Monthly portfolio packs tend to arrive in different shapes, on different dates and with different versions of the same KPI. The review problem is rarely a shortage of slides. It is finding the two changes that deserve a call before the board meeting.
AI can standardize the first pass: extract reported numbers, compare them with budget and prior periods, locate the stated explanation and draft follow-up questions. The process works only after the reporting schema is stable and every alert can be traced to a source cell or page.
Define the minimum reporting schema
Start with a common file that every company can map into. Keep the first version small:
- revenue, gross profit, EBITDA and cash
- budget and prior-year comparatives
- volume, price and mix where relevant
- customer retention and concentration
- working-capital metrics
- capex and headcount
- debt, liquidity and covenant headroom
- two or three operating KPIs specific to the business
Define each metric. “Churn” may mean logo churn, gross revenue churn or net revenue retention. A cross-portfolio view built on mixed definitions creates false comparisons.
Add fields for unit, currency, period, source, definition version and confidence. The definition version is useful when a company changes systems or restates a KPI.
Create a controlled monthly folder
A practical structure is:
/portfolio-monitoring/
/company-a/2026-08/
/company-b/2026-08/
kpi-dictionary.csv
alert-rules.csv
portfolio-summary-template.xlsx
Claude Cowork or ChatGPT Work can work through board packs and spreadsheets in a selected project. Claude Code or Codex is useful when the files need repeatable scripts, validation and version control. Use only an approved setup. Board materials, employee data and customer-level detail may carry tighter restrictions than a standard monthly P&L.
Keep original files read-only. Write extracted data and exceptions into new files.
Ingest the pack with evidence tags
Read the August files for Company A and kpi-dictionary.csv.
Extract the required metrics into company-a-2026-08.csv.
For every value, include:
- reporting period
- metric definition
- value and unit
- budget value
- prior-month and prior-year value where available
- source file, page or cell
- whether the value is reported, calculated or inferred
- confidence
Use UNKNOWN for missing values. Do not calculate an inferred KPI unless the
formula is defined in kpi-dictionary.csv. List definition conflicts and units
that changed. Preserve management's explanation as a quote with its source.
Review a fixed sample before scaling. Pick 15 rows covering financials, an operating KPI, a percentage, a covenant and a calculated metric. Check the source, unit, period and definition. A 98% extraction rate can still miss the one covenant figure that matters.
Separate movement from explanation
The harness can calculate variance. It should not treat management commentary as the cause.
Store three fields:
- observed movement
- management explanation
- investor follow-up
For example, gross margin may be 280 basis points below budget. Management may cite mix. The pack may show product mix moving only slightly. That leaves a follow-up on discounting, input cost or labour efficiency. Keeping the fields separate prevents a supplied explanation from becoming an accepted conclusion.
Write explicit alert rules
Rules should combine thresholds with context. Examples:
- revenue more than 8% below budget for two months
- EBITDA miss above 10% in one month
- cash runway below six months
- covenant headroom below 20%
- top-customer revenue down more than 15% year on year
- overdue receivables up more than five percentage points
- headcount growth ahead of revenue growth for three months
Put the rules in alert-rules.csv with owner, severity and required action. Avoid asking the model to identify “material” issues without a definition.
A worked early-warning prompt
Using company-a-2026-08.csv, alert-rules.csv and the prior six monthly extracts:
1. Calculate current-month, budget, prior-month and prior-year movement.
2. Test every rule and record PASS, TRIGGERED or NOT TESTABLE.
3. For triggered rules, quote the source evidence and management explanation.
4. Check whether the issue appeared in either of the prior two months.
5. Draft up to three follow-up questions. Each question must request a specific
bridge, schedule or operating action.
6. Flag any apparent data-definition change before comparing periods.
Write early-warning-log.csv and a one-page owner brief. Do not rank an issue
below MEDIUM because management says it is temporary. Do not infer covenant
compliance from liquidity alone.
That prompt creates a review queue. It does not decide whether the fund should intervene.
Reconcile covenants independently
Covenant reporting needs its own check. Capture the document definition, reported calculation, period and headroom. Tie debt and EBITDA inputs back to the monthly pack and credit schedule.
If the pack reports only a headroom percentage, ask for the underlying numerator, denominator and permitted adjustments. Do not let the harness reverse-engineer a covenant definition from a chart.
Use a separate calculation to test the arithmetic. Any difference goes to a person who has read the credit agreement.
Build a cross-portfolio taxonomy
Standard categories help the operating team see repeated issues:
- demand
- pricing
- customer concentration
- delivery and capacity
- labour
- gross margin
- working capital
- liquidity and covenant
- systems and reporting
Map each alert to one primary category and, if needed, one secondary category. Keep the company-specific evidence attached. The purpose is pattern recognition, not a league table of unrelated businesses.
A cross-portfolio view might reveal that three companies are hiring ahead of plan while revenue conversion is slowing. That can justify a focused review even if no individual company has crossed a severe threshold.
Draft the board-pack commentary after review
Once the numbers and alerts are signed off, the harness can draft a short commentary. Give it the approved facts and require evidence tags.
Draft the August portfolio commentary from approved-alerts.csv and
company-a-2026-08.csv.
Use four headings: performance, cash and covenant, operating issues, actions.
Every sentence containing a number must include an evidence tag in brackets.
Label management explanations as such. Do not add adjectives such as strong,
healthy or temporary unless the approved file uses them.
Keep the draft under 350 words.
Remove the tags only after a reviewer checks them. The tagged working draft should remain in the folder.
Verify the monthly run
Use three checks:
- Source audit: re-open the evidence behind every high-severity alert.
- Completeness check: compare expected metrics with extracted metrics and list missing rows.
- Change check: identify formulas, definitions or units that differ from the prior month.
Then sample quiet companies. Pick one with no alerts and inspect its largest three movements. This is the portfolio-monitoring version of a false-negative test. A rule set that never fires can look reassuring for the wrong reason.
Keep a decision log
Record the alert, reviewer conclusion, action, owner and closure evidence. This matters when the same issue returns three months later. It also shows which rules create noise and which ones catch real problems early.
Review thresholds quarterly. Do not tune them after every false alarm. A changing rule set makes trend data hard to interpret.
Human escalation
The harness can tell you that cash conversion deteriorated and the explanation is unsupported. A person decides whether to call the CFO, ask for a 13-week cash flow or wait for the next close.
Escalation also depends on context absent from the pack: a systems migration, customer negotiation, planned inventory build or leadership change. Keep that context in an investor note with an owner and date. Do not let it become a permanent excuse that suppresses alerts.
Handle late and revised packs
Monthly monitoring needs version rules. Record received time, reporting period and file version. When a company sends a revised pack, preserve both copies, rerun the extract and produce a difference report. The alert log should show which conclusions changed.
Do not mix preliminary and final numbers without a label. A late company should appear as missing, not quietly carry forward the prior month. The absence of a pack can be an operating signal in its own right, especially when reporting has usually been timely.
Tune for leading indicators
Financial statements confirm what has already happened. Add a small number of operating indicators that lead revenue, margin or cash for each business: booked orders, pipeline conversion, service backlog, utilization, customer tickets, labour hours or inventory ageing.
For every leading indicator, document the expected link to the financial result and the lag. Then test whether the relationship holds. A metric that moves every month without changing the forecast creates noise. Remove it or lower its prominence.
The harness can plot and test relationships. The deal team decides whether the business logic makes sense and whether the data has been measured consistently.
Review actions, not only variances
Carry prior actions into the next month’s run. Each action should have an owner, date, expected effect and closure evidence. Ask which overdue actions relate to a triggered alert. This prevents the board pack from raising the same issue in fresh language without showing what happened after the last discussion.
A closed action should link to evidence: a revised forecast, customer recovery plan, headcount freeze or cash schedule. “Discussed with management” is an activity, not closure. The portfolio team can then distinguish a new issue from an old issue that never received a response.
Use commentary as a data-quality test
When the draft cannot explain a movement from the extracted fields, inspect the pack before adding prose. The missing item may be a bridge, a definition or an operating metric. Record that gap and ask the company for the smallest schedule that resolves it.
Over time, the gap log improves the reporting package. Repeated questions about price, volume and mix suggest that the template needs that bridge every month. Repeated confusion about working capital may call for a standard ageing and inventory schedule. Automation is most useful when it makes these recurring holes visible.
In practice
Portfolio monitoring improves when the file is boring: stable definitions, repeatable extracts, explicit rules and a short queue of exceptions. AI helps produce that file across companies that report differently. The early-warning value comes from the review discipline around it.
The monthly question is simple: what changed, what evidence explains it and what action follows? If the system cannot answer all three, it is producing commentary rather than control.
This is the ownership-stage workflow in How to Use AI in Private Equity. For the broader ownership context, see The Life Cycle of a Private Equity Fund.
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