SQMS 1 and AI: Documenting AI Tools in Your Firm's Quality System

A CPA firm should document AI tools in its SQMS No. 1 system as technological resources: list each tool, assess the quality risks it creates, record the responses that control them, monitor those responses, and show how engagement review catches AI errors. Keep the evidence together so a reviewer can follow it.

The dates shape the work. Firms had to design and implement their systems by the 15 December 2025 effective date, and the standard requires the first evaluation within one year of implementation, so by 15 December 2026 at the latest, then every year after. Every firm is now past implementation, so an AI tool added since then is a change to a live system: it needs a documented risk assessment before that evaluation, not a separate project.

Follow me on Instagram@sagnikteaches

Where AI sits in the standard

SQMS No. 1 replaced the old quality control standard with a risk-based approach. The AICPA's practice aid on the standard works through eight components: the risk assessment process, governance and leadership, relevant ethical requirements, acceptance and continuance, engagement performance, resources, information and communication, and monitoring and remediation. For each, the firm sets quality objectives, identifies and assesses quality risks that could stop those objectives being met, and designs responses. A quality risk, in the standard's sense, is one with a reasonable possibility of occurring and of adversely affecting one or more quality objectives, so not every theoretical AI worry needs an entry.

Connect on LinkedInSagnik Bhattacharya

The resources component is where most AI tools land. AICPA's own webcast material on this component describes the resources a quality management system has to manage as human, technological, intellectual and financial. An AI assistant that drafts a memo, a tool that reads invoices, or a bookkeeping platform with a machine-learning feature is a technological resource, and the vendor behind it is an outside supplier whose terms you need on file, so each needs an owner, a purpose and controls.

Subscribe on YouTube@sagnikteaches

You do not need a separate document for AI. The AICPA practice aid, Establishing and Maintaining a System of Quality Management for a CPA Firm's Accounting and Auditing Practice, includes libraries of example quality risks and responses for small and medium firms and for sole practitioners, plus a fillable spreadsheet for documenting the risk assessment. If your firm built its system on that spreadsheet, add the AI rows there, so the evaluation reads one record instead of two.

But AI tools touch more than resources. A tool that summarises a client's board minutes touches engagement performance. A tool that drafts client emails touches information and communication. A tool that stores client data touches relevant ethical requirements through confidentiality. That is why a single register entry should name every component it affects, not just one.

The register: one row per tool

Start with an inventory, because you cannot assess what you have not listed. Ask every partner and staff member, then check the firm's software licences, browser extensions, practice-management add-ons and meeting tools. AI is also built into products the firm already owns, so look at the feature lists of the tax, audit, document and accounting packages, not only at standalone chatbots.

Keep the register in the same place as the rest of the system documentation. A firm that stores policies in one folder and AI approvals in a chat thread will not find the evidence when it is asked for.

Writing risk entries that connect to objectives

The standard asks the firm to identify risks to its quality objectives, assess them by likelihood and effect, and design responses. A good AI risk entry therefore has four parts: the objective at risk, the risk, an assessment, and the response. Keep each short and specific.

Take a firm of eight, an illustration. Its quality objective for engagement performance says that engagement work is performed in accordance with professional standards. Here are four AI risk entries.

1. Confidentiality of client data. Risk: staff paste client data into a consumer chatbot that retains prompts. Assessment: likely without a rule, serious if it happens. Responses: approved-tool list, a business plan with training use switched off, a written rule, and an annual acknowledgement by all staff.
2. Inaccurate output in workpapers. Risk: an AI summary of a lease is wrong and enters a workpaper unchecked. Assessment: possible, moderate. Responses: AI-drafted text is labelled, a preparer ticks source checks, and the reviewer samples AI-assisted workpapers.
3. Over-reliance and loss of skill. Risk: junior staff accept AI output without understanding. Assessment: possible over time. Responses: training on tool limits, and a rule that conclusions are written by the preparer.
4. Vendor change. Risk: the vendor changes its data terms or retires a feature. Assessment: possible. Responses: a named owner subscribes to vendor notices, and the register has a review trigger.

Number the risks, link them to the register rows, and keep the language plain. The test of an entry is whether a reviewer who has never seen your firm can tell what could go wrong and what you do about it.

A prompt to draft entries, with a sample output

An assistant can draft first-pass entries for you to edit, provided you give it only the tool description and never client data.

I run quality management for a [SIZE]-person accounting firm. We use [TOOL] for [TASKS].
Draft quality risk entries. For each give: objective affected, risk, likelihood (low/medium/high),
effect (low/medium/high), proposed response, and how we would monitor the response.
Cover: confidentiality, accuracy, supervision and review, staff competence, vendor change,
client communication. Keep each entry under 60 words. Do not invent vendor facts.

An illustrative result for a meeting-notes tool: "Risk: transcripts of client calls stored on vendor servers beyond the retention we have disclosed to clients. Likelihood: medium. Effect: high. Response: record only with consent, set retention to the shortest available, delete after notes are filed. Monitor: quarterly check of the tool's retention settings." What you would fix: confirm the real retention options on the vendor's page, and check the response against the firm's confidentiality policy and client letters. The model cannot know either.

Responses: the five controls most small firms need

Responses should operate and be evidenced. Think of five controls that cover most small-firm AI use.

• Approved-tool control. Only listed tools may touch client work. Evidence: the register and a signed policy.
• Data rule. A list of what may and may not be entered, for example no names, tax IDs or account numbers into tools without an approved data agreement. Evidence: policy, staff acknowledgements.
• Output verification. A named step for checking AI output against source documents. Evidence: workpaper tick marks or a field in the checklist.
• Training and competence. Short training on approved tools, their limits and the rules. Evidence: attendance records and dates.
• Vendor due diligence. Reading data terms, retention, training use, sub-processors and security documentation before approval, and again on change. The AI vendor questions tutorial has a list to start from. Evidence: a dated note with links to the terms reviewed.

Engagement-level review: where AI shows up in the file

The standard's engagement performance component is about work being done and reviewed properly. For AI, the practical change is small and visible: the file should make clear when AI helped and who checked it.

Add three lines to the engagement checklist:

1. Was an approved AI tool used on this engagement? If so, which and for what?
2. Was AI output used in a workpaper, memo or client communication, and has the preparer verified it against source documents?
3. Did the reviewer see which sections were AI-assisted and test a sample?

Here is an illustrative case. A preparer uses an assistant to summarise a 40-page supplier contract for a review engagement. The workpaper note says: "Summary drafted with approved assistant; verified against clauses 4, 7, 12 and schedule B; payment terms corrected from 45 to 30 days." The reviewer sees the correction, which shows the check worked, and the evidence stays in the file. Without that line, the same error would have been invisible.

For engagements that fall under an engagement quality review, check SQMS No. 2, the AICPA's separate standard on engagement quality reviews, and your firm's criteria for which engagements need one. Whether AI use changes the risk profile for a particular engagement is a judgement to document, not a blanket rule.

Monitoring: how you know the responses operate

Monitoring and remediation is a component of its own. For AI, three simple activities give you evidence.

• Quarterly register check. The owner confirms the tool is still approved, the vendor terms are unchanged, and the user list is correct. Takes 20 minutes per tool.
• Sample file review. Each quarter, pick two or three completed files where AI was used. Check the AI line was completed, the verification was evidenced, and no prohibited data was entered. Log what you found.
• Shadow-AI check. Review browser extensions, expense claims for AI subscriptions and a short anonymous staff question: "Have you used an unapproved AI tool for work?" The point is to find out, not to punish.

When monitoring finds a deficiency, the standard expects you to investigate the cause and remediate. Write the finding, the root cause, the fix and the date. "Two staff used a free chatbot for email drafts because the approved tool had no mobile app" leads to a fix that changes the system, not just a reminder email.

Evidence pack: what a reviewer will ask to see

The last row ties back to governance. Leadership has to evaluate the system as a whole, and your AI register, risk entries and monitoring log are inputs to that conclusion. If they are tidy, the evaluation takes minutes.

A walked example: an eight-person firm, a Friday afternoon

Illustration only. The firm has three tools: a general-purpose assistant, a meeting-notes tool and an AI feature inside its tax software. The quality lead takes about four hours.

1. Inventory (45 minutes). Survey of eight staff finds a fourth tool, a free browser extension a manager installed.
2. Register (60 minutes). Four rows created. The extension is marked not approved and removed.
3. Vendor terms (60 minutes). The quality lead reads data and retention terms for each and records the date and the setting changes needed, such as training use off.
4. Risk entries (45 minutes). Six entries drafted with the prompt above, edited by the lead.
5. Checklist lines (15 minutes). Three AI lines added to the engagement checklist.
6. Policy and training (30 minutes). One-page rule and a 30-minute session scheduled.
7. Monitoring plan (15 minutes). Quarterly diary entries for the register check and file sample.

None of this needed special software. It needed one owner and a fixed hour each quarter.

AI inside software you already own

The register entries people forget are the ones where AI arrived as a feature update. A tax package adds a "summarise this notice" button. A document platform starts suggesting classifications. A meeting tool offers action-item extraction. Nobody bought "an AI tool", so nobody assessed it.

Handle these with a standing trigger: every time a vendor announces a new feature, the tool owner asks three questions and writes the answers in the register. Does the feature send client data to a third party that was not already covered by the vendor's terms? Can it be switched off by the firm, and who has the admin setting? Is it on by default? An illustrative case: a document platform enables AI classification for all accounts on a Monday. The owner reads the notice on Wednesday, finds the feature processes uploaded client documents, turns it off for the firm until the terms are reviewed, and records the dates. The register shows the risk was spotted and acted on within two days, which is exactly the kind of evidence monitoring is meant to produce.

The tools tutorial for small accounting firms lists the common packages, which is a useful prompt when you are working out which of your existing products have AI features.

Acceptance and continuance: clients who say no

The acceptance and continuance component is about taking on and keeping clients where the firm can meet its quality objectives. AI adds one practical question: does the client have rules on how its data may be processed? Some clients, such as those in regulated sectors or with contractual confidentiality duties, forbid the use of third-party AI tools on their information.

Record this at acceptance. A simple field on the client acceptance form, "AI tools permitted on this client's data: yes, no, restricted", does the job. For a "no", the engagement team needs to know, and the engagement checklist line about AI use should reflect it. A realistic slip: a staff member uses the meeting-notes tool on a client call where the client had asked for no recording. The monitoring sample catches the mismatch because the file says "AI used" and the acceptance form says "not permitted". That is a deficiency, a root cause (the acceptance flag never reached the team) and a fix (the flag is now printed on the engagement checklist).

A before-and-after register entry

Weak entry: "AI tool: ChatGPT. Use: general. Risk: confidentiality. Response: be careful." Nothing in it can be tested.

Better entry: "General-purpose assistant, business plan, owner: operations manager. Approved for: drafting internal and client emails from templates, summarising published guidance. Not approved for: entering client names, identification numbers or financial statements; forming audit or tax conclusions. Training use: switched off, checked on the quarterly register check. Risk R-14: client data entered in error. Responses: data rule, annual acknowledgement, quarterly sample of prompt history where the business plan allows it. Next review: next quarter-end or on a change to the vendor's data terms."

The second version names a person, sets a boundary and states what will be checked. A reviewer can compare it against what happened.

A staff survey that finds the shadow tools

Run this short survey, anonymous if the firm is large enough for that to be meaningful, before you finalise the register. Ask: which AI tools have you used for work in the last three months; what for; did it involve client information; which free accounts do you use; and what would make an approved tool more useful than the one you use now? The last question matters most, because the root cause of shadow use is usually that the approved option is slower or missing.

In an illustrative eight-person firm, the survey finds three people using a free chatbot on their phones to polish emails. The fix is not a stern memo. It is an approved assistant on a business plan, with mobile access, a one-page rule and a 30-minute walk-through.

Common gaps in firms' AI documentation

• A policy exists but no register, so nobody can say which tools are in use.
• The register exists but vendor terms were never read, or were read once two years ago.
• Risk entries describe generic AI risk, not the firm's tools and tasks.
• No link between AI use and engagement files, so a reviewer cannot see what AI touched.
• No monitoring evidence, only the intention to monitor.
• AI built into existing software is ignored because it was not "bought" as AI.

Client-facing consequences

Your engagement letters may need wording about the use of AI-assisted tools and confidentiality safeguards. The AI engagement letters tutorial covers drafting. Check with your professional advisers on wording and on what your regulator or insurer expects, because requirements differ between firms and engagement types.

Also check your professional liability insurer's position on AI use, and record the answer. It is not an SQMS requirement, but it affects how you describe risk responses to partners.

Five jobs before the first annual evaluation

1. Put the first evaluation date in the calendar (within a year of implementation, so by 15 December 2026 at the latest) and work back from it.
2. Run the inventory and build the register.
3. Assess risks for each tool and link them in the system file.
4. Add the three engagement checklist lines.
5. Set the quarterly monitoring entries and run the first sample review.

If you would rather map this with someone, my AI implementation consultation covers finding the AI jobs in your firm and writing the controls around them. Treat this tutorial as a working method, not as professional advice on the standard: your firm's quality leader, advisers and the standard itself decide what is required.

SQMS 1 and AI: questions firm leaders ask

Does SQMS No. 1 mention AI?

It does not need to name AI to apply. It requires firms to identify quality risks and respond to them, and its resources component covers technological resources. An AI tool the firm uses on engagements is a technological resource, so it belongs in your risk assessment. Check the standard's current text.

Do we have to document free tools staff use on their own?

Ideally you prohibit unapproved tools for client work and document that rule as a response. If staff use them anyway, that is a quality risk you identify, and your monitoring should look for it. An undocumented tool touching client data is worse than a documented, restricted one.

How detailed must the AI tool inventory be?

Detailed enough that a reviewer can see what each tool does, who owns it, what data it touches, what engagements it is used on and what checks apply. A one-page register per tool or a single sheet with a row per tool usually works for a small firm. Scale to your size and risk.

Who signs off an AI tool for use?

The firm decides, but the standard requires the managing partner or equivalent to be assigned ultimate responsibility and accountability for the system. A sensible small-firm pattern is a named person who owns the tool, the managing partner or quality lead who approves it, and a written date. Keep the approval and the reasons together in the system file.

Further reads

• AI Readiness Checklist for Accountants, Solicitors, Consultants — A wider readiness check before any tool is approved.
• What an AI Implementation Looks Like in a Small Accounting Firm — What a rollout looks like inside a small firm.
• How Much Does AI Cost a Small Accounting Firm? — Budget the tool and the review time together.
• What to Check in an AI Tool's Privacy Policy and Terms — Read data and retention terms before approving a tool.
• How to Chase Missing Client Records With AI Before Deadlines — A common AI use case to put through the register.
• How Accountants Use AI to Explain Tax to Clients in Plain English — Client-facing AI drafting that needs review controls.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Journal of Accountancy (AICPA & CIMA) articles on SQMS No. 1, its effective date, eight components, evaluation timing and the AICPA quality management practice aid; AICPA SQMS No. 2, Engagement Quality Reviews; AICPA quality management webcast description of the resources component.