Your loan officer’s AI chat history is Exhibit A.

“I told a borrower he needed to sell all his grain by the end of the week or we would foreclose on his operation. Could that statement give rise to lender liability?”

Bankers hear that AI conversations are discoverable and file the warning under legal trivia. Do not. A loan officer can create the borrower’s next exhibit simply by asking whether he has already gone too far.

Picture a farm borrower with a troubled credit and grain in storage. The officer demands a sale by Friday and threatens foreclosure. After the call, he wonders whether the ultimatum crossed a line. Before calling counsel, he types the question above.

The prompt records the officer’s own account: the demand, the deadline, and the threatened consequence. If the borrower later claims the bank forced a damaging sale, his attorney can compare that account with the call notes, grain-sale records, and the officer’s testimony. The bank has supplied a starting point in its own words.

The question may be more damaging than the answer. The officer does not have to ask AI to hide anything. A candid request for legal guidance can preserve facts the borrower would otherwise have to establish through documents and testimony.

The ultimatum does not establish liability by itself. The loan documents, default status, required notices, and the bank’s conduct matter. Where Farmer-Lender Mediation applies, the timing of enforcement matters too. Those are questions for counsel before the threat, not an AI reassurance afterward.

Suppose the chat continues: “He says selling now will lock in a loss, but I need the principal paid down this month. Does that change your answer?” The officer has added the borrower’s objection and his own objective to the record. The more context he gives the chatbot, the more context the other side may obtain.

Protect the bank’s record

An ordinary business chat does not become privileged because the officer asks a legal question. A chat app is not bank counsel. Nor does an enterprise subscription, confidentiality setting, or “privileged” label settle the issue. Legal protection depends on the work and the circumstances in which it was created.

A discovery request can reach prompts, outputs, uploaded loan documents, exported conversations, and AI-created drafts copied into the file. Deleting the visible conversation may leave screenshots, downloads, audit data, and emails untouched. The record often lives in more than one place.

Do not let cleanup become the case

When litigation is reasonably anticipated, preserve relevant AI material along with the loan file. If records are lost because reasonable preservation steps were not taken, the court can impose corrective measures. Destruction intended to deprive the borrower of evidence can support adverse inferences or case-ending sanctions.

Stop deletion, preserve the account and device, identify where the output went, and involve counsel. Preservation does not concede that everything must be produced. It keeps that decision from being made for the bank by a missing record.

Do not fix a damaging chat by starting a second one: “Rewrite this so it cannot be used against us.” That instruction may become the next exhibit.

Use the borrower’s record

Consider the borrower who forms a new LLC, moves equipment and receivables into it, and leaves the debt behind six weeks before default. The paperwork calls it a routine reorganization.

He may have asked AI how to move those assets without making it look like he was avoiding the bank. In a fraudulent-transfer claim, intent is often assembled from timing, insider control, and the surrounding facts. A prompt may put that intent in plain English.

The chat does not prove the entire claim. It can explain why the transaction was structured that way and connect the new entity, moved equipment, and redirected payments into one chronology.

Ask your litigator to pursue the relevant AI histories. Identify the platforms and accounts, preserve uploads and outputs, and follow the drafts into emails and transaction documents. Expect the borrower to make the same requests of the bank.

The record starts now. The lawsuit comes later.

The credit has not reached a formal workout. An officer is weighing an advance request, a collateral release, or a tougher repayment demand. Calling counsel feels premature. Asking AI feels like a quick way to think it through.

“Can we freeze further advances until he pays down the line?” “What if we require him to sell equipment first?” “How should I document the call if he says we pressured him?” These are legal-risk questions arising inside routine credit management.

To get a useful answer, the officer supplies context: what the bank knew, what the borrower said, what the officer wants to accomplish. Follow-up prompts add options, doubts, and proposed language. The conversation becomes a parallel account of the credit decision.

An abandoned idea is not a bank decision. But a later claim may put both the idea and the actual decision under scrutiny. The borrower’s attorney can ask what the bank considered, what it rejected, and whether the final file memo tells the same story.

A confident answer can make a premature decision feel safe. If AI misses a restriction in the loan documents or a required notice, the officer may act on the reassurance. The bank then has both the action to defend and a chat explaining how it got there.

Documentation creates a second risk. AI may turn “the borrower discussed a sale” into “the borrower agreed to sell.” Once that polished sentence enters a call summary or credit memo, an unsupported inference starts to look like an established fact.

Bring counsel in earlier

Make the referral trigger the substance of the question. “Can we do this?” “Could this expose the bank?” “What should we say about what happened?” Those questions belong with counsel before the bank acts or revises its account of events.

Keep factual loan records accurate and complete. Check AI-assisted summaries against their sources. If a chat already exists, preserve it and tell counsel. Calling the attorney later does not retroactively protect an earlier business chat.

Ask counsel before you ask AI.

Federal Rule of Civil Procedure 37(e) addresses lost electronically stored information that should have been preserved and cannot be restored or replaced. Measures under subsection (e)(1) require prejudice; adverse-inference and case-ending sanctions under subsection (e)(2) require intent to deprive another party of the information’s use. The rule does not make every deletion sanctionable.

In United States v. Heppner, No. 25 Cr. 503 (JSR) (S.D.N.Y. Feb. 17, 2026), Judge Rakoff rejected attorney-client privilege and work-product protection for the defendant’s self-directed exchanges with a public AI platform. The court considered the absence of counsel’s direction, the platform’s confidentiality terms, and whether the documents reflected counsel’s strategy. The ruling illustrates the risk discussed here; it does not decide protection for every lawyer-directed or enterprise AI workflow.

These source notes accompany the web edition and were not part of the original newsletter layout.

Matthew Bialick represents financial institutions in banking, creditors’ rights and commercial litigation. Contact him at 952-239-3095 or matthew@mjblawmn.com.

Related reading: Banking and creditors’ rightsPreparing for a commercial disputeAgricultural finance