An M&A legal team should pause the affected AI workflow when a change could invalidate its previous approval for accuracy, confidentiality, permissions, or human review. Triggers include a model replacement, a material prompt revision, changed document retrieval, expanded access, altered data-handling terms, or removal of a review step. Freeze reliance on affected outputs, record what changed, and rerun tests against authoritative deal records before restarting. If the change cannot be identified or its impact bounded, treat that uncertainty as a pause trigger. Restart should require approval from the accountable legal workflow owner, with technical and confidentiality clearance where relevant, rather than an automatic release or a vendor's general assurance.
Preserve the approval that the change may invalidate
Start with the workflow already approved, not a fresh evaluation of every available AI product. For an illustrative contract-review task, the approval might permit preparing an internal draft schedule of assignment and change-of-control provisions. Record the permitted document set, output format, model configuration, prompt version, retrieval settings, confidentiality choices, and reviewing lawyer. Approval for that task is not approval to negotiate terms, send communications, or give client advice.
ABA Formal Opinion 512, issued in July 2024, explains that technological competence is not static and that independent verification depends on the tool and task. Lawyers remain responsible for their work. The opinion interprets ABA Model Rules; responsible qualified counsel should check the professional-conduct rules and ethics guidance applicable to the lawyers and matter. This article offers a general operating framework, not individual legal advice. [1]
Keep the last approved configuration and test results together. Without that baseline, a team cannot tell whether the changed workflow still meets the conditions supporting its previous approval.
Treat changes to inputs, behavior, and responsibility as material
A material change is one that could alter the evidence the system sees, the answer it produces, who can access information, or who checks the result. Do not limit the change register to model names. Record provider updates, model substitutions, fine-tuning, and changes to settings controlling how much source material the system can consider. If a provider changes an unspecified component, ask what changed and whether the affected configuration can be identified.
Prompt changes also deserve assessment. Replacing 'quote the clause and identify exceptions' with 'state whether consent is needed' changes the task from extraction toward a legal conclusion. Retrieval changes include adding documents, replacing an index, changing document parsing, or adjusting which passages are selected. NIST's July 2024 generative-AI profile recommends reassessing risks after fine-tuning or retrieval-augmented generation implementation and for uses not covered by initial testing. [2]
Workflow changes count too. Removing lawyer review, sending results into another matter, expanding user access, or changing retention and training-use terms can invalidate approval even when the model stays the same.
Decide whether to pause one task or the whole workflow
Use a documented triage decision. A formatting-only change may continue after the owner confirms that extraction, source references, access, and review remain unchanged. A change with a plausible effect on substantive output should pause reliance on that task until relevant checks pass. A cross-matter access problem, unknown data destination, or bypassed review control warrants broader suspension because the impact may extend beyond one output type.
Set concrete stopping conditions before accepting the change. In the illustrative clause-review workflow, stop if the system invents a quotation, omits a known material exception, retrieves a document from another matter, or produces an output that cannot be matched to its source. Stop as well if the required reviewer is unavailable. These are proposed operating criteria, not universal legal thresholds.
NIST's voluntary AI Risk Management Framework recommends safely ceasing deployment when risks are unacceptable until they can be sufficiently managed, and assigning responsibility for disengaging systems that depart from intended use. Pause the smallest scope whose boundary can be established reliably; do not assume an uncertain impact is confined. [3]
Open a change record and preserve a manual fallback
The workflow owner should open a record identifying the old and proposed configurations, the change reason, when it became effective, affected tasks, and affected matters. Distinguish a planned change from one discovered after deployment. Record the last known approved run and any uncertainty about the intervening period. A release announcement is evidence of a change, not proof that the team's own workflow remains acceptable.
Give the technical owner responsibility for configuration evidence and the supervising lawyer responsibility for substantive impact. Identify who can suspend access and who will notify affected internal users. Preserve relevant prompts, source identifiers, and outputs under the firm's applicable confidentiality and records policies; do not spread sensitive content into unrestricted incident notes.
Tell users what is paused and what work can continue manually. For contract review, the fallback is lawyer review of the underlying agreements and amendments, not silent reliance on cached AI summaries. NIST's generative-AI profile recommends deactivation communication plans addressing reasons, workarounds, access removal, and alternative processes. [2]
Rerun the benchmarks that the change can affect
Reuse the relevant tests that supported approval, then add cases aimed at the changed component. A retrieval change calls for checking whether the right agreement, amendment, and clause are returned. A prompt change calls for testing quotations, exceptions, and the distinction between extraction and legal interpretation. A model change may require broader regression testing across the workflow's approved tasks. A permission change needs access testing, not just better answers.
Use a counsel-reviewed reference set drawn from permitted records or deliberately created synthetic examples. For the illustrative task, include an agreement with an exception, a later amendment, and a document where the requested clause is absent. The source of truth is the governing document set, not the previous model's answer. A responsible lawyer should confirm expected results before comparing the changed outputs.
Score missing provisions, unsupported conclusions, wrong document versions, invented quotations, and failed source references separately. Do not let an improved average conceal a critical failure. Preserve the old acceptance conditions unless the accountable owner explicitly approves a justified revision. NIST recommends performance criteria tied to deployment approval. [2]
Recheck confidentiality and permissions independently of accuracy
A correct answer does not establish appropriate data handling. Compare the current terms and settings with those previously reviewed: who receives inputs and outputs, whether content is retained or used for training, which people can access it, and whether new processing parties or destinations are involved. Ask a knowledgeable technical or privacy reviewer to resolve uncertainty before confidential matter information returns to the changed workflow.
ABA Opinion 512 recommends understanding a tool's terms, privacy policy, and related contractual provisions, or consulting expertise that has analyzed them. Its confidentiality analysis is fact-specific. Responsible counsel should reassess any consent, communication, or authorization requirements under applicable jurisdictional rules and the matter's circumstances; neither an old engagement clause nor a generic 'private' label resolves every change. [1]
Keep initial retesting away from live confidential material where feasible. Synthetic records are often preferable; redaction should be checked because remaining details may still identify a transaction. Test that users outside the authorized matter cannot retrieve its documents. If the team cannot verify that boundary, do not upload more documents to discover whether it works.
Review work produced since the last known approved state
Pausing new runs does not resolve reliance on earlier outputs. Map the exposure period to draft schedules, issue lists, internal summaries, and any downstream materials that used them. If deployment timing is uncertain, expand the review window to cover the uncertainty. Prioritize work connected to imminent decisions or already shared, while keeping a record of items still awaiting review.
The reviewing lawyer should check affected assertions against the relevant agreement versions and amendments. Correct internal materials and decide whether prior recipients need an explanation or replacement. Any communication to a client or counterparty remains a human decision made under applicable professional duties and the engagement, not an automated incident response. ABA Opinion 512 emphasizes independent verification and retaining professional judgment. [1]
Do not presume that reverting a model repairs existing work. Restoration is a configuration decision; correcting prior reliance is a separate legal-work decision. Track which outputs were checked, which were withdrawn, and which remain unresolved so the restart decision does not hide unfinished remediation.
Require a bounded trial and accountable restart approval
After relevant tests pass, run a limited trial on the approved task with the proposed configuration fixed. Use permitted records and a named reviewing lawyer who checks every trial output against the underlying sources. Preserve the same stopping conditions used in reassessment. A critical omission, source mismatch, access failure, or missing review should stop the trial and return the issue to its owner.
The restart record should identify the configuration, test results, remaining limitations, corrected work, permitted task scope, and required review. The legal workflow owner approves substantive restart. The technical or information-security owner clears affected configuration and access issues, and qualified counsel resolves outstanding ethics or confidentiality questions. A small team may combine roles, but should still record each responsibility and decision.
Approval should say what may resume and what remains prohibited. Passing clause extraction tests does not authorize autonomous negotiation, disclosure, outreach, or approvals. NIST recommends clear deactivation responsibilities and escalation procedures; the named restart decision here is a practical counterpart to those controls, not a NIST certification. [2]
Make the new approval visible to everyday users
Update the permitted-use instructions, source set, required review, and known limitations before ordinary use resumes. Give users a way to flag unexplained behavior changes, missing sources, and newly unavailable documents. Assign someone to review vendor change notices and revisit the workflow when matter scope or document access changes. NIST's framework calls for ongoing monitoring of third-party resources and post-deployment risk treatments. [3]
Keep change-triggered checks alongside regular monitoring. The practical outcome is a repeatable decision: identify the change, pause the affected use, retest against authoritative records, review exposed work, and obtain accountable human approval. These are recommended implementation controls, not assertions about features already supplied by any platform.
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. The software is user-operated and has no success fee. Lawyers remain responsible for legal judgment, permissions, confidentiality decisions, and approval of work; Acquivo does not represent clients or provide legal advice.
Key Points
- Tie continued approval to a defined task, configuration, permitted source set, and accountable human reviewer.
- Assess model, prompt, retrieval, access, data-handling, and review changes, not just new model names.
- Pause reliance when a critical control fails or the change's impact cannot be bounded; preserve a manual fallback.
- Rerun affected benchmarks against authoritative documents and test confidentiality permissions separately from answer accuracy.
- Review outputs produced during the exposure period and require documented human restart approval after a bounded trial.
Sources and references
Source retrieval dates record research access, not article publication. Read our editorial and date policy.
- [1] ABA Formal Opinion 512: Generative Artificial Intelligence Tools, July 29, 2024 · Accessed
- [2] NIST AI 600-1: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, July 2024 · Accessed
- [3] NIST AI 100-1: Artificial Intelligence Risk Management Framework, AI RMF 1.0, January 2023 · Accessed
Choose your next step
Acquivo is user-operated software. Owners and their professionals remain responsible for decisions, communications, and professional review.
Educational information, not individual legal, tax, or investment advice. Consult qualified professionals for your circumstances. Acquivo provides software, not representation or negotiation on your behalf.