---
title: "AI in financial modelling and due diligence"
description: "Where AI helps in 2026 AI tools are now widely embedded in spreadsheets, document platforms and data services, and organisations deploy approved…"
url: https://optimizeall.com/learn/project-finance-and-financial-modelling/ai-in-modelling-and-diligence
updated: 2026-10-05
---

Project Finance & Financial Modelling · AI-enabled analysis with governance · lesson 18 of 20 · 14 min

# AI in financial modelling and due diligence

## Where AI helps in 2026

AI tools are now widely embedded in spreadsheets, document platforms and data services, and organisations deploy approved enterprise assistants. In project finance, the most valuable uses combine large document volumes, repetitive analysis and pattern detection.

| Task | How AI helps | Human responsibility |
|---|---|---|
| Contract review | Extract key terms (tenor, tariff formula, LDs, caps, termination payments) into a term matrix; flag inconsistencies between contracts | Lawyers verify every extracted term and interpret meaning |
| Model review | Explain formulas, detect hard-codes, inconsistent formulas across rows, broken links | Modellers confirm and fix; model auditor remains independent |
| Model building support | Draft formula logic for standard blocks (indexation, flags, sculpting) | Modeller tests against known answers |
| Scenario analysis | Rapidly generate and summarise sensitivity runs | Analyst checks inputs and interprets |
| Due diligence Q&A | Summarise data rooms; answer questions with citations | Advisers verify against source documents |
| Monitoring | Detect anomalies in operating data, predict maintenance needs, forecast CFADS | Asset manager validates and decides |
| Report drafting | Draft lender reports and variance commentary from verified data | Finance lead reviews and signs |

## Worked example: contract term extraction

*Illustrative.* A fictional infrastructure fund in London reviews a 400-page data room for a Gulf water project. An approved AI tool extracts a term matrix: PPA tenor, tariff indexation, availability deductions, EPC LDs and caps, O&M penalties, termination compensation. It flags that the EPC delay LD cap appears lower than the offtaker's late-COD penalty plus debt service for a six-month delay. The legal team verifies the clauses; the flag is correct, and the issue becomes a negotiation point. The AI saved days of reading, but the conclusion and negotiation strategy came from people.

## Worked example: formula consistency checks

*Illustrative.* A modeller asks an AI assistant to scan a 30-sheet model for rows where the formula changes partway across the timeline. It finds 14 inconsistencies; 11 are intentional (phase changes handled with flags would have been better) and 3 are genuine errors, including a tax row that stops indexing after year 10. The modeller fixes the errors and restructures the intentional changes using flags.

## Prompting and workflow tips

- Provide context: model purpose, structure, conventions and the loan agreement definitions.
- Ask for **citations** to page and clause, or to cell references, for every claim.
- Ask the AI to list what it could not determine or is uncertain about.
- Test AI-drafted formulas against hand calculations or known outputs.
- Never paste confidential deal data into tools not approved by your organisation; many data rooms have strict confidentiality obligations.

## Limitations to respect

- AI can **misread** complex clauses, especially defined terms that cross-reference other documents.
- It can produce **plausible but wrong formulas**, such as discounting with the wrong period convention.
- It lacks context about negotiations, relationships and market practice unless provided.
- Outputs can vary between runs; results used for decisions must be verified and recorded.

## Machine learning for forecasting

For operating portfolios, machine-learning models can forecast output (e.g., from weather data), predict equipment failures and estimate opex. These can improve CFADS forecasts, but lenders' base cases are typically set by independent advisers using transparent methods. Use ML forecasts as inputs to discussion, with clear documentation of data, validation and limitations.

## Common mistakes

- Treating AI-extracted terms as verified.
- Using AI to "fix" models without understanding the change.
- Losing the audit trail of what AI changed.
- Assuming AI can replace the independent model audit.

## A practical workflow

A sensible workflow for AI-assisted diligence has four steps. First, define the questions and the output format, such as a term matrix with clause references. Second, run the AI extraction on approved infrastructure. Third, have a qualified reviewer verify each item against the source, marking it confirmed, corrected or unclear. Fourth, store the verified matrix, the AI draft and the reviewer's notes together. This produces faster diligence and a clear audit trail, and over time it shows where the AI is reliable and where it tends to err.

## Quick self-check

For any AI-assisted output you are about to rely on, ask: could I show a sceptical reviewer exactly where each fact came from? If not, it is not ready.

## Prompt template: term extraction with citations (approved tool and approved data only)

```text
CONTEXT: Lender-side review of a project finance deal. Documents: PPA, EPC contract, O&M agreement (pasted/attached).
Key defined terms appear in each contract's definitions clause; follow them.
TASK: Extract the terms below into a table.
FORMAT: Term | Value | Verbatim quote (max 40 words) | Document and clause | Depends on definition (clause)
TERMS: PPA tenor; tariff formula and indexation (base date, lag); availability deductions; late-COD penalties;
EPC price; completion date; delay LDs (rate, cap); performance LDs (rate, cap); aggregate liability cap;
O&M availability guarantee; termination compensation (each default type).
RULES: cite a clause for every value; do not infer values not stated; list items you could not determine under
"UNRESOLVED"; flag any inconsistency between documents under "POSSIBLE GAPS" with both clause references.
```

## Hands-on: a verification log in Python

```python
import csv
from datetime import date

rows = [  # term, ai_value, source_clause, reviewer_value, status
    ("EPC delay LD cap", "10% of contract price (all LDs)", "EPC 20.6", "10% of price, delay LDs only", "corrected"),
    ("PPA tenor", "25 years from COD", "PPA 3.1", "25 years from COD", "confirmed"),
    ("Tariff indexation lag", "not found", "-", "CPI with 3-month lag, PPA 7.4(b)", "corrected"),
]
with open("verification_log.csv", "w", newline="") as f:
    w = csv.writer(f)
    w.writerow(["term", "ai_value", "source_clause", "reviewer_value", "status", "reviewer", "date"])
    for r in rows:
        w.writerow([*r, "A. Reviewer", date.today().isoformat()])
confirmed = sum(r[4] == "confirmed" for r in rows)
print(f"{confirmed}/{len(rows)} confirmed without change; review every 'corrected' item with counsel")
```

Tracking the share of AI extractions that needed correction, by term type, shows where the tool is reliable and where extra review is needed.

## How to measure success

- Every term in the final matrix has a clause reference and a reviewer status.
- Correction rate tracked by term type; high-error types get mandatory specialist review.
- No confidential deal data processed outside approved tools.

## Video lecture: AI in financial modelling and due diligence

Lecture coming soon · 9 chapters · about 9 minutes. Read the full transcript below.

1. Four hundred pages by Friday
2. Why it matters
3. The concept: a task map
4. Worked example one: a formula consistency scan
5. Worked example two: a Gulf water term matrix
6. Watch me do it: a prompt with citations and uncertainty
7. Limitations and machine learning
8. A four-step workflow and common mistakes
9. Recap and try this now

## Lecture transcript

### Four hundred pages by Friday

A fund manager is given a four-hundred-page data room on a Friday afternoon and asked for a view on the deal by Monday. Ten years ago, that meant a weekend of reading. Today, an approved AI tool can extract the key terms into a matrix in minutes. That's genuinely transformative. It's also genuinely dangerous if nobody checks the output, because AI can misread defined terms, miss cross-references and produce confident, plausible, wrong answers. In this lecture you'll learn where AI adds real value in project finance, two worked examples on contract term extraction and model formula checks, how to prompt for citations and uncertainty, the limitations you must respect, and a four-step verification workflow. By the end, you'll be able to use AI to go faster without giving up accuracy.

### Why it matters

Why does this matter? Project finance lives on large document sets and large models: hundreds of pages of contracts, dozens of model sheets, and months of diligence. In a competitive bid, speed is an advantage. AI can compress the reading and checking time dramatically. But in this field, a single unverified error, such as a misread damages cap or a wrong discounting convention in an AI-drafted formula, can move debt capacity or returns by millions, or leave the SPV exposed to a risk nobody priced. So the goal isn't to use AI or avoid it. It's to capture the speed while keeping every fact traceable and every figure verified.

### The concept: a task map

Here's the map. For contract review, AI can extract key terms, such as tenor, tariff formula, damages, caps and termination payments, into a term matrix, and flag inconsistencies between contracts. Lawyers verify every term and interpret meaning. For model review, AI can explain formulas and help detect hard-codes, inconsistent rows and broken links. Modellers confirm and fix, and the model auditor stays independent. For model building, it can draft standard blocks like indexation, flags or sculpting, which the modeller tests against known answers. It can generate and summarise sensitivity runs. It can answer diligence questions with citations to the data room. For operating assets, machine learning can detect anomalies, predict maintenance needs and forecast output. And it can draft lender reports from verified data. In every row, a qualified human owns the result. Think of AI as a very fast, very well-read junior analyst who has never worked on a deal and occasionally makes things up.

### Worked example one: a formula consistency scan

Let's start with a simple example from the lesson. A modeller asks an AI assistant to scan a thirty-sheet model for rows where the formula changes partway across the timeline. It finds fourteen. Eleven turn out to be intentional, where the modeller handled a phase change by writing a different formula, which would have been better handled with flags. Three are genuine errors, including a tax row that stops indexing after year ten. That one error would have understated tax in the later years and overstated equity returns. The modeller fixes the errors and restructures the intentional changes using flags, so next time the scan comes back cleaner. Notice the ratio: most flags weren't errors, but the few that were mattered a lot. That's exactly the kind of task where AI saves time and a human decides.

### Worked example two: a Gulf water term matrix

Now a realistic scenario, adapted from the lesson. A fictional infrastructure fund in London reviews a four-hundred-page data room for a Gulf water project. An approved AI tool extracts a term matrix: PPA tenor, tariff indexation, availability deductions, EPC delay damages and caps, O and M penalties, and termination compensation, each with a clause reference. It flags that the EPC delay damages cap appears lower than the offtaker's late-COD penalty plus debt service for a six-month delay. That's a real bankability issue worth raising. Then the lawyer verifies every row against the source. Most are right. One isn't: the AI read a cap as covering all liquidated damages, but a defined term elsewhere meant performance damages sat under a separate cap. The matrix is corrected, and the reviewer's notes are saved alongside the AI draft. The AI made the first pass fast. The lawyer made it right.

### Watch me do it: a prompt with citations and uncertainty

Let me show you how I prompt for this kind of work, in whichever enterprise assistant my organisation has approved for confidential deal data. I start with context: the purpose, the structure of the documents, and the key defined terms. Then the task: extract these specific terms into a table. Then the format: term, value, verbatim quote, clause reference. Then the rules. Cite the clause for every item. If a term depends on a definition elsewhere, follow it and cite that too. List anything you couldn't determine, or were uncertain about, in a separate section. Don't infer values that aren't stated. When I ask for a formula, say a sculpting block, I also ask it to state the period and discounting convention it assumed, and then I test it against a hand calculation before it goes anywhere near the model. And I never paste confidential data room content into a tool that isn't approved: data rooms often carry strict confidentiality obligations.

### Limitations and machine learning

Let's be clear about the limits. AI can misread complex clauses, especially defined terms that cross-reference other documents. It can produce plausible but wrong formulas, such as discounting with the wrong period convention, which we've seen is an easy mistake even for humans. It lacks context about negotiations, relationships and market practice unless you provide it. And outputs can vary between runs, so anything used for a decision must be verified and recorded. On machine learning for operating portfolios: models can forecast output from weather data, predict equipment failures and estimate opex, which can improve CFADS forecasts. But lenders' base cases are typically set by independent advisers using transparent methods. Use ML forecasts as an input to discussion, with documented data, validation and limitations.

### A four-step workflow and common mistakes

Here's a workflow you can defend to any reviewer. First, define the questions and the output format, such as a term matrix with clause references. Second, run the extraction on approved infrastructure. Third, have a qualified reviewer verify each item against the source, marking it confirmed, corrected or unclear. Fourth, store the verified matrix, the AI draft and the reviewer's notes together, so anyone can see what the AI produced and what a person changed. The common mistakes are the reverse: treating AI-extracted terms as verified, using AI to fix a model without understanding the change, losing the audit trail of what AI changed, and assuming AI can replace the independent model audit. It can't, and lenders won't accept it.

### Recap and try this now

Let's recap. AI is most valuable in project finance where there are large document volumes, repetitive checking and pattern detection: term extraction, model consistency checks, scenario summaries, diligence questions and report drafting. It can be wrong in convincing ways, so require citations to clauses or cells, ask for uncertainties, test any drafted formula against a hand calculation, and have a qualified person verify every item. Keep the AI draft, the verified output and the reviewer's notes together as your audit trail. Here's the self-check: could you show a sceptical reviewer exactly where each fact came from? Your try-this-now: use an approved AI tool, or simulate the task manually, to build a key-terms matrix from a sample contract, then verify each term against the source text and mark it confirmed, corrected or unclear.

## Key takeaways

- AI accelerates contract extraction, model review, scenario summaries, diligence Q&A, monitoring and drafting.
- Demand citations, uncertainty lists and testing against known answers.
- Humans verify every term and formula and own conclusions and negotiations.
- Use only approved tools for confidential deal data; AI does not replace the independent model audit.

## Try it

Use an approved AI tool (or simulate the task manually) to build a key-terms matrix from a sample contract, then verify each extracted term against the source text.

- [Previous: Cost management and financial reporting after close](https://optimizeall.com/learn/project-finance-and-financial-modelling/construction-and-operations-reporting)
- [Next: Governing AI in financial analysis](https://optimizeall.com/learn/project-finance-and-financial-modelling/governing-ai-in-finance)
- [All lessons of Project Finance & Financial Modelling](https://optimizeall.com/learn/project-finance-and-financial-modelling)