Bankers should treat an AI-assisted milestone tracker as a review aid, not as the approved sale-process plan. Compare each proposed date with the controlling record, label it planned, confirmed, overdue, or blocked, and record the source and reason for every material change. Assign one human process owner to each milestone, test dependencies and time zones, and keep commitments, approvals, bidder communications, and escalations with accountable people. A bounded pilot should stop when the tool cannot show provenance, produces unexplained changes, exposes confidential material, or creates false confidence. FINRA says firms using generative AI remain subject to existing obligations and should address governance, data integrity, reliability, and accuracy in their controls. [2] Exact legal and supervisory implications depend on the firm, role, use case, and jurisdiction.
Start with the approved process plan
Begin with the record that the deal team already treats as authoritative: the approved sale-process plan, engagement instructions, internal project plan, or dated decision log. Do not ask the model to decide which document controls. The reviewer should identify the version, effective date, owner, and scope before comparing any generated tracker rows.
Define the milestone fields before extraction. A useful row includes the milestone name, target date and time zone, status, accountable owner, predecessor or dependency, source reference, last confirmed date, and notes about assumptions. Keep planned dates separate from confirmed dates. A model can summarize a discussion that mentions a target without converting that target into a commitment.
This separation prevents a common error: treating an email, meeting note, or draft calendar invitation as equivalent to the approved plan. If records conflict, preserve both references and route the conflict to the process owner. The tracker should show an unresolved discrepancy rather than silently selecting the more recent or more plausible date.
Use four status states with explicit tests
Use four practical states. Planned means the date appears in an approved plan but has not been confirmed by the responsible owner. Confirmed means the owner has verified the date and the prerequisite conditions are understood. Overdue means the confirmed or approved target has passed without a recorded completion or approved change. Blocked means progress depends on an unresolved condition, such as missing information, a decision, a document, or an external response.
Write the test for each state in plain language so another reviewer can reproduce it. A confirmed management presentation date requires the date in the plan, an identified owner, and a confirmation recorded in the chosen source. A blocked diligence request requires the open dependency and the person responsible for resolving it. Do not use a single confidence score as a substitute for these states.
A status change should be an event, not just an overwritten cell. Record the old state, new state, timestamp, reviewer, source, and reason. This lets the deal team distinguish a legitimate schedule revision from an extraction error. NIST describes AI risk management through governance, mapping, measurement, and management, which supports assigning responsibility and documenting how outputs are evaluated. [1]
Require provenance for every material date
For every material date, capture a source that a human can reopen. The reference might be a page and section in the approved process plan, a dated internal decision record, or a meeting note with its author and date. A citation such as ‘latest email’ is not enough. If the source is a file, preserve its version or stable internal identifier under the firm’s records practices.
Ask the reviewer three questions: What exact wording supports the date? Who supplied or approved it? What changed since the previous tracker version? If the answer is unclear, label the row planned or unresolved rather than confirmed. A generated explanation can help locate evidence, but it is not evidence by itself.
Where dates come from documents with different time zones or business-day conventions, normalize the display without erasing the original value. Show the source time zone and conversion rule used by the team. If a deadline depends on receipt, approval, or another event, record that dependency instead of presenting a fixed date as unconditional.
Check dependencies and ownership, not just dates
A tracker is useful only when it explains what must happen next and who can act. For each milestone, identify the accountable process owner, contributor, and predecessor conditions. Avoid assigning ownership to a generic group when one person must confirm the state. The owner can delegate work, but the tracker should still show who is accountable for review.
Review the dependency chain from both directions. Ask what must be complete before the milestone and what downstream work relies on it. Look for circular dependencies, missing owners, impossible ordering, and dates copied forward after a prerequisite moved. AI can flag likely relationships, but the deal team should accept, reject, or amend each material dependency.
Keep external actions outside automated execution. The tracker may identify that a buyer list, data-room item, management call, or counsel review needs attention. It should not send outreach, change an approved process, negotiate a deadline, approve a disclosure, or represent that a bidder has committed. Human process owners make those decisions and record the result.
Design a source-checking review loop
A workable review loop has five passes. First, an authorized user supplies only records permitted for the use case. Second, the tool extracts candidate milestones without changing source files. Third, a reviewer checks each candidate against the source and assigns a state. Fourth, the owner confirms material dates, dependencies, and blockers. Fifth, the team publishes the reviewed tracker version and retains the change log.
Use a small, representative pilot before relying on the workflow across a process. Include ordinary milestones, revised dates, conflicting notes, missing owners, and at least one blocked dependency. Compare the output with a human-created sample. The pilot is not evidence that the system will perform reliably in every transaction or document set.
FINRA advises member firms to evaluate generative AI tools before deployment and notes that policies should address technology governance, model risk management, data privacy and integrity, and reliability and accuracy when the tool is part of a supervisory system. [2] Even when a boutique bank uses a tool for internal coordination rather than supervision, these are useful control questions.
Set stopping criteria before the pilot
Decide in advance when to pause or stop the workflow. Stop if the tool cannot provide a source for a material date, repeatedly merges separate milestones, changes dates without an identifiable source, or labels a dependency complete when the record says otherwise. Stop if the reviewer cannot determine which documents were used or if the output cannot be removed from an unauthorized workspace.
Also stop when confidentiality controls do not match the records involved. Do not place confidential deal materials into a service unless the firm has approved the service, its terms, access controls, retention behavior, and handling of sensitive information for that use. A general statement that a tool is secure is not a substitute for the firm’s own review. FINRA identifies privacy, accuracy, and misleading or corrupted data as concerns in AI use, while its securities-industry report warns that unusual conditions can fall outside model training. [2] [3]
Resume only after the owner documents the issue, adjusts the workflow or permissions, and reruns the affected sample. If the team cannot explain the correction, keep the tracker manual for that process. A manual fallback is a control, not a failure.
Keep human review tied to the decision boundary
The accountable reviewer should be close enough to the process to understand what a date means. A coordinator can verify whether a source contains a date, while a senior deal-team member may need to decide whether an owner’s confirmation changes the approved plan. Escalate conflicts that could affect sequencing, disclosure, bidder treatment, or a client commitment to the appropriate process owner and, where applicable, compliance or counsel.
Do not describe the tracker as proving compliance, privilege, accuracy, or closing certainty. NIST’s framework is voluntary, and FINRA’s notice says it does not create new legal or regulatory requirements or remove existing ones. [1] [2] Obligations that apply to a particular firm or transaction require qualified review of the facts and jurisdiction.
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. In this context, the practical implementation outcome is a source-checked milestone tracker with confirmed owners, dependencies, and date-change records. The software is user-operated, so the bank’s process owners remain responsible for review and action.
Key Points
- Treat the approved process plan and dated decision records as the source of truth; never let a model choose between conflicting sources.
- Separate planned, confirmed, overdue, and blocked states, and define the evidence required for each state.
- Require a human owner, dependency record, source reference, and change reason for every material milestone.
- Pilot on a bounded sample and stop for missing provenance, unexplained changes, confidentiality mismatches, or material false positives.
- Keep scheduling commitments, approvals, outreach, negotiation, and bidder-facing communications with human process owners.
Sources and references
Source retrieval dates record research access, not article publication. Read our editorial and date policy.
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.