What Should Associates Practice Before Using AI on a Live Sell-Side Mandate?

Start with a narrow task and a named reviewer

Begin with one low-consequence task, such as extracting selected fields from a fictional financial schedule or tracing statements in a mock company overview. Name the associate who performs the task and the supervisor who reviews it. Define the permitted output before the exercise starts. For example, the output may be a draft list of revenue, debt, and customer-concentration figures, each linked to a page and line reference.

The reviewer should assess the work product and the process used to produce it. Did the associate preserve the source file, record the tool and prompt version, and distinguish an extracted value from an interpretation? Did the associate stop when the source was incomplete? This mirrors the practical point in FINRA's guidance that firms should evaluate an AI tool before deployment and maintain supervision appropriate to the business use. [1]

Exercise 1: extract, then reconcile

Give the associate a synthetic income statement, debt schedule, and short management summary containing a few deliberately repeated figures. Ask the AI tool to extract specified fields into a review sheet. The associate must then compare every extracted field with the source record instead of accepting the output as a transcription.

Require a second pass for arithmetic and labels. Revenue should not be confused with bookings. Total debt should not be replaced by net debt. A period-end amount should not be presented as a full-year amount. If the source uses parentheses, thousands, or a nonstandard label, the associate should preserve the source convention and flag the ambiguity. A useful competency standard is that every reported field has a source locator, unit, period, and status such as verified, unclear, or missing.

Introduce a controlled discrepancy, such as one debt balance differing between two synthetic schedules. The associate passes only if the discrepancy is visible in the review record and escalated rather than silently normalized.

Exercise 2: trace every material claim

Use a fictional company description and ask the associate to produce a short draft paragraph about operations, customers, and financial performance. The exercise is not a writing contest. The associate must build a claim trace beside the draft. Each material statement should point to a source excerpt, filing-style page, or approved internal record. Statements that are inferences must be labeled as inferences.

Include a source that supports only part of a sentence. A schedule may support a customer percentage but not the reason for a change. The associate should split the sentence, retain the supported fact, and remove or qualify the unsupported explanation. The reviewer should look for inflated certainty, unsupported adjectives, and claims generated from the prompt rather than from the source.

This exercise teaches a durable human-in-the-loop habit: AI may propose language, but a person remains responsible for deciding whether the source actually supports the language.

Exercise 3: test permission boundaries before the prompt

Create three input bins: synthetic data that is always allowed, approved internal material with a stated handling rule, and live or client-confidential information that is out of scope. Ask the associate to classify each item before using any tool. The decision record should state what may be entered, into which approved environment, for which task, and who can view the result.

Include an apparently harmless request that contains a confidential attachment, a pasted customer name, or an unnecessary personal identifier. The correct action is to stop, remove the unnecessary detail only if policy permits, or ask the supervisor for direction. It is not enough to say that a vendor or model is secure. FINRA identifies data privacy and integrity as technology-governance considerations for member firms using generative AI, while noting that rules remain applicable when a third party supplies the technology. [1]

Do not treat this exercise as proof of privilege, confidentiality, or security. The associate should know the approved data-handling policy and escalation path for an uncertain classification. The supervisor should record whether the boundary was respected without relying on an AI-generated assurance.

Exercise 4: resolve conflicting evidence

Provide two synthetic source records with a genuine conflict, such as different fiscal periods, a revised balance, or inconsistent customer counts. Ask the associate to produce a conflict log with source names, dates, definitions, and the exact difference. The associate should not choose the more convenient figure merely because the AI output sounds confident.

The expected action is to check the source hierarchy and task instructions, identify whether the conflict is explainable, and escalate when it is not. If a revised statement supersedes an earlier draft, the record should say why. If neither source is clearly authoritative, the draft should preserve uncertainty and avoid a single definitive number. A supervisor can score whether the associate separated facts, assumptions, and unresolved questions.

This matters in sell-side materials, where a polished sentence can hide a reconciliation problem. A human reviewer should decide whether the issue affects the workstream, requires a new source request, or stops further drafting.

Exercise 5: refuse to invent missing values

Give the AI a mock prompt that asks for a missing EBITDA margin, customer-growth percentage, or transaction assumption while withholding the underlying inputs. The associate must instruct the tool to state that the value is unavailable, identify the missing fields, and specify what source would be needed to calculate it. The answer should not contain a plausible-looking estimate unless the exercise explicitly authorizes an assumption and labels it.

Add a second scenario in which the source uses a blank, a dash, or an ambiguous abbreviation. The associate should preserve the original notation and ask what it means rather than converting it to zero. If an illustrative calculation is authorized, show the formula and label the result as illustrative. The supervisor should check that the associate distinguishes a missing value from zero and a reported figure from an estimate.

A refusal is a useful outcome, not a failed use of AI. It shows that the associate understands when the source record is insufficient for a live deliverable.

Use stopping criteria, not enthusiasm, to end the pilot

Before practice begins, write stopping conditions. Stop if the tool exposes restricted input, returns an uncited material claim, changes a unit or period, repeats a conflict without flagging it, or produces a value that cannot be traced. Stop also when the reviewer cannot reproduce the result from the recorded source and prompt. Continuing to prompt is not a substitute for resolving the problem.

Define what happens after a stop: preserve the evidence, remove or quarantine the output according to policy, notify the named reviewer, and decide whether the task returns to manual handling. Do not promise that a pilot is compliant, secure, accurate, or privileged merely because no incident occurred. NIST describes its Generative AI Profile as a voluntary risk-management resource for incorporating trustworthiness into design, use, and evaluation, which supports treating these criteria as an evaluation aid rather than a safe harbor. [2]

Record observed competency and supervisor sign-off

The competency record should identify the task, allowed data class, approved tool or environment, source-of-truth records, test cases, observed errors, corrections, escalation decisions, and reviewer conclusion. It should say exactly what the associate demonstrated, such as extraction review for synthetic financial schedules, and what remains outside scope, such as live client uploads or external communications.

Use a small rating scale with evidence: demonstrated, demonstrated with correction, or not demonstrated. A sign-off should include the date, reviewer role, task boundary, and any expiration or re-review trigger required by the firm's policy. Sign-off authorizes only the defined task under the stated controls. It does not authorize autonomous outreach, negotiation, approvals, disclosure, or use of unapproved data.

This documentation prevents a common governance mistake: treating a course completion badge or prompt familiarity as proof that an associate can safely handle every AI use case. FINRA says applicable obligations depend on how a member firm deploys the technology, so competency should be scoped to the actual workflow. [1]

Move from practice to a controlled live workflow

After sign-off, run a limited live pilot with pre-approved inputs and a named human reviewer. Start with a task where the source set is known and the output can be checked before it reaches a client, buyer, or counterparty. Keep the first workflow read-only: no automated disclosure, outreach, negotiation, approval, or change to a system of record.

Set a review gate for every material figure, claim, attachment, and recipient. The associate should retain the source trace and record corrections. If the pilot produces repeated unsupported claims, unexplained extraction changes, permission uncertainty, or reviewer overload, pause the task and return to practice. A pilot is evidence about one workflow, not evidence that every future use is safe.

Public descriptions should be equally precise. In 2024, the SEC announced charges against two investment advisers over false or misleading statements about purported AI use. The lesson for a boutique bank is not a universal rule about every adviser, but a warning to describe only capabilities and controls that actually exist. [3]

Make the checklist part of the handoff

The practical handoff is a one-page checklist attached to the workstream. It should ask: What exact task is AI assisting? What data class is permitted? What are the source-of-truth records? Who checks each material output? What errors must stop the task? Where are conflicts and missing values logged? Who approves the next step?

The associate should complete the checklist before using the tool, and the reviewer should close it after checking the result. If the answer to any question is unknown, the task remains in practice or manual handling. This makes the implementation outcome concrete: an associate practice checklist plus a supervisor-reviewed competency record for approved AI-assisted tasks.

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. It is user-operated software, so a firm's own permissions, review steps, and professional responsibilities still govern how it uses any AI-assisted workflow.