Sixty-five years ago a federal court put an accountant inside the attorney-client privilege, because accounting is a foreign language to lawyers.
For a lawyer's supervised AI agent, that door is already open.

An accountant employed by a tax law firm was jailed for contempt after refusing to answer grand-jury questions about a firm client. Judge Friendly, writing for the court, held him inside the privilege:

The privilege extends, the court held, where the third party's presence is "necessary, or at least highly useful, for the effective consultation between the client and the lawyer." And the limit sits on the same page: where what is sought is only accounting service — where the advice sought is the accountant's rather than the lawyer's — no privilege exists.

From Kovel and its progeny, four operating conditions. The engagement runs from counsel, not the client. Direction and control stay with counsel, documented. The purpose is making the legal consultation effective — nothing else. And the work is held in confidence, segregated from all non-legal work. An AI agent has no independent interests to serve; the honest pressure on the analogy comes from Kovel's own circuit, which later confined the doctrine to translators — we take that up on the deep page rather than paper over it.

United States v. Heppner, No. 25 Cr. 503 (S.D.N.Y. Feb. 17, 2026) (Rakoff, J.), did not hold that using AI waives privilege.

It denied a pro se defendant's privilege claim over his solo use of a consumer AI platform, on three traditional grounds: no attorney was involved; the platform's data-retention and training terms defeated any reasonable expectation of confidentiality; and no legal advice from counsel was being sought. The work-product claim failed too — the material was created on the defendant's own volition, not at any attorney's direction.

Read as a checklist, each failure inverts into a requirement: counsel engagement, documented. Contractual zero-retention, no-training confidentiality terms. Use directed at obtaining legal advice through counsel. Attorney direction, recorded as it happens. A counsel-engaged, zero-retention, audit-logged supervised agent answers all three grounds — which is precisely the structure Kovel built for accountants.

Two things we say plainly. No court has squarely held that a supervised AI agent receives Kovel treatment; Heppner is negative-space guidance — it tells you what failure looks like, not what success is guaranteed to look like. And the memorandum itself has not yet been retrieved; until we have read it, this site describes the ruling only through the published analysis it cites.

That is incorporation by reference. Whatever the common law of attorney-client privilege comes to hold about a lawyer's supervised AI agent flows into §7525 automatically, by the statute's own text. No separate ask on the accounting side; no rule change; no petition.

And the perforations are severe, so we state them bluntly. The privilege is assertable only in noncriminal tax matters before the IRS and noncriminal tax proceedings in federal court brought by or against the United States — no criminal matters, no state courts or agencies, no private litigation, no non-tax federal proceedings. No written communication promoting participation in a tax shelter is covered. And it never reaches return preparation: the Seventh Circuit in United States v. Frederick, 182 F.3d 496, 500–01 (7th Cir. 1999), held return-prep information unprivileged and stripped dual-purpose documents outright — which is why the supervised agent instance that facilitates legal advice must never also prepare returns.

By certifying the public reports that depict a corporation's financial status, the Court held, the independent auditor assumes a public responsibility transcending any employment relationship with the client — and so audit workpapers get no work-product immunity, and no accountant-client privilege exists under federal law at all.

The audit function is defined by non-confidentiality: the auditor's loyalty runs to the investing public, not the client. No privilege campaign belongs there, for anyone, human or AI. Everything on this site concerns the advice side only — tax advice, controversy representation, litigation support. Never the attest side.

The Kovel engagement letter — a template implementing the counsel-engaged structure for a supervised AI agent: retention by counsel, purpose recited in Kovel's register, return prep and audit expressly excluded, contractual zero-retention and no-training terms, documented direction, a tamper-evident audit trail, one dedicated instance per matter. Read it at tax.org.ai/kovel-letter; the adaptable text itself is at /kovel-letter.md.

The confidentiality compliance memo — an analysis template for a CPA firm's professional-standards function: AICPA Rule 1.700.001 and its service-provider interpretation, the two compliance paths, consent mechanics, the honest privilege posture, and the audit wall. Read it at tax.org.ai/confidentiality-memo; the text itself is at /confidentiality-memo.md.

Both artifacts carry the same notice, and it applies here too: this is educational material from a Foundation program, not legal advice; it forms no attorney-client or accountant-client relationship; no court has squarely blessed AI-agent Kovel status; have your counsel adapt every clause to your engagement and jurisdiction.

The list, in one paragraph.

26 U.S.C. §7525 and United States v. Arthur Young & Co., verbatim from law.cornell.edu. United States v. Ackert and United States v. Frederick, full text with pin cites from law.resource.org. AICPA Rule 1.700.001 as published by the Journal of Accountancy. State law text-verified in three states only — Colorado's C.R.S. §13-90-107(1)(f), Missouri's RSMo §326.322, and Idaho Rule of Evidence 515 — and no other state is cited anywhere on this site. Kovel's own mirrors refused our fetch, so its quoted passages are cross-verified through full-text secondary sources; and the Heppner memorandum has not yet been retrieved, so the ruling is described only through the published analysis of counsel who read it. Until we have read a document, we do not cite it as if we had.