|
EXPLAINER |
AI FOR ACCOUNTANTS · PART 15 OF 16
Where client data goes when an AI assistant answers your question
Zero-retention clauses sound protective, but the subprocessor list is where the real risk often lives.
When a practitioner pastes a trial balance or a client's bank reconciliation into an AI assistant and hits send, that text travels somewhere before an answer comes back. Most accountants have a rough sense that this happens, but the mechanism matters more than the feeling. The data moves to a server, gets processed by a model, and - depending on which product and which plan - may or may not be stored, logged, or used to train future model versions. The clause that governs this is called a zero-retention or no-training agreement, and not every plan includes one by default.
A zero-retention clause is a contractual promise that the vendor will not store your inputs after the session ends and will not feed them into training data. On paper, that handles the obvious concern: the client's payroll figures or unreleased acquisition detail do not end up in a model that someone else queries next month. In practice, two things complicate the picture. First, zero-retention terms usually apply to the primary vendor, but most AI products rely on subprocessors - third-party infrastructure or model providers sitting behind the interface. The clause covers what the vendor does; it may say less about what the subprocessor retains, where that subprocessor operates, or which law governs a breach. Second, zero-retention is often a feature of enterprise or API tiers, not the consumer-facing product a staff accountant might open in a browser on a Tuesday afternoon.
The data region question is a separate layer. When a practitioner in one jurisdiction uploads a workpaper containing a client's entity structure, the processing may route through servers in another jurisdiction entirely. That matters for engagements where the client's data residency requirements, privacy law, or their own contracts restrict cross-border transfers. GDPR is the example most accountants have heard of, but sector-specific rules and client-side data processing agreements can impose the same kind of constraint. A vendor's terms of service may specify a default region and permit the vendor to change it on notice - language easy to miss on the first read.
The practical difference between a consumer tier and an enterprise tier shows up in exactly these details. A sole practitioner using a free or low-cost plan to draft a memo is typically on terms that permit the vendor to use that session for model improvement. An enterprise agreement with a data processing addendum, a list of approved subprocessors, and a named processing region is a different document with different obligations. Both might run on identical underlying technology, but only one gives the practitioner anything to show an engagement partner, a client's legal team, or a regulator asking about data handling.
What to do is not complicated, but it requires a deliberate step before the first use, not after. Pull the vendor's data processing agreement and subprocessor list - both are usually published and linked from the privacy policy. Check whether zero-retention or no-training terms require a specific plan or a signed addendum. Note the default processing region and whether the vendor reserves the right to change it. Then match that against the sensitivity of the data you intend to use: a public-facing industry schedule carries different risk than a client's draft tax position. For the latter, if the terms do not clearly cover it, the conservative answer is to anonymize the data before it leaves your system or use a different tool that does provide the contractual coverage.
WORKED EXAMPLE
In practice
A practitioner is drafting a response to a client query about intercompany eliminations. The supporting workpaper contains the client's entity names, transaction amounts, and a draft footnote. Before pasting anything, they want to know what the assistant will do with it.
What came back. The assistant identified that zero-retention was listed under the enterprise tier only, that two named subprocessors operated under US Standard Contractual Clauses with EU processing permitted, and that one clause allowed region changes on 30 days notice rather than requiring consent. It flagged that ambiguity as a gap. One item needed correcting: it initially read a defined term as applying to all plans when the definition was scoped to Section 4 of the DPA, which covered enterprise customers only.
How it was checked. The practitioner opened the DPA PDF, located Section 4, and confirmed the defined term's scope matched the assistant's corrected reading before noting the gap in the engagement file.
A constructed example. The prompt is usable as written; the figures show the shape of a result, not a measured one.
WHEN TO USE IT
| WHEN NOT TO
|
WHAT TO TAKE FROM THIS
| Zero-retention clauses often require an enterprise plan or signed addendum - check before using client data. | |
| Subprocessor lists, not just the primary vendor's terms, determine where your data actually goes. | |
| Anonymize sensitive client detail if the processing region or retention terms are unclear. |
SPONSORED
QUESTIONS THIS ANSWERS
What is a zero-retention clause in an AI vendor agreement?
A contractual promise that the vendor will not store your inputs after the session ends and will not use them to train future model versions. It typically applies to the primary vendor, not necessarily to subprocessors.
Does the free version of an AI assistant have the same data protections as the paid version?
Usually not. Zero-retention and no-training terms are often reserved for enterprise or API tiers. Consumer-facing plans may permit the vendor to use session data for model improvement.
Why does data region matter for accounting work?
Processing a workpaper through servers in another jurisdiction can conflict with client data residency requirements, privacy law, or the client's own data processing agreements. The vendor's default region and their right to change it should be checked before use.
SOURCES
Where this comes from
What the accounting job market is actually asking for.
GO DEEPER
Go deeper
IN THIS SERIES
Previously: Why you probably do not need to fine-tune a model
Next: Prompt injection, the new phishing for AI workflows (coming)
An explainer, not a study: it carries no statistics on purpose. Examples are illustrative.
How accountants are using AI, automation and smarter workflows to close faster, audit cleaner, and free up time for real work.
Accounting Stack · Audit Friendly Data · Accounting & Finance Jobs
Audit Friendly · modernaccounting.ai