Claude in Tax Practice: Client Confidentiality, Verification and Anthropic's Usage Policy

Claude, the AI assistant made by Anthropic, is now part of the working day in many tax offices: summarising judgments, drafting replies to notices, explaining a provision, checking a working. Two sets of rules apply when a chartered accountant uses it on client work. The first is the profession’s own, in the Chartered Accountants Act, 1949. The second is Anthropic’s: its Consumer Terms, its Commercial Terms and its Usage Policy, which deals with tax advice by name. The following sets out what each says, and the dos and don’ts that follow from them.

The profession’s duty of confidence

Clause (1) of Part I of the Second Schedule to the Chartered Accountants Act, 1949 treats a chartered accountant in practice as guilty of professional misconduct if he:

discloses information acquired in the course of his professional engagement to any person other than his client, without the consent of his client or otherwise than as required by any law for the time being in force;

The clause predates AI tools by decades, and no ruling or ICAI pronouncement on AI services was among the sources reviewed for this article. Whether typing a client’s figures into a third-party service is a “disclosure” in this sense is therefore open, not settled. The clause itself leaves only two routes for client information to leave the engagement: the client’s consent, or a law requiring it. A practice that wants to stay clear of the question either removes the identifying details before anything is entered, or takes the client’s consent and uses a plan whose terms treat the content as confidential.

Clause (7) of the same Part, which covers a member who “is grossly negligent in the conduct of his professional duties”, is the other provision in point. The duty of care attaches to the member, whichever tool helped produce the work.

Individual plans and client data

Anthropic’s Consumer Terms govern Claude.ai, Claude Pro and other products for individuals. Under them, Anthropic may use the material in a chat to train its models unless the user opts out of training through account settings. Opting out does not cover everything. Two kinds of material are still used for training: feedback the user gives, and material flagged for safety review. Rating an answer with the thumbs up or thumbs down icon stores the related conversation as feedback.

Anthropic’s announcement of the updated consumer terms, which apply to the Free, Pro and Max plans, gives the retention periods. Where the user allows training, data is kept for five years; where not, the existing 30-day retention period continues. A deleted conversation is not used for future model training.

On an individual plan, then, the model-training setting in Privacy Settings is the first thing to check, and the thumbs icons are best left alone on any chat that holds client material.

Business plans and the API

The Commercial Terms govern Anthropic API keys and other offerings that refer to them, and the position on training is stated in one line:

Anthropic may not train models on Customer Content from Services.

Customer Content is treated as the customer’s Confidential Information, which Anthropic may use only to exercise its rights and perform its obligations under those Terms. Anthropic’s announcement lists Claude for Work, “which includes our Team and Enterprise plans”, together with the API, among the services under the Commercial Terms.

A practice that intends to run client documents through Claude as a matter of routine should settle which of the two sets of terms it is working under before anything else.

Anthropic’s rule for tax advice

Anthropic’s Usage Policy contains a section titled High-risk Use Case Requirements. Its list of high-risk areas includes, in finance:

advising an individual on their own taxes, such as what to report, deduct, or claim; or preparing a tax return for them to sign or file.

The wording speaks of an individual. It does not, in terms, reach advice to a company or a firm, nor a practitioner’s own research that never becomes advice; the review discipline below costs nothing to apply there as well.

Where Claude is used to produce such a recommendation, two controls apply. The first is review. A qualified person must meaningfully review the recommendation, with the authority to change it, before it is delivered as advice. The policy describes a qualified person as someone with the training or experience to evaluate the output, holding a licence where the law requires one, and states that this person “remains responsible for the accuracy and appropriateness of what is delivered or decided”. The second is disclosure. The individual who receives the advice must be clearly told that AI was used to produce it, although Anthropic, Claude and the model need not be named.

Anthropic has published a revised Usage Policy that takes effect on 12 November. Its announcement states that the high-risk requirements have not changed; the section has been rewritten to list which recommendations are covered and which are not.

The review requirement is what the profession expects anyway. The disclosure requirement is the one few practices have yet written into their advice notes or engagement letters, and it is worth adding a line for it now.

Outputs are a draft, not an authority

The Consumer Terms are direct about accuracy. Outputs “may contain material inaccuracies even if they appear accurate because of their level of detail or specificity”, may not reflect correct, current or complete information, and are not to be relied on without independent confirmation.

In tax work the risk sits in exactly the details that look most authoritative: a notification number, the date of a circular, a section reference, a case citation. A notification issued last month may be unknown to the model, and an answer about it can be fluent and wrong. Every such reference in an AI-assisted draft has to be traced to the notification, circular, judgment or bare Act before the draft leaves the office.

Where it earns its place

The work that gains most is work where the practitioner supplies the source and Claude does the reading or the first draft:

• summarising a long judgment or order that has been uploaded, with the paragraph numbers behind each point;
• comparing two versions of a provision, or a draft reply against the show cause notice it answers;
• a first version of a reply, a client letter or an internal note, from facts the practitioner states;
• checking a working, a reconciliation or a spreadsheet formula.

Answers drawn from an uploaded document are far easier to check than answers from the model’s memory, and asking for the paragraph relied on turns the check into a matter of finding that paragraph.

Dos and don’ts

Do:

• check the model-training setting on an individual plan, or use a plan under the Commercial Terms for client work;
• replace names, PAN, GSTIN, Aadhaar and bank details with placeholders before uploading;
• upload the notification, circular or order, rather than asking about it from memory;
• keep the reviewed version, with the sources checked, on the client file;
• tell the client that AI was used, where the advice concerns the client’s own taxes.

Don’t:

• rate a chat holding client material with the thumbs icons;
• cite a case, circular or notification because Claude named it, without reading it;
• let advice or a return go out on the output alone, without a qualified person’s review.

This article is intended for educational purposes only and should not be construed as professional advice. It is based on the law as understood on the date of publication; laws and regulations are subject to change, and it may not reflect later developments. Readers are advised to consult a qualified tax professional on their specific situation before taking any action based on it. CA. Mohit Gaba takes no responsibility for any action taken or decision made on the basis of this article.