Lesson objective

Use COSO's eight GenAI capability types to choose which accounting tasks a session covers, and match each chosen task to the risks and the behaviour it calls for.

Plain-language explanation

Why this matters now

A session titled "AI for Accountants" has no edges. It could cover research, drafting, reconciliations, journal entries, close monitoring, client chat and agents, and in sixty minutes it will cover each of them badly. Designing backwards requires a scope: which tasks, for which people, with which behaviour.

A useful way to draw that scope comes from internal control guidance.

Professional guidance (voluntary): COSO, "Achieving Effective Internal Control Over Generative AI (GenAI)", released 23 February 2026. Commissioned by COSO and written by Scott Emett, Marc Eulerich, Jason Guthrie, Jason Pikoos and David A. Wood, it adapts "the five components and 17 principles" of COSO's 2013 Internal Control-Integrated Framework to GenAI "instead of creating a separate governance framework." It organises GenAI uses into eight capability types: ingestion, transformation, posting, orchestration, judgment, monitoring, knowledge retrieval, and human-AI interaction. Risks it names include hallucination, prompt-injection attacks, opaque reasoning, model drift, rapid configuration changes, bias, third-party or vendor opacity, and "shadow AI" outside formal IT oversight. It is voluntary guidance, not authoritative for SEC or PCAOB purposes, and it does not amend the 2013 framework. The COSO press release lists "regulatory intelligence" where the publication says "knowledge retrieval"; use the publication's list.

The capability types were written for internal control, not for teaching. They are useful to an instructor because they sort accounting tasks by what the tool is doing, and what the tool is doing decides what the learner must check.

Professional guidance (staff observations, not binding): PCAOB staff Spotlight, July 2024, announced 22 July 2024. Limited outreach, mainly to larger firms, found that current GenAI integration in audits by the firms interviewed "appears to be focused primarily on administrative and research activities," and that some firms did not allow GenAI when performing audit or attest procedures, citing data privacy and reliability concerns. It represents the views of PCAOB staff, "not a rule, policy, or statement of the Board," and it is a mid-2024 snapshot of mostly large firms, not a description of practice in 2026. It is a reason to ask your room what they actually use, not to assume.

A question to sit with

Of the eight things a GenAI tool can do in an accounting process, which one or two does your room actually do this month, and what goes wrong when they do it?

What goes wrong

1. Everything, briefly. Eight capability types, eight slides, no practice. Learners leave knowing the tool can do many things and knowing how to check none of them.

2. The scope follows the vendor. The session covers what the firm's tool is marketed to do rather than the tasks the room performs.

3. The high-risk task is taught as a convenience. Posting entries or orchestrating a workflow is shown as a time saver, with no approval gate and no named approver. Level 5 of this programme covered automation; a teaching session that touches it must carry its controls with it.

4. The snapshot is taught as current practice. "Audit firms only use AI for admin" is a mid-2024 staff observation from limited outreach. It is not a rule, and it is not a description of your room in 2026.

5. The capability types are taught as a requirement. COSO's guidance is voluntary. Using its categories to scope a session is a design choice, not compliance.

The practical answer

Best practice: use the capability types as a scoping grid. The accounting examples below are fictional illustrations of each type; the risks are drawn from the risks COSO names. The behaviour column is practice, not a COSO requirement.

Capability type Fictional accounting example Risks most in play Behaviour a session might teach
Ingestion Reading a stack of vendor invoices into structured fields Hallucination; prompt injection hidden in a document Compare extracted fields with the source document on a sample
Transformation Summarising a lease or rewriting a memo Hallucination; opaque reasoning Check each summary statement against the source text
Posting Proposing journal entries into the ledger Hallucination; rapid configuration changes Approve each proposed entry before it posts, as a named approver
Orchestration Routing a document request through several steps Configuration changes; vendor opacity Know where the workflow stops for a person, and stop it when something looks wrong
Judgment Suggesting a classification or an accounting treatment Hallucination; bias; opaque reasoning Treat the suggestion as a proposal and document the professional's own conclusion
Monitoring Flagging unusual transactions during close Model drift Investigate flagged items and sample unflagged ones
Knowledge retrieval Answering a question from tax or accounting guidance Hallucination (fabricated citations) Verify every citation against the primary source
Human-AI interaction A chat assistant used by staff or clients Shadow AI; confidentiality Use only the approved tool, with only the data it is approved for

Best practice: choose one or two capability types per session. A sixty-minute session can teach one behaviour well. Choose by three questions: Which tasks does this room perform this month? Which carry the most risk to a client or a filing? Which have a firm procedure the behaviour can attach to?

Best practice: let the capability type choose the check. Knowledge retrieval calls for citation verification, which OPR Alert 2026-19 describes for practitioners before the IRS: verifying "facts, citations, and calculations produced by AI." Posting calls for approval before an entry lands, by a named person. Judgment calls for the professional's own conclusion, documented. A session that mixes types should make the different checks visible, not blur them into "always check".

Best practice: treat scope as a promise. Put the scope on the first slide: "Today: using AI to answer tax research questions, and checking what it cites. Not today: posting, agents, or client-facing chat." Learners can plan around a clear edge; they cannot plan around a vague one.

AI Coalition Network recommendation: scope the first session a team receives to the capability types the firm has approved for that team. Teaching staff to post entries or run workflows the firm has not approved invites shadow AI, which COSO names as a risk. It is our recommendation, not a requirement.

Remember this

  • COSO's eight capability types are ingestion, transformation, posting, orchestration, judgment, monitoring, knowledge retrieval and human-AI interaction.
  • COSO's guidance is voluntary and not authoritative for SEC or PCAOB purposes; using it to scope a session is a design choice.
  • Choose one or two capability types per session, by what the room does, the risk and the procedure available.
  • The capability type chooses the check: citations for retrieval, approval for posting, the professional's own conclusion for judgment.
  • The PCAOB staff Spotlight is a mid-2024 snapshot of mostly large firms; ask your room instead of assuming.

Accounting example

Rhys Adebayo, a fictional controller who trains the finance team at the fictional Northmoor Components Inc., is asked to run a sixty-minute AI session for 16 people: 7 in accounts payable, 5 in general ledger and close, and 4 in financial planning and analysis. Seven plus five plus four makes sixteen. His pre-session form asks what each person uses a tool for. Accounts payable staff use the company's approved tool to read invoices into fields: ingestion. Close staff use it to draft variance commentary: transformation. Planning staff ask it questions about accounting guidance: knowledge retrieval. Nobody is approved to have the tool post entries, and the company's procedure says entries are proposed and then approved by the assistant controller. Rhys scopes the session to ingestion and transformation, which together cover 12 of the 16 attendees, and gives the planning team a separate thirty-minute session on verifying citations. His first slide reads: "Today: reading invoices and drafting commentary, and checking both. Not today: posting entries or workflows." In practice, accounts payable staff compare extracted fields with fictional invoices on a sample and find one invoice where a planted note in the document text tried to change the payee field; Rhys names prompt injection as a risk COSO lists, shows the check that caught it, and gives no technique for writing such a note. Close staff check each sentence of a fictional commentary against a fictional ledger. Rhys labels the capability types as COSO's voluntary guidance and tells the room it is a way to organise the session, not a rule.

Responsible-use warning

Never try to cover all eight capability types in one session, and never scope a session to what a vendor markets rather than what the room does. Never teach posting or orchestration without the approval point and named approver that control them. Never present COSO's guidance as a requirement, and never present the PCAOB staff Spotlight's mid-2024 observations from mostly large firms as a rule or as current practice.

Human-review requirement

The instructor owns the scope and the accuracy of every statement about COSO and the PCAOB staff Spotlight. Before delivery, the person who owns the firm's or company's AI procedures confirms which capability types are approved for the audience, and that nothing in the session teaches a use the procedures do not permit. A second reviewer checks that COSO is labelled as voluntary guidance and the Spotlight as a mid-2024 staff snapshot. The controller or engagement owner remains responsible for any output produced with the tool in real work.

Practical exercise

Scope a session with the capability grid. Use a fictional firm or company, fictional staff and fictional tasks throughout.

  1. Write a fictional audience of 12 to 25 people by team, and check that the counts add up.
  2. For each team, list two AI-assisted tasks they might perform, and label each with one of COSO's eight capability types.
  3. For each task, name one or two risks from the list COSO gives, and explain in one sentence why they are in play.
  4. Write down which capability types the fictional organisation has approved for this audience.
  5. Choose one or two capability types for the session, and justify the choice with the three questions: what the room does, the risk, and the procedure available.
  6. For each chosen type, write the behaviour and the check the session will teach.
  7. Write the scope line for your first slide, stating what the session covers and what it does not.
  8. Write one sentence that labels COSO's guidance correctly, and one that describes the PCAOB staff Spotlight with its date and limits.
  9. List the capability types you left out, and for each write where a learner who needs it will be sent instead: a later session, a reference handout or the owner of the firm's AI procedures.
  10. Check your scope line against the approvals in step 4 and remove any use the fictional organisation has not approved.

Knowledge check

  1. 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.

  2. Is COSO's GenAI guidance a requirement for companies or auditors?

    Answer guidance: No. It is voluntary, commissioned guidance that adapts the 2013 framework, is not authoritative for SEC or PCAOB purposes and does not amend the framework.

  3. How does the capability type help choose what a session teaches?

    Answer guidance: It shows what the tool is doing, which decides the check: citation verification for retrieval, approval for posting, the professional's own conclusion for judgment.

  4. How should an instructor describe the PCAOB staff Spotlight's finding about audit use?

    Answer guidance: As a mid-2024 staff observation from limited outreach to mostly larger firms that use appeared focused primarily on administrative and research activities, not a rule or a picture of 2026.

Sources and further reading