Lesson objective
Write an agent authority register entry that states the capability type, rung, permitted tools, scope, stop conditions and named owners before an automation touches live client data.
Plain-language explanation
Why this matters now
Lessons 1 and 2 produced two decisions for every automation: the rung it runs on, and the tools and reach it gets. Decisions held in a meeting, an email thread or a vendor's settings screen drift. An administrator widens a scope to fix a problem in deadline week. A new manager assumes an agent was cleared for their client because it runs on the client next door. The integrator moves on and nobody remembers the rules.
The fix is unglamorous: write the authority down, in one place, before the automation touches anything live, and treat any change to what it may do as a change to that entry first.
Two sources support writing it down, and neither prescribes a format.
- Professional guidance (IRS Office of Professional Responsibility, Alert Issue Number 2026-19, distributed 24 June 2026; interprets Circular 230, which binds practitioners before the IRS, and creates no new rule): under Circular 230 § 10.36, the alert says firms must have internal AI policies and procedures covering staff training, secure data handling and AI accuracy monitoring, and vetting of outsourced or third-party AI tools, and that those steps should be documented. Its best practices include documenting AI use and verification processes.
- Professional guidance (COSO, Achieving Effective Internal Control Over Generative AI, 23 February 2026; voluntary): it adapts "the five components and 17 principles" of the 2013 Internal Control Integrated Framework to GenAI "instead of creating a separate governance framework", and organises GenAI uses into eight capability types. Its intended users include controllers, internal audit and "External auditors assessing GenAI-related controls." It names "shadow AI" outside formal IT oversight among its risks.
The register in this lesson is an AI Coalition Network recommendation. It is not an official agent-authority standard, because none exists. Organising it by COSO's capability types is this programme's choice, made so that a controller, an internal auditor or an external auditor can read it in a vocabulary they may already use.
A question to sit with
If the partner who approved your first agent left tomorrow, could a new partner tell from one page exactly what that agent may do, for which clients, and who said so?
What goes wrong
The configuration is the only record. What the agent may do is whatever its settings say today. Nobody can tell whether a permission was granted deliberately or crept in.
The register describes intentions, not configuration. The entry says "read only" and the connector has write access. A register that is never reconciled to the live settings is a comfort document.
Shadow automations. A team builds a small workflow in a spreadsheet add-in or a personal automation account. It never reaches the register because nobody thought of it as an agent. COSO names shadow AI as a risk for exactly this reason.
Changes skip the register. Scope is widened in the settings first and the register is "updated later", which in deadline season means never.
No stop conditions. The entry says what the agent does on a good day and nothing about when it must stop and hand to a person.
The practical answer
Classify each automation by capability type. COSO's publication lists eight: ingestion, transformation, posting, orchestration, judgment, monitoring, knowledge retrieval and human-AI interaction. The accounting illustrations below are this programme's examples, not COSO's definitions.
| COSO capability type | Illustrative firm use | Default ceiling this programme suggests |
|---|---|---|
| Ingestion | Reading client uploads, bank statements and invoices into the firm's system | Rung 2 inside the firm |
| Transformation | Mapping a client trial balance to the firm's workpaper lead schedules | Rung 2 inside the firm, output reviewed |
| Posting | Proposing entries to a client ledger | Rung 1 (propose); a person posts |
| Orchestration | Routing a task, a request or a document through a workflow | Rung 2 inside; rung 3 for anything outward |
| Judgment | Suggesting a classification, a tax position or a materiality flag | Rung 1 only |
| Monitoring | Watching feeds or deadlines and raising alerts | Rung 2 for alerts to staff |
| Knowledge retrieval | Finding guidance, prior-year files or firm policy | Rung 1, every citation checked |
| Human-AI interaction | Answering staff or client questions in a chat interface | Rung 1 for staff; client-facing answers need a release |
Write one entry per automation. The fields below are this programme's template; the downloadable register uses them.
| Field | What it records |
|---|---|
| Automation name and ID | A stable name, and the automation identity it runs under |
| Capability type | One or more of COSO's eight |
| Purpose | The task in one sentence |
| Trigger | The event that starts it |
| Rung | 0 to 3, with the reason |
| Permitted tools | Each tool by name; everything else disabled |
| Scope | Clients, entities, folders and systems it may reach, read or write |
| Data handled | Including whether it touches tax return information or attest client data |
| Stop conditions | When it must halt and hand to a person |
| Never list | Acts it must not perform (Lesson 4) |
| Automation owner | The person accountable for how it runs |
| Output owner | The professional who owns what it produces |
| Approved by and date | Named, dated approval |
| Last reconciled | Date the entry was checked against live configuration |
Write stop conditions a machine can apply. "Stop if unsure" is not a stop condition. Write ones the automation can test: the document is not from a known client address; the amount differs from the purchase order by more than a stated tolerance; the bank details on an invoice differ from the vendor master; the client has no current consent on file for this use; the task would touch a second client. Each stop sends the item to a named person's queue.
Change the register first. Amend the entry, have it approved by the same owners, then change the configuration, keeping the prior version. The configuration should always be something the current entry permits.
Bring shadow automations in. Ask each team, at least yearly, what runs on its own: add-ins, scheduled scripts, personal automation accounts, email rules that forward to an AI tool. Register them or switch them off.
Link the register to the firm's other records. OPR's alert expects documented procedures; the register is one natural home for the automation part of them. Link each entry to the tool's vetting file (Level 4), the training record for staff who release its outputs and the action log specification (Module 11).
Remember this
- No official agent-authority standard exists; the register is this programme's recommendation.
- Classify each automation by COSO's eight capability types so controllers and auditors can read it.
- Each entry names the rung, tools, scope, stop conditions, never list and named owners.
- Change the register first, then the configuration, and reconcile the two on a schedule.
- OPR Alert 2026-19 expects firms' AI procedures to be documented; the register is one part of that.
Accounting example
Hollins Marsh CPAs is a fictional 25-person firm. Its operations partner, Ingrid Salo, finds that the firm runs four automations and nobody can list them in one place. She builds a register. The first entry is the portal intake agent: capability types ingestion and orchestration; purpose, file client uploads and open a preparer task; trigger, a new portal upload; rung 2; permitted tools, read the portal, write to the engagement folder, create a task; scope, the portal and document system only, no ledger or tax software access; data, includes tax return information; stop conditions, upload from an unrecognised account, a file that names a second client, or a client with no engagement letter for the current year; never list, no messages to clients, no deletion. The automation owner is the IT manager, Joel Brandt; the output owners are the engagement managers. Ingrid then reconciles each entry against the live settings with Joel. The intake agent's portal connector turns out to have a "notify client of receipt" function switched on, which sends an automatic message the register never allowed. Joel turns it off, and Ingrid records the finding and the fix. In the survey of teams she also finds a fifth automation: a senior in the tax department had set up an email rule that forwards organiser questions to a personal AI account. It is not registered and uses a tool the firm has not vetted. Ingrid has it switched off that day, records the incident for the firm's information security lead to assess, and asks the senior to propose the use properly through the register so it can be assessed.
Responsible-use warning
Do not let the register become a description of how you hope the automation behaves. Every entry must be reconciled against the live configuration on a schedule, and every change must go through the register first. An entry that says read only while the connector has write access is worse than no entry, because reviewers will rely on it. Unregistered automations built by individuals, including email rules and personal accounts, are the ones most likely to hold client data outside any oversight.
Human-review requirement
The automation owner maintains each entry and runs the reconciliation. The output owner approves the rung, scope and stop conditions for their work: the engagement partner or licensed CPA for attest engagements, the signing preparer for tax, the controller or responsible accountant for a client's books. The partner with principal authority for the firm's tax practice confirms that the register sits within the firm's documented AI procedures. Approvals are named and dated.
Practical exercise
Use a fictional firm, or your own firm with nothing confidential.
- List every automation you know of in the firm, including add-ins, scheduled scripts and email rules that send data to an AI tool.
- Ask two colleagues in different teams what runs on its own in their work, and add anything new to the list.
- Classify each automation by one or more of COSO's eight capability types.
- Choose the automation that handles the most sensitive data and complete every register field from this lesson for it.
- Write at least four stop conditions for it that the automation itself could test, and name the queue each stop sends work to.
- Write its never list.
- Open its live configuration, or a description of it, and compare every permission with your entry. List each difference.
- For each difference, decide whether the configuration changes or the register changes, and who must approve that.
- Set a reconciliation date and name the three people who would attend.
- Write two sentences explaining how this register connects to the firm's documented AI procedures and the tool's vetting file.
- Note which of your entries an external auditor or internal auditor could read without asking you a question, and rewrite one field in any entry that fails that test.
Knowledge check
-
What are COSO's eight GenAI capability types as listed in the publication?
Answer guidance: Ingestion, transformation, posting, orchestration, judgment, monitoring, knowledge retrieval and human-AI interaction. The press release says regulatory intelligence where the publication says knowledge retrieval; cite the publication.
-
What makes a stop condition usable?
Answer guidance: The automation itself can test it, for example bank details differing from the vendor master or a document naming a second client, and each stop sends the item to a named person's queue.
-
In what order should a change to an automation's scope be made?
Answer guidance: Amend the register entry first, have the same named owners approve it, then change the configuration, keeping the prior version so the live settings always match what the entry permits.
-
What does OPR Alert 2026-19 say firms should do about their AI procedures under Circular 230 section 10.36?
Answer guidance: Have AI policies and procedures covering staff training, secure data handling and accuracy monitoring, and vetting of third-party AI tools, and document those steps. The alert applies existing rules and creates no new one.
Sources and further reading
- Committee of Sponsoring Organizations of the Treadway Commission (COSO) — Achieving Effective Internal Control Over Generative AI (GenAI) (2026-02-23). Accessed 2026-09-15.
- 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.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-15.
- Cybersecurity and Infrastructure Security Agency (CISA), Australian Signals Directorate's Australian Cyber Security Centre, and U.S. and international partners — Careful Adoption of Agentic Artificial Intelligence (AI) Services (2026-05-01). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.