The best management-meeting questions usually come from details that do not quite line up. Revenue grows 18%, headcount grows 6%, service levels improve and churn is described as “stable.” Somewhere in that set is a question worth asking.
AI is useful before a management meeting because it can hold the CIM, diligence notes and open issues in one working structure. The job is to build a question tree with evidence behind it, then test whether the branches survive plausible answers. A generic list of 50 questions is easy and rarely helpful in the room.
Build the evidence pack
Create a controlled folder with:
- the CIM or approved extracts
- the latest model and operating case
- customer, cohort and KPI analyses
- expert-call notes
- the open-issues list
- the meeting agenda
Use Claude Cowork, ChatGPT Work or another approved knowledge-work environment with access limited to that folder. Check what the tool uploads, what stays local and which connectors are enabled. Keep personal data and restricted third-party material outside the workspace unless the firm’s rules allow it.
Add an evidence-index.csv with columns for source, page or cell, date, topic and short description. The index gives the harness a route back to the material and makes citations easier to audit.
Start from gaps, not topics
A topic list produces “Tell us about your growth strategy.” A gap list produces “Price contributed six points of growth last year. Which customer cohorts accepted the increase, and what happened to unit retention in the next two quarters?”
Ask the harness to identify four kinds of gaps:
- internal contradictions
- claims unsupported by data
- changes in definition or time period
- model assumptions that require management evidence
Every gap should carry a source tag. If the source is unclear, the gap stays unverified.
A worked question-tree prompt
Read the files in /management-meeting-pack and the evidence index.
Build a management-meeting question tree for a 90-minute session.
For each topic:
1. State the diligence gap in one sentence.
2. Cite the source page, file or model cell that created the gap.
3. Draft one primary question in plain language.
4. Draft follow-up branches for:
a. a specific answer with supporting evidence
b. a broad answer without evidence
c. an answer that conflicts with the pack
5. State what evidence would close the issue after the meeting.
6. Label priority HIGH, MEDIUM or LOW.
Cover revenue quality, customer retention, pricing, sales capacity, delivery,
people, cash conversion and forecast credibility. Do not fill a topic when the
pack contains no real issue.
Put no more than 12 questions in the core agenda. Move the rest to an appendix.
Write the result to question-tree.md and a one-page chair sheet to chair-sheet.md.
Do not invent benchmark data or management views.
The 12-question limit forces prioritization. It also leaves room for follow-ups, which is where most of the useful information appears.
Make each branch earn its place
A follow-up should react to the answer. Avoid three versions of the same question.
Take a customer-concentration issue. The primary question might ask why the top customer fell from 14% to 9% of revenue. If management says the customer shrank, ask about retention and gross-margin effect. If management says other customers grew faster, ask for the cohort bridge. If management disputes the figure, confirm the definition used in the CIM and request the source schedule.
Write those branches in advance. In the meeting, the lead investor can listen instead of inventing the next question while management is still speaking.
Stress-test the questions
The first draft often contains questions that sound clever and produce little evidence. Run a mock answer pass.
Act as a well-prepared management team answering question-tree.md.
Use only facts contained in the evidence pack. Where the pack is silent, give a
plausible but non-committal answer and label it SIMULATED.
For each primary question:
- draft the strongest likely answer
- identify which planned follow-up still works
- flag questions that can be answered with a slogan
- rewrite any question that contains two or more separate asks
- flag wording that sounds accusatory or assumes the conclusion
Write the critique to question-stress-test.md. Do not add the simulated answers
to the evidence index.
Keep simulated answers physically separate from facts. Otherwise they have a habit of reappearing later in an IC memo.
Check the tree against the model
Several questions should tie to the underwriting. For every major assumption, ask what management evidence would move it.
If the model assumes 8% volume growth, the tree might cover pipeline conversion, salesperson ramp, capacity and churn. If the answer would not change a model input or a diligence conclusion, the question may belong below the line.
Create a mapping from question ID to model assumption, memo section and owner. That turns meeting preparation into a controlled handoff rather than a standalone document.
Use contradiction hunting carefully
Tell the harness to quote both sides of a conflict. “Growth figures are inconsistent” is too vague. “The CIM states 2025 organic growth of 12% on page 18; the operating case shows 9% in cell G42” is reviewable.
Some differences will come from currency, perimeter, pro forma adjustments or a fiscal-year change. Mark those as reconciliation items. Do not frame them as management credibility issues before checking the definitions.
Plan the room
A strong question tree also says who asks, who follows up and which questions can be dropped if time runs short. Assign one note taker and one person to maintain the gap log.
Keep the chair sheet short:
- five issues that matter to the thesis
- the exact evidence conflict
- the opening question
- the must-get follow-up
- the closing request
The full tree can sit in the appendix. Nobody can run a conversation from a 20-page prompt book.
Capture notes into the same structure
After the meeting, put notes under the existing question IDs. Claude Cowork can turn a transcript or approved notes into a gap log, provided the meeting material is permitted in the environment.
Map management-meeting-notes.docx to the question IDs in question-tree.md.
For each ID, record:
- answer given
- supporting evidence promised
- open follow-up
- owner
- due date if stated
Tag every statement as MANAGEMENT STATEMENT, DOCUMENTED FACT or TEAM JUDGMENT.
Do not upgrade a management statement to documented fact.
Write post-meeting-gap-log.csv and list any new claim that affects the model.
That tag discipline is worth keeping through the rest of diligence. A confident answer in the room remains a management statement until the promised schedule arrives.
Verify the output
Review all high-priority questions against the original source. Then sample five medium-priority questions and check the same things:
- source citation is correct
- question follows from the evidence
- wording contains one ask
- follow-up branches are genuinely different
- the requested evidence can be obtained
Also search the tree for numbers. Every number needs a source. Dates, percentages and customer counts are common places for a clean-sounding question to drift away from the pack.
After the meeting
Close the loop within a day. Update the evidence matrix, model-assumption log and IC issues list. Separate items answered in the room from items supported later by documents.
The meeting should leave a shorter, sharper issue list. If the harness produces more prose and no change in what the team believes, it has mainly reformatted notes.
Build questions around evidence requests
A question is stronger when the team knows what it wants back. “How do you think about churn?” may start a discussion. “Please bridge beginning customers to ending customers by cohort and show gross revenue retention for the last eight quarters” creates evidence that can enter the underwriting.
For each high-priority question, add a closing request. Management may answer verbally and still agree to send a schedule. Capture the promised file, owner and timing. The request should fit the issue. Avoid asking for broad data dumps when a one-page bridge will settle it.
Protect the meeting from the prompt book
Question trees can make teams rigid. Mark the three questions that must be answered and treat the remaining branches as support. If management volunteers the evidence, skip the prepared wording. If a new issue appears, follow it.
The tool has no feel for tone, fatigue or the difference between a CEO working through an answer and avoiding one. Agree in advance who can interrupt, who watches time and who decides to move on. Those room mechanics matter more than an extra page of generated questions.
Run a post-meeting contradiction check
Once the notes are mapped, compare management statements with the pre-meeting pack. Ask the harness to quote the earlier evidence and the new statement side by side. Classify each item as reconciled, new information, definition difference or unresolved conflict.
Do not ask it to judge honesty. A conflict may come from perimeter, timing or shorthand in the room. The useful output is a clean list of what changed and what document would settle it. Feed confirmed items into the model and IC evidence matrix. Keep unresolved verbal claims tagged until the support arrives.
Order the agenda around the thesis
Open with enough business context to let management settle, then move to the two or three issues that can change the case. Leave lower-priority confirmatory questions for the end or a written follow-up. The order should follow the investment thesis, not the layout of the CIM.
For a service business, customer retention and labour capacity may belong ahead of market growth. For a project business, backlog quality and cash conversion may lead. Ask the harness to propose an order and show its link to the thesis. The deal lead makes the final call.
In practice
The value of the workflow is preparation under time pressure. It turns scattered diligence into a dozen questions with evidence, branches and a clear request behind each one. The investor still reads the room, decides when to press and distinguishes an honest operational answer from a polished deflection.
A good question tree helps that judgment. It never supplies it.
This workflow picks up from AI for CIM Review and feeds the evidence into the IC memo workflow. The full lifecycle map is in How to Use AI in Private Equity.
Tool references
Claude Cowork: https://www.anthropic.com/product/claude-cowork
ChatGPT Work and Codex: https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex