Lesson objective
Draft the elements of a firm AI governance programme at a depth that fits the firm, and build a permitted-use inventory that records every AI use, its status and its owner.
Plain-language explanation
Why this matters now
Module 1 named who is accountable. This module writes down what that person is accountable for: the firm's AI governance programme. Three 2026 publications describe, from different directions, what such a programme tends to contain.
- Professional guidance (IRS OPR, June 24, 2026; applies Circular 230, which binds practitioners before the IRS): Alert 2026-19 says that under §10.36 firms must have AI policies and procedures covering staff training, secure data handling, accuracy monitoring and vetting of third-party AI tools, with those steps documented. Its listed best practices include documenting AI use and verification, never uploading sensitive data to unsecured sites, and treating AI text as drafts.
- Professional guidance (Texas only; newsletter article, not a rule): the Texas Board's August 2026 Board Report lists common elements of firm AI policies: approved applications and vendors, confidentiality restrictions, review procedures, documentation, training, cybersecurity, and monitoring of guidance.
- Professional guidance (COSO, February 23, 2026; voluntary, not a standard): Achieving Effective Internal Control Over Generative AI (GenAI) names risks including hallucination, prompt-injection attacks, opaque reasoning, model drift, rapid configuration changes, bias, third-party opacity, and "shadow AI" outside formal IT oversight. It is written for controllers, financial reporting groups, internal audit, board committees and external auditors assessing GenAI-related controls.
None of the three sets a size. A sole practitioner and a 900-person firm read the same alert. Sizing is the firm's judgement, and the programme should record how it was made.
A question to sit with
If you listed every AI use in your firm today, including features switched on inside software you already license, how many would have a named owner and a recorded decision behind them?
What goes wrong
The programme is copied, not sized. A 12-person firm adopts a 60-page policy written for a national firm. Nobody reads it, nobody can operate it, and its controls exist only on paper.
The programme is too thin to operate. A one-paragraph memo says "use AI responsibly." It lists no approved tools, no restricted data and no reviewer. It cannot be followed, audited or enforced.
The inventory lists tools, not uses. "We use Tool X" hides the difference between drafting an internal email and drafting a tax position memo. The risk sits in the use, not in the product name.
Embedded features are missed. AI features inside tax, audit and practice-management software are switched on by a vendor update. Nobody bought a new tool, so nobody added a line.
The inventory is a one-off. It is compiled for a partners' meeting and never updated. Six months later it describes a firm that no longer exists.
The practical answer
Start from a crosswalk of the elements the sources describe. The left two columns report what each source says. The right-hand column is an AI Coalition Network recommendation for the programme section that covers it.
| Programme section | OPR Alert 2026-19 (under §10.36 and best practices) | Texas Board Report (common elements) |
|---|---|---|
| 1. Accountability and decision rights | Final decisions rest with qualified professionals | Not listed as an element |
| 2. Permitted-use inventory and approved tools | Vetting third-party AI tools; secure, enterprise-approved AI | Approved applications and vendors |
| 3. Data rules | Secure data handling; never upload sensitive data to unsecured sites | Confidentiality restrictions; cybersecurity |
| 4. Review and verification | Accuracy monitoring; verify facts, citations and calculations; treat AI text as drafts | Review procedures |
| 5. Training | Staff training | Training |
| 6. Records | Steps documented; document AI use and verification | Documentation |
| 7. Monitoring and review cycle | OPR says its guidelines are introductory and it may issue more | Monitoring of guidance |
Seven sections. Every element either source names has a home, and section 1 comes first because Module 1 showed that nothing else works without it.
Size the programme by the firm's exposure, not its headcount alone. The sizing grid is an AI Coalition Network recommendation. Place the firm by the highest row that applies.
| Exposure | Typical profile | Programme depth |
|---|---|---|
| Low | AI used only for internal drafting; no client data enters any tool | Short written programme; inventory reviewed twice a year; one owner |
| Moderate | AI touches client data in tax or CAS work; no attest use | Full seven sections; quarterly inventory review; owner plus deputy; training records |
| High | AI used on attest engagements, written tax advice, or issuer clients' reporting | Full seven sections inside the quality system (Lesson 4); monthly inventory changes reviewed; separate stop authority; annual independent review |
A 10-person firm doing written tax advice with AI sits in the high row. A 200-person firm that uses AI only to draft internal emails may sit in the low row. Headcount changes how much effort each section takes, not which sections exist.
Build the inventory one use at a time. Each line records a use, not a product.
| # | Field | What to record |
|---|---|---|
| 1 | Use | What the AI does, in one sentence |
| 2 | Tool and version | Product, feature and vendor, as named in the contract |
| 3 | Service line | Tax, attest, advisory, CAS or internal |
| 4 | Data | The most sensitive class of information the use touches |
| 5 | Capability type | One of the eight COSO GenAI types (below) |
| 6 | Status | Permitted, permitted with conditions, prohibited, or not yet assessed |
| 7 | Conditions | Review depth, data limits, who may use it |
| 8 | Owner | The named professional who owns the use |
| 9 | Decision reference | Link to the decision record (Module 3) |
| 10 | Next review | Date, and the events that bring it forward |
Classify by COSO capability type. COSO's publication organises GenAI uses into eight capability types: ingestion, transformation, posting, orchestration, judgment, monitoring, knowledge retrieval, and human-AI interaction. (COSO's press release names one type "regulatory intelligence"; the publication says "knowledge retrieval," and that is the list to cite.) Level 3 used these types to locate uses inside a workflow. At firm level they do a different job: they show leadership where the portfolio is concentrated. A firm whose uses cluster in posting and judgment has a different risk profile from one whose uses are mostly transformation. COSO's guidance is voluntary and not authoritative for SEC or PCAOB purposes.
Use four statuses and define them.
- Permitted: approved for the stated use with the firm's standard review.
- Permitted with conditions: approved only within written limits, such as no client identifiers, or partner review of every output.
- Prohibited: not to be used for this purpose.
- Not yet assessed: discovered or requested, and not to receive client information until assessed (Lesson 2).
Remember this
- A programme has seven sections; the elements named in the OPR alert and the Texas Board Report each have a home in them.
- Size by exposure: written tax advice or attest use puts a small firm in the high row.
- Inventory uses, not products, and include AI features inside existing software.
- Four statuses, with "not yet assessed" meaning no client information until someone decides.
- COSO's eight capability types show leadership where the firm's AI use is concentrated.
Accounting example
Linmoor & Hart CPAs is a fictional 18-person firm with tax, CAS and a small review practice. Tax partner Beatrice Okonkwo, named accountable for AI in Module 1, drafts the programme. She first places the firm on the sizing grid. Staff use a drafting assistant on written tax advice memos, which puts the firm in the high row despite its size. She then builds the inventory by interviewing each service line and checking vendor release notes. She finds 11 uses across 5 products. Tax has 4 uses: drafting advice memos, summarising IRS notices, extracting data from client documents, and drafting engagement emails. CAS has 4: bank reconciliation matching, categorising transactions, drafting monthly commentary, and answering client questions through a portal chatbot. The review practice has 1: summarising client contracts. Internal has 2: meeting notes and marketing copy. That is 4 + 4 + 1 + 2 = 11. Classified by COSO capability type, 5 are transformation (memo drafting, notice summaries, engagement emails, commentary, contract summaries), 2 ingestion (document extraction, bank matching), 1 posting (transaction categorisation feeding the ledger), 1 human-AI interaction (the portal chatbot) and 2 more transformation for internal use, so 7 transformation in all: 7 + 2 + 1 + 1 = 11. Beatrice sets statuses. Meeting notes and marketing copy are permitted. Memo drafting and document extraction are permitted with conditions: partner review of every citation and a ban on pasting full returns. The posting use is permitted with conditions and daily review. The portal chatbot is prohibited, because it answers clients directly with no reviewer, and the contract summary feature, switched on by a vendor update, is not yet assessed. Each line gets an owner and a next review date. Her programme runs to nine pages, not sixty.
Responsible-use warning
Do not copy another firm's AI policy and call it a programme, and do not inventory product names instead of uses. A long policy the firm cannot operate is as weak as a one-line memo nobody can enforce. Size the programme by exposure, record each use with its data, capability type, status and owner, and include AI features that arrived through software updates. An inventory that is not kept current soon describes a firm that no longer exists.
Human-review requirement
The partner accountable for the programme owns the sizing decision and the inventory, and signs both. They confirm that each use line names a qualified owner: a licensed CPA for attest and reports carrying the firm's name, the signing preparer or practitioner for tax, and the responsible controller or CAS lead for client books. They check that statuses are justified, that "not yet assessed" uses receive no client information, and that the inventory is reviewed at the cadence the firm's exposure requires.
Practical exercise
Use a fictional firm only.
- Describe a fictional firm in three sentences: size, service lines, and whether AI touches client data, written tax advice or attest work.
- Place the firm on the sizing grid and write one sentence justifying the row.
- Build the seven-section crosswalk for your firm, noting which sections already exist in some form and which are missing.
- List at least ten AI uses across the firm, including at least two features that arrived inside existing software.
- Complete all ten inventory fields for each use.
- Classify every use by COSO capability type and count the uses per type. Check that the counts add up to the total number of uses.
- Assign each use one of the four statuses, and write the conditions for every "permitted with conditions" use.
- Identify the capability type where the firm's uses are concentrated and write one sentence on what that means for review depth.
- Set the inventory review cadence and the events that would trigger an early review.
- Name who signs the inventory and what they confirm before signing.
- Pick one "not yet assessed" use and write the first three questions its assessment must answer.
- Estimate how many hours a year the inventory reviews will take at your chosen cadence, and show the arithmetic.
Knowledge check
-
What common elements of firm AI policies does the Texas Board Report list?
Answer guidance: Approved applications and vendors, confidentiality restrictions, review procedures, documentation, training, cybersecurity, and monitoring of guidance. It is a Texas newsletter article, not a rule.
-
Why should the inventory record uses rather than products?
Answer guidance: Risk depends on what the AI does, on which data, for which service line. The same product can be fine for internal drafting and unsuitable for written tax advice.
-
What are COSO's eight GenAI capability types?
Answer guidance: Ingestion, transformation, posting, orchestration, judgment, monitoring, knowledge retrieval, and human-AI interaction, as listed in the publication.
-
Why can a 10-person firm sit in the high row of the sizing grid?
Answer guidance: The grid sizes by exposure, not headcount. Using AI on written tax advice, attest engagements or issuer reporting puts a firm in the high row whatever its size.
Sources and further reading
- Internal Revenue Service, Office of Professional Responsibility (OPR) — Introductory Guidelines for Responsible AI Use in Federal Tax Practice (Alerts from the Office of Professional Responsibility, Issue Number 2026-19) (2026-06-24). Accessed 2026-09-15.
- Texas State Board of Public Accountancy — AI and the Texas CPA: Professional Obligations in a Changing Environment (Texas State Board Report, Vol. 168, August 2026) (2026-08-01). Accessed 2026-09-15.
- Committee of Sponsoring Organizations of the Treadway Commission (COSO) — Achieving Effective Internal Control Over Generative AI (GenAI) (2026-02-23). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.