functions yourself; calculations are redone in the spreadsheet or the software. See our article on calculations.
2. Letting AI reach a conclusion without professional review
A tax conclusion commits whoever signs it. The Code of ethics requires CPAs to retain a critical mindset (s. 27) and to consider the extent of their competencies (s. 19). The Ordre sums it up: professional responsibility cannot be delegated to a technological tool. Even the CRA warns that the answers of its generative AI assistant should not be taken as professional tax advice.
3. Putting client files in an unapproved tool
This is the least visible risk: an employee uploads a statement to a consumer assistant. Yet professional secrecy (Professional Code, s. 60.4) and the duty of protection (s. 40 of the Code of ethics) apply to that act. The Ordre classifies the social insurance number (SIN), the Quebec enterprise number (NEQ) and tax data tied to a named person as highly sensitive data, to be kept away from any external AI tool, except in a secure professional environment chosen after a thorough risk analysis. The CRA itself asks users not to give its AI assistant any personal information. See our article on shadow AI.
4. Citing a reference nobody has checked
A language model can confidently present inaccurate information as fact: this is what is called a hallucination. That is why the Ordre’s validation method starts with the source, and the more precise the information, the higher its priority for checking. The Code prohibits signing a document you know, or should know, contains false or misleading information (s. 26).
What the CRA can take away from a firm
The CRA can suspend a filer’s electronic filing (EFILE) privileges, in particular for failing to protect clients’ confidential information, failing to report its loss or disclosure, or failing to ensure that source documents fully support what is reported. In most cases a warning letter comes first, but suspension can be immediate.
What professional ethics and the authorities already require
The Code of ethics of chartered professional accountants (CQLR, c. C-48.1, r. 6.1) replaced the previous code (r. 6) on May 9, 2024: a complete overhaul (professional secrecy moves from s. 48 to s. 39). It does not mention AI, but the Ordre states, in its June 2026 guide, that CPAs’ ethical obligations apply fully to its use:
- Secrecy and protection: a professional may be released from secrecy only with the client’s authorization or where so ordered or expressly authorized by law (Professional Code, s. 60.4; Code of ethics, s. 39), and takes reasonable steps to protect information at every stage, in particular with respect to the persons working with them (s. 40).
- Competence and record: act with due care and keep competencies up to date (ss. 17 and 18), keep in the records the reasons supporting the document (s. 21), supervise anyone for whom you are immediately responsible (s. 22). The Ordre concludes that you should not rely on AI to make up for a lack of competence.
- Responsibility: a CPA cannot exclude, or attempt to exclude, professional liability (s. 52). “The tool got it wrong” relieves no one.
- CRA: the electronic filer deals directly with the client, verifies the client’s identity, ensures documents support every amount, protects confidentiality and reports any loss or unauthorized disclosure. Form T183 (T183CORP for a corporation) is signed before transmission; the firm and the client each keep a copy for at least six years.
- Revenu Québec: Form TP-1000.TE-V or CO-1000.TE authorizes online filing; the client and the firm keep it for six years. The codes for My Account for professional representatives are reserved for their holder, and every transaction made with them has the same legal effect as written, signed instructions from the firm.
Before signing a deliverable prepared with AI, apply the Ordre’s six-point validation method, adapted to a tax file:
- Source: every reference cited exists and says what it is quoted as saying.
- Calculation: every amount has been redone in the tax software or the spreadsheet.
- Assumptions: the taxpayer, the year, the province and the options chosen are the right ones.
- Consistency: the result makes sense compared with last year.
- Trace: the file shows what came from AI, and with which tool.
- Approval: a colleague has reviewed the important deliverables.
Before the 2027 season: a plan for a 10- to 30-person firm
The tax calendar is in our 2027 tax season guide; for AI, here is the plan:
November 2026
Take stock
Ask everyone, without blame, which AI tools they already use, and with what data.
December 2026
Write the rule, choose the tool
One page is enough: what is allowed, with which tool, and what never is; since April 2025 the Ordre has offered a model guideline to adapt. The approved tool has a written contract, does not train its models on your data, says where it processes them and gives individual access. Carry out the privacy impact assessment the law requires (s. 3.3, and s. 17 if data leaves Quebec): see our article on Law 25.
January 2027
Launch two pilot tasks
Reading documents and drafting client requests: two tasks that are easy to control. Set the confidence threshold and the review rule, then test it, in the approved tool, on three of last year’s files with identifying information removed.
February to April 2027
Measure every week
Data-entry and review time per file, the share of values sent for human review, corrections made in review, incidents. Compare with last year.
May 2027
Keep, extend or stop
Keep what saved time without adding corrections, extend it to another task in the table, stop the rest.
What is never delegated
The tax conclusion. The signature and the authorization to file. The tax calculation, which belongs to the tax software. Sending the client a text nobody has read. The access codes for EFILE and for My Account for professional representatives.
Frequently asked questions
- Can a CPA use generative AI to prepare an income tax return?
- Yes, to read documents, draft requests or find a rule, with a tool the firm has approved. The calculation stays in the tax software, every amount is supported by a document, and the CPA remains responsible for the return (s. 52 of the Code of ethics).
- Is a paid AI assistant subscription enough to handle client files?
- No. According to the Ordre des CPA du Québec’s guide, a paid subscription does not mean professional, confidential use: only professional accounts govern the use of data by contract. Before any use, the Ordre asks you to assess the sensitivity of the data and the tool’s privacy settings.
- What does a firm risk if a client’s information is exposed?
- The CRA can suspend its EFILE privileges. And Quebec’s private-sector privacy act requires, after a confidentiality incident, measures to reduce the risk, notice to the Commission d’accès à l’information and to the people concerned when there is a risk of serious injury, and a register.
- Should a firm tell clients it uses AI?
- Several texts point that way. Quebec’s private-sector privacy act requires informing people of the purposes and means of collection (s. 8). The Code of ethics requires CPAs to explain their services to clients and obtain their consent (s. 54). And the Ordre’s guide lists, among unethical situations, presenting a result as entirely your own analysis without transparency about the tool’s role.
Sources
Every fact in this article links to its official source, consulted on October 29, 2026.
- LégisQuébecCode of ethics of chartered professional accountants (CQLR, c. C-48.1, r. 6.1) (opens in a new tab)
- LégisQuébecProfessional Code (CQLR, c. C-26) (opens in a new tab)
- LégisQuébecAct respecting the protection of personal information in the private sector (CQLR, c. P-39.1) (opens in a new tab)
- Ordre des CPA du QuébecCode of ethics of chartered professional accountants: concordance table with the previous code (in French only) (opens in a new tab)
- Ordre des CPA du QuébecKey principles (CPA guide to good practices in artificial intelligence, in French only) (opens in a new tab)
- Ordre des CPA du QuébecEthical obligations (CPA guide to good practices in artificial intelligence, in French only) (opens in a new tab)
- Ordre des CPA du QuébecValidating results: a six-point method (CPA guide to good practices in artificial intelligence, in French only) (opens in a new tab)
- Ordre des CPA du QuébecData protection: classification and […] (CPA guide to good practices in artificial intelligence, in French only) (opens in a new tab)
- Ordre des CPA du QuébecCompliance, professional conduct and ethics (CPA guide to good practices in artificial intelligence, in French only) (opens in a new tab)
- Ordre des CPA du QuébecWhat’s new (CPA guide to good practices in artificial intelligence, in French only) (opens in a new tab)
- Ordre des CPA du QuébecModel guideline on the use of generative AI in an organization, announced April 8, 2025 (Word document, in French only) (opens in a new tab)
- Canada Revenue AgencyEFILE for electronic filers: eligibility, responsibilities and suspension (opens in a new tab)
- Canada Revenue AgencyForm T183 and authorizing a representative (opens in a new tab)
- Canada Revenue AgencyForm T183CORP, Information Return for Corporations Filing Electronically (opens in a new tab)
- Canada Revenue AgencyAuto-fill My Return for professional tax preparers (opens in a new tab)
- Canada Revenue AgencyCRA’s online chat services (opens in a new tab)
- Revenu QuébecProfessional representatives: filing for an individual (in French only) (opens in a new tab)
- Revenu QuébecProfessional representatives: filing for a business (in French only) (opens in a new tab)
- Revenu QuébecProfessional representatives: preparing an individual’s return, downloading tax data (in French only) (opens in a new tab)
- Revenu QuébecTerms of use: Représentant PRO (in French only) (opens in a new tab)
- Statistics CanadaAnalysis on artificial intelligence use by businesses in Canada, second quarter of 2026 (opens in a new tab)
- Statistics CanadaAnalysis on planned use of artificial intelligence by businesses in Canada, third quarter of 2026 (opens in a new tab)
- Statistics CanadaThe Daily: Use of generative artificial intelligence tools among Canadian workers, March 2026 (opens in a new tab)
- Statistics CanadaNAICS Canada 2022 Version 1.0: structure of sector 54, Professional, scientific and technical services (opens in a new tab)
- Microsoft LearnInterpret and improve model accuracy and confidence scores (Azure Document Intelligence) (opens in a new tab)
This article provides general information. It is not legal, tax or accounting advice for your situation.
Keep reading
- Tax
Tax season 2027: what an accounting firm can automate right now
Deadlines, electronic filing, the revised T183, document collection, reading slips: a firm’s plan between now and January.
8 min read
Your case deserves thirty minutes.
A conversation to understand your context and tell you, frankly, whether we are the right choice. The diagnostic that follows is free of charge.
Book a call