How Should M&A Attorneys Set AI Data Retention and Deletion Requirements?

Approve the task and accountable owners before uploading

Start with a bounded task, such as extracting assignment clauses from selected contracts for an attorney's review. Specify which documents may enter the tool, which outputs may be retained, and where approved work belongs. Keep the signed contracts in the controlled matter repository as the source of truth. The AI summary is a working aid, not a replacement for the contract or the attorney's analysis.

Assign the matter attorney responsibility for permitted use, client communication, confidentiality choices, and legal exceptions. Give the records owner responsibility for the retention inventory and closure checklist. The security or IT owner verifies permissions, settings, and deletion behavior. Procurement should obtain the applicable agreement and vendor responses. These roles can overlap in a small practice, but the written plan should identify who makes each decision.

ABA Formal Opinion 512, issued in July 2024 under the ABA Model Rules, calls for tool-specific competence, confidentiality assessment, and appropriate independent review. It also addresses supervision and vendor retention after services end. Its guidance is not a substitute for the rules governing a particular lawyer. Counsel should review those rules before enabling live matter data. [1]

Inventory the full information path, not just chat history

Trace an illustrative contract-review request from the user's device through the application and any model service. Ask whether document parsing creates a temporary copy, whether retrieval creates a search index, and whether application logs capture excerpts. Include generated summaries, revision history, downloaded exports, shared links, support attachments, and recovery copies wherever those features exist. Do not assume every tool creates each category; require a specific answer for the actual workflow.

For every location, record the data category, controlling organization, authorized users, processing region, relevant subprocessor, and deletion owner. Give derived records enough detail to be found later: a matter identifier and file identifier may be more useful than a list of filenames alone. Search indexes can include embeddings, which are numerical representations used for retrieval. Treat their retention as a separate question rather than assuming deletion of the original file also removes them.

Official OpenAI API documentation distinguishes abuse-monitoring logs from application state retained to support features. It lists different retention behavior for different endpoints, including files and vector stores. This illustrates why the inventory must follow the selected feature and configuration rather than the vendor's brand name. It does not establish how another application, or Acquivo, handles data. [2]

Specify retention triggers and contractual commitments

A usable requirement says when the clock starts, what happens when it expires, and who verifies completion. Possible triggers include completion of the approved task, deletion of an uploaded document, matter closure, or termination of the service. Choose triggers appropriate to each record category. A temporary processing copy and an approved diligence memorandum need not follow the same schedule.

Ask the vendor to identify retention periods for customer content, metadata, security logs, support records, and backups. Clarify whether a period is a maximum, a default setting, or a customer-configurable option. Request the exceptions, the process for changing settings, and the effect of account termination. Document which contract controls if a policy page and signed agreement differ. Resolve uncertainty before relying on a deletion commitment.

OpenAI's API documentation, as retrieved in November 2026, says API data is not used for training or model improvement unless customers opt in. Separately, it describes default abuse-log retention of up to 30 days, with stated exceptions. Some features may store application state even when Zero Data Retention is enabled. These are provider-specific terms, not a recommended retention period or a universal definition of zero retention. [2]

Separate training and reuse permissions from storage

Request explicit answers about model training, fine-tuning, product improvement, evaluation, human review, and sharing with subprocessors. Determine whether those uses apply to inputs, outputs, feedback submissions, or support interactions. Ask who can opt in and whether a user-level choice can override an approved organization setting. A prohibition on training is useful only within its stated scope; it does not answer every storage or access question.

For the proposed workflow, limit uploads to information the matter attorney has approved and restrict access to the authorized matter team. Consider synthetic material first, then carefully minimized excerpts if live use is authorized. Removing names alone should not be the approval test: counsel should consider whether transaction details still reveal confidential information. Do not assume private deployment, a privilege label, or a retention setting settles privilege or permissible disclosure.

Formal Opinion 512 advises lawyers to understand terms, privacy policies, and related agreements, or consult someone who has analyzed them. Where informed consent is required under the applicable framework, generic engagement-letter boilerplate may be insufficient. The attorney should document the actual permission and client-communication analysis, rather than treating acceptance of vendor terms as client authorization. [1]

Reconcile deletion requests with scoped retention exceptions

Maintain an exception register separate from the routine retention schedule. For a preservation hold or other retention obligation, counsel should identify the basis, covered records, authorized access, responsible reviewer, and release condition. Technical staff should not decide whether a client request overrides an obligation, and an AI tool should not release a hold or approve destruction.

Distinguish the authoritative matter file from unnecessary processing copies. ABA Model Rule 1.16(d) addresses protecting client interests at termination, including surrendering papers and property to which the client is entitled. It does not prescribe a universal file-retention period or decide whether every AI prompt must be retained. Governing local rules, applicable law, and the circumstances require counsel's assessment before relevant material is destroyed. [4]

For personal data within UK GDPR scope, ICO guidance explains that erasure is conditional and identifies exceptions for necessary processing, including legal obligations and legal claims. This is not a blanket permission to retain all transaction data. The ICO page also states that its guidance is under review following the Data (Use and Access) Act, so qualified counsel should check the applicable position when deciding a request. [3]

Define deletion evidence and distinguish pending copies

Require a deletion record that identifies the requested scope, authorization, affected objects, action taken, result, and unresolved categories. Keep verification metadata proportionate: reproducing the sensitive prompt in the deletion ticket can create another copy. A matter identifier, object identifiers, timestamps, and a reference to the approved exception may provide a clearer record without duplicating the content.

Distinguish three conclusions: the team observed removal from accessible live systems; the vendor confirmed removal from specified downstream systems; or copies remain pending under a documented schedule or exception. A screen showing no chat history proves only what that screen exposes. Ask how deletion covers associated indexes, retained outputs, support copies, subprocessors, and backups, and whether a vendor confirmation excludes any of them.

The ICO's UK guidance says a valid erasure request with no exemption requires steps addressing backups as well as live systems. It recognizes that backup data may remain until overwritten, with the key issue being to put it beyond use and avoid other processing, subject to the circumstances. [3] Accordingly, ask for the expiry schedule and restore procedure. A recommended control is to reapply recorded deletions before restored data becomes available for normal use, rather than silently reintroducing it.

Run a synthetic-data trial with explicit stopping conditions

Use an illustrative contract containing unique test phrases and a known assignment clause. The attorney compares the AI extraction with the approved test document, checking omitted qualifications, incorrect quotations, and invented references. IT records the tested account, project, features, and settings. Use only approved test accounts when checking that an unauthorized matter team cannot retrieve the test content.

Delete the uploaded test record and check the relevant history, search results, stored outputs, exports, and available administrative records against the inventory. Where behavior cannot be observed, obtain scoped vendor evidence instead of reporting a successful test as proof of backend erasure. OpenAI's documentation allows project-level retention choices to differ from organization settings, reinforcing the need to verify the actual project rather than rely on a general account label. [2]

Stop the trial or withhold live-data rollout if retained locations cannot be explained, reuse permissions are unresolved, access crosses approved matter boundaries, deleted content remains retrievable, or the exception process is unclear. Resume only after the responsible human records the correction and verifies the affected control. A failed summary check should also stop reliance on that task until corrected. The trial evaluates a defined workflow, not overall compliance or security.

Make closure and change review part of the requirements

At matter closure, the attorney decides what belongs in the matter file and confirms whether exceptions remain. The records owner applies the approved schedule to processing copies. IT revokes unnecessary access and verifies configured deletion actions. Keep pending backup expiry or vendor confirmation open until the assigned owner records its resolution. Closing a transaction is not evidence that all associated AI data has disappeared.

Recheck requirements when the team changes a model, enables a new feature, grants broader access, or accepts revised terms. Preserve the version of the requirements and evidence supporting each approval. The useful deliverable is a documented retention and deletion plan with named verification responsibilities, not a general assurance that the tool is safe. This framework supports professional review; it is not individual legal advice.

Acquivo is a software platform that delivers the entire sell-side M&A process, valuation, marketing materials, buyer matching, and deal management, on one platform for business owners, M&A attorneys, and boutique investment banks.

Acquivo is user-operated software with no success fee. The workflow above is a requirements framework, not a claim that Acquivo supplies particular retention controls or integrations. Users and their qualified advisers retain responsibility for legal judgments, disclosures, and approvals; software does not represent clients or guarantee privilege, security, or compliance.