Responsible AI under Circular 230

The IRS says you may use AI, but you may not hand it your judgment. CitedTax is built on the same rule.

On June 24, 2026 the IRS Office of Professional Responsibility issued Alert 2026-19, Introductory Guidelines for Responsible AI Use in Federal Tax Practice. It creates no new rules; it applies the duties practitioners already carry under Circular 230 to the tools they use. This page states, duty by duty, what the product does to meet each one. Everything here describes what the system actually does today.

Duty by duty

Due diligence

Circular 230 §10.22

What OPR expects. Review every AI-generated document before it reaches a client or the IRS. Verify the facts, the citations and the calculations. AI output is a starting point, never a finished product.

What CitedTax does. Each finding carries an evidence grade. VERIFIED means the citation exists, the engine fetched the cited text and confirmed it supports the claim as written, and the sentence it quotes is in that text; CPA TO CONFIRM means the analysis stands but the best citation is secondary or could not be machine-confirmed; a citation the gate rejects is shown as rejected or withheld, never silently dropped. Every finding lists what the reviewing professional must confirm, and amber items cannot enter a client report until the professional marks them confirmed.

Competence

Circular 230 §10.35

What OPR expects. Understand how the tool generates content, where it fails, and whether its output suits an IRS matter.

What CitedTax does. The method is documented on this page and in every set of papers: retrieve primary authority from a closed index, ground each position in the retrieved text, check every citation against the source, grade it, and show the fact-to-law chain and the computation. The model provider is named below. The engine can be wrong; it is built to be checked, and it says so on every cover.

Firm procedures

Circular 230 §10.36

What OPR expects. Firm leaders must keep documented procedures: staff training, secure data handling, accuracy monitoring, and vetting of third-party tools, with the steps documented.

What CitedTax does. Every set of working papers ends with a diligence record: when it was generated, the vintage of the law index, how many passages were retrieved, the authorities relied on with their weight, and every citation the verification gate rejected or withheld, by name. Every memo carries the same record on its cover; every letter keeps it in the job file. The client report names the professional who reviewed it and the date. The activity log records every upload, download, approval and purge and can be exported by an administrator. On request we provide a security summary for your written information security plan, our data-processing addendum, and this page as the vendor-vetting record.

Written advice

Circular 230 §10.37

What OPR expects. Written advice must rest on reasonable factual and legal assumptions that the practitioner has independently verified. If a system's logic is opaque, relying on it may itself be unreasonable.

What CitedTax does. The output is not opaque. Each position shows the conditions the authority requires and the facts that satisfy each, the operative sentence quoted from the source, the computation, the limitations found in the authority, and the grade. Figures are labeled as estimates for professional validation on every cover.

Fees

Circular 230 §10.27

What OPR expects. Efficiencies from AI should reach the client. Billing for time the tool replaced, or billing twice for AI-assisted work, may be an unconscionable fee. Disclose the AI activities performed.

What CitedTax does. Every request has a fixed, visible price before it is submitted, so the cost of the AI-assisted work is a known figure the firm can pass through or absorb. The client report states that it was prepared with research and drafting software, that each cited authority was compared with its source text, and that the firm reviewed every item before release.

Confidentiality

Circular 230 §10.51(a)(15), IRC §§6713 and 7216

What OPR expects. Never place tax return information in an unsecured or public AI system. Use only secure, enterprise systems, and obtain the consents the Code requires.

What CitedTax does. Identifiers are replaced with placeholders before any text is analyzed. Every stored object is encrypted with a key unique to the firm. The source PDF is deleted once extraction is complete. Analysis runs through our model provider under commercial terms that forbid training on your data. Our data-processing addendum states our position under §7216, and the firm obtains any consent the section requires before uploading.

What the engine does

Retrieves primary and official authority from a closed, curated index: statutes, regulations, rulings, guidance, court decisions, forms and instructions, for the United States, the carried states, Canada and the United Kingdom.
Grounds every position in the retrieved text and checks each citation against the source before grading it.
Shows the full chain: the question, the facts, the conditions the authority requires, the operative text, the computation, the limitations, and what the professional must confirm.
Keeps the reviewing professional in control: nothing reaches a client until a licensed professional curates and approves it under the firm's name.
Writes a diligence record into every set of papers and keeps an exportable activity log.

What it does not do

Give tax advice to anyone. CitedTax is a research and drafting tool for licensed professionals; the firm validates every item before reliance.
Rely on open internet content. The index is built from primary and official sources, and only those sources are cited.
Train on your data, aggregate it across firms, or share it. See the security page for retention and subprocessors.
Hide its reasoning. A finding the engine cannot support is withheld with the reason stated, not written around.

The diligence record

Every set of working papers ends with a record the firm can file: the date and time it was generated, the vintage of the law index it was generated from, how many passages were retrieved, the authorities relied on with their weight (statute, regulation, ruling, guidance, court decision, form or publication), and every citation the verification gate rejected or withheld, named with the reason. A memo carries the same record on its cover; a letter keeps it in the job file; the client report names the professional who reviewed it and the date. This is the documentation §10.36 asks firms to keep, produced by the system rather than reconstructed later.

Vendor vetting in one place

OPR asks firms to vet third-party AI tools before use. The facts a reviewer needs: the model provider is Anthropic, reached through its API under commercial terms that do not permit training on your data; the law-text index and embeddings run on Voyage AI and Upstash and receive queries only, never the return; the full list of subprocessors, encryption, retention and access controls is on the security page; a security summary for your written information security plan and our data-processing addendum are available on request at security@citedtax.com.

Source: IRS Office of Professional Responsibility, Alert 2026-19, Introductory Guidelines for Responsible AI Use in Federal Tax Practice (June 24, 2026), published through the IRS GovDelivery bulletin service. The duties are paraphrased; read the alert itself. This page describes CitedTax; it is not advice on any firm's obligations under Circular 230.