An LBO model can be made to balance while still being wrong. That is the main problem with asking an AI agent to build one from a short prompt.
The useful applications are narrower and more practical: normalize source data, set up repetitive schedules, inspect formulas, generate sensitivities and explain movements. The deal team still owns the capital structure, operating case and investment judgment.
Choose the right tasks
Good candidates have a defined input and a testable output:
- map historical financials into a standard chart of accounts
- create a sources-and-uses schedule from agreed terms
- roll a debt schedule using specified tranches and cash-sweep rules
- find inconsistent formulas across a row or period
- rebuild a sensitivity table
- reconcile returns to entry, operations, leverage and exit
Avoid one-line instructions such as “build an LBO for this company.” The agent will make choices about debt, fees, cash, circularity and exit assumptions that may be invisible in the finished workbook.
Work on a copy
Create a local project folder with the model copy, source schedules and a written convention file. Claude Code or Codex can inspect spreadsheet files, run scripts and write revised copies from a terminal-based workspace. Check firm approval before putting deal data into any environment. Disable unrelated connectors and use a clean folder.
The convention file should state:
- sign convention
- fiscal-year dates
- units and currency
- hardcode and formula colours
- treatment of cash, fees and minimum cash
- circularity setting
- debt sweep priority
- return dates and day-count approach
- cells that may be changed
That document is the model mandate. The agent should stop when an instruction is missing.
Normalize the source data first
Most model problems start before the formulas. A historical P&L may mix reported and pro forma results, use shifting segment labels or include an adjusted EBITDA bridge whose items do not tie to the accounts.
Use a staging file. Keep original values, normalized values, mapping rule and source reference in separate columns.
Read historical-financials.xlsx and model-conventions.md.
Create normalized-historicals.xlsx.
For every source row:
- preserve the original label and value
- map it to the standard chart of accounts
- record source sheet and cell
- record any sign change, unit conversion or perimeter adjustment
- mark ambiguous mappings REVIEW REQUIRED
Do not net revenue and cost lines. Do not treat management adjustments as
accepted EBITDA adjustments. Produce a reconciliation from reported EBITDA to
the normalized output and stop if it does not tie.
Review the mapping before using it in the model. Once a mislabeled line flows through five years of projections, it becomes harder to spot.
Build modules, not a black box
Ask for one schedule at a time. Start with sources and uses, then operating case, debt schedule, tax and returns. Save a new file after each stage.
For a debt schedule:
In lbo-working-copy.xlsx, build the debt schedule only in the blank range
Debt!B8:N45.
Use the tranche terms in debt-terms.xlsx. Follow model-conventions.md.
Show beginning balance, mandatory amortization, optional repayment, cash
interest, PIK interest, new borrowing and ending balance for each tranche.
Apply excess cash in the stated priority after minimum cash.
Do not create or enable circular references. Put unresolved choices in
Debt!B48:B60 as review notes.
After writing the schedule, run these checks:
- beginning plus movements equals ending balance
- no repayment exceeds available balance
- cash sweep does not use cash below minimum
- interest rate and base-rate floor follow the term sheet
- maturity dates do not extend past the exit date without a warning
Save as lbo-debt-v1.xlsx. Do not overwrite the input.
A bounded range and a new filename make review easier. They also reduce the risk of the agent “helpfully” changing unrelated tabs.
Use formula audits where they work
Agents are good at pattern detection. Ask them to identify formulas that break a row pattern, hardcodes inside forecast blocks, inconsistent absolute references and blank cells feeding outputs.
Audit formulas in Operating Case, Debt and Returns.
Create formula-audit.csv with sheet, cell, issue type, current formula, expected
pattern and confidence.
Flag only. Do not repair formulas.
Treat deliberate subtotal, first-period and terminal-period differences as
possible exceptions. Group repeated issues caused by one root cell.
Review the list before authorizing repairs. A formula that differs from its neighbours may be the only correct one in the row.
Reconcile the return
A model review should explain returns, not only recalculate them. Build a bridge from entry equity to exit equity and from entry multiple to gross MOIC and IRR.
At minimum, reconcile:
- EBITDA growth
- margin change
- cash generation
- debt paydown
- multiple movement
- add-on equity or leakage
- holding period
Use a separate script or workbook check to recalculate MOIC and IRR from dated cash flows. If the independent result differs, stop before writing commentary.
Generate sensitivities from agreed assumptions
Once the base case is locked, an agent can build two-way tables and scenario copies. Specify the exact variables, ranges and step sizes. Avoid broad requests to “stress the model.” That invites arbitrary downside cases.
Useful tests include exit multiple versus EBITDA, revenue growth versus margin, interest rate versus leverage and hold period versus exit multiple. Check that each corner case changes the intended inputs and that the base-case intersection ties to the returns page.
Run hallucination tests
There are several fast ways to find hidden assumptions:
- Remove one debt term and rerun the setup. The agent should stop or flag the gap.
- Insert a deliberately inconsistent unit label. It should identify the conflict.
- Compare every hardcode written by the agent against the source file.
- Trace a sample of ten formulas from output to input.
- Recalculate the workbook in Excel and check for errors, stale values and unsupported functions.
Keep a write log with every changed cell. A model cannot be reviewed from a polished summary alone.
What not to trust
Do not delegate the investment case, the appropriate leverage, covenant interpretation or whether an adjustment belongs in EBITDA. The same applies to management-case credibility and exit multiple selection.
Be careful with macros, add-ins, named ranges and iterative calculations. A script may preserve cell values while damaging behaviour that only appears when the workbook opens in Excel.
A visually complete debt schedule also may mishandle daily interest, stub periods, commitment fees, tax limitations or PIK toggles. Those terms need direct review against the credit documents.
Learning model versus production model
AI can be useful for learning: explain a formula, show a simple schedule or critique a practice case. A production underwriting model has a different bar. It needs the firm’s conventions, a source trail, locked review points and a full calculation check.
If somebody is still learning the mechanics, a structured course or reviewed template can save time. Wall Street Prep’s LBO materials are a natural reference here because they teach the schedule relationships that an agent can otherwise hide. The modeler needs to understand the answer before automating it.
Final review
Before the model goes to IC, confirm:
- source data ties to the approved financial pack
- every new hardcode has a source or assumption owner
- sources and uses balance
- debt rolls correctly and respects cash constraints
- taxes and interest follow the agreed mechanics
- sensitivities point to the intended inputs
- returns recalculate independently
- no agent-generated note is being treated as evidence
- the file has been opened, recalculated and reviewed in Excel
Control the workbook writes
Before any agent changes a production-style workbook, define permitted sheets, ranges and file names. Protect source tabs. Ask for a cell-level change log with old formula, new formula and reason. If the tool cannot produce that log, use it for audit and code generation rather than direct workbook edits.
Open every generated file in Excel with calculation set correctly. Check named ranges, links, data tables, conditional formatting and print areas. These features sit outside the obvious grid and can break even when the headline returns still look right.
Test downside mechanics
Downside cases often expose model errors hidden by the base case. Push revenue and margin low enough to reduce cash, turn off optional repayment and approach covenant limits. Check that revolver draws, interest expense, cash sweep and minimum-cash logic behave in the right order.
Then test an upside with faster debt paydown. A schedule that allows a negative balance or continues charging interest after repayment has a structural problem. Review the dates around acquisition, refinancing and exit, especially when the hold period includes stubs.
Ask the agent to describe the expected movement before changing the inputs. Compare that expectation with the workbook. A mismatch can identify a broken reference faster than inspecting every formula.
Review explanations like formulas
Generated commentary should tie to a return bridge. “Returns improve due to operational performance and deleveraging” says almost nothing. Require the contribution from EBITDA growth, cash generation, multiple movement and holding period.
Check that the explanation follows the model direction. I have seen commentary describe deleveraging as the main driver when the debt balance barely changed and nearly all value came from multiple expansion. The paragraph may read well while the economics are wrong. Keep commentary downstream of a signed-off model and evidence table.
Keep model assumptions in one place
A harness works better when assumption cells and source notes are explicit. Create an assumptions register with owner, value, unit, effective date, source and model cell. Use it to compare the workbook against the approved case.
When a number changes, update the register and model together. Then ask for a difference report. This avoids the familiar problem of one scenario carrying an old tax rate or debt margin because the change was hardcoded in several tabs. The register also gives the IC memo a clean route back to the underwriting.
In practice
The agent earns its keep on setup, repetition and testing. It can shorten a formula audit from hours to minutes and produce a useful first version of a bounded schedule. The final workbook still needs a modeler who can explain every important line and spot an answer that is mechanically neat but economically wrong.
That is the test I use: if the team cannot rebuild the return bridge without the agent, the automation has gone too far.
For the rest of the lifecycle, see How to Use AI in Private Equity. Once the model is signed off, its evidence should flow into the IC memo workflow.
Tool references
Claude Code: https://www.anthropic.com/claude-code
OpenAI Codex: https://openai.com/codex/