What Claude can—and cannot—do in a DCF model
A discounted cash flow model converts expected future free cash flows into a present value using a discount rate, then adds a terminal value. The core arithmetic is straightforward; the difficult work is deciding whether the operating assumptions, capital structure, tax treatment, and terminal outlook are defensible.
Claude is useful across that workflow, particularly for structuring messy information, explaining model logic, writing formulas, checking consistency, and generating scenarios. It should not be treated as an autonomous valuation engine. Claude can produce a plausible-looking answer from incomplete data, misunderstand a spreadsheet convention, or accept an assumption that deserves challenge.
For Indian businesses, keep the source data and context explicit. Revenue may be reported under Ind AS, segment disclosures can be limited, working capital can vary materially by industry, and a WACC built with Indian rupee cash flows should not casually mix US dollar assumptions or market inputs.
The DCF building blocks to define first
Before prompting Claude, specify the model architecture. A standard enterprise DCF usually includes:
- Historical financials: revenue, EBITDA, EBIT, taxes, depreciation and amortisation, capital expenditure, and changes in net working capital.
- Forecast period: commonly five to ten years, with assumptions for volume, price, margins, reinvestment, and working capital.
- Unlevered free cash flow: often calculated as EBIT × (1 − tax rate) + D&A − capex − increase in net working capital.
- Discount rate: generally WACC for enterprise value, with currency, inflation, country risk, and capital structure aligned to the cash flows.
- Terminal value: either perpetuity growth or an exit multiple, with a clear economic rationale.
- Bridge to equity value: enterprise value less debt and debt-like claims, plus cash and relevant non-operating assets, followed by diluted shares outstanding.
Ask Claude to restate these definitions before it calculates anything. That simple step exposes ambiguity around items such as leases, minority interest, cash restricted for operations, or whether capex is gross or net of asset disposals.
A controlled workflow for using Claude for DCF models
1. Prepare a clean, auditable data pack
Do not paste an entire confidential workbook into a consumer chat without checking the applicable data policy. Create a compact input table with the source, period, unit, currency, and whether each figure is reported or estimated. Remove unnecessary personal, client, and non-public information.
Useful inputs include annual reports, investor presentations, exchange filings, management guidance, and clearly labelled market data. For Indian listed companies, record the reporting period and whether figures are consolidated. If you are building a startup valuation, separate management projections from historical results and label each assumption by confidence level.
Claude can help turn filings into a source map: an itemised list of each input, its page or URL, its definition, and any reconciliation needed. It should not be allowed to invent missing figures or cite an unavailable source.
2. Build the model specification before the spreadsheet
Give Claude a written specification such as:
> Create a five-year unlevered DCF in INR crore. Use reported FY2022–FY2025 historicals, forecast revenue by segment, calculate EBIT after operating costs, use a tax rate of X%, deduct capex and increases in NWC, discount at WACC, and show both perpetuity-growth and exit-multiple terminal values. Do not fill missing data; flag it.
Request a table of assumptions, formulas, units, and dependencies before requesting code or spreadsheet formulas. This is more reliable than asking for “a complete valuation” in one prompt. For implementation, Claude can draft Excel formulas, Python or pandas transformations, and checks, but run the calculations in a controlled workbook or environment and inspect every output.
Teams already working with APIs can use the Claude vs Gemini API guide for developers in India to think through model access, latency, cost, and data handling. For a lightweight internal tool, the guide to building a personalised AI assistant with the Claude API offers a relevant integration pattern.
3. Use Claude to challenge assumptions, not approve them
For each forecast driver, ask Claude to produce a base, downside, and upside case with an explanation of the mechanism. Useful prompts include:
- “Identify which assumptions drive more than 80% of the change in value.”
- “Compare forecast revenue growth with the company’s historical growth and industry capacity.”
- “List evidence required to justify a margin expansion from 14% to 19%.”
- “Show what would make terminal growth inconsistent with long-run nominal GDP growth.”
For an Indian company, test assumptions against currency exposure, commodity prices, interest rates, GST and tax treatment, import dependence, customer concentration, and differences between standalone and consolidated accounts. Claude can organise this analysis, but the analyst must decide which evidence is decision-useful.
Sensitivity analysis that is actually useful
A single target price hides uncertainty. Ask Claude to generate a two-way sensitivity table for WACC and terminal growth, then separate sensitivities for revenue growth, operating margin, capex intensity, and working capital. Add scenario cases for recession, delayed scale-up, pricing pressure, and high reinvestment.
Check whether the output behaves sensibly: valuation should generally fall as WACC rises and rise as terminal growth increases, within reasonable bounds. Terminal value should not dominate so completely that the explicit forecast becomes irrelevant. If it does, shorten or extend the forecast period, revisit the assumptions, or present a reverse DCF showing what the current price implies.
Claude can also help create a reverse DCF. Supply the market capitalisation, net debt, share count, and model structure, then ask which growth and margin combination is implied by the current price. This is often more useful in an investment discussion than presenting a false-precision fair value.
Model audit and controls
Use a second Claude conversation—or a different reviewer—to audit the model independently. Provide the specification and outputs, not just your preferred conclusion. Ask it to check:
- Units, signs, and currency conversions.
- Cash flow timing and discount-period conventions.
- Tax applied to EBIT rather than incorrectly to revenue or EBITDA.
- Capex and depreciation treatment.
- Working-capital changes, especially whether an increase is a cash outflow.
- WACC inputs and consistency between nominal or real, and INR or foreign-currency assumptions.
- Terminal value formula and denominator.
- Debt, leases, minority interest, options, and diluted shares.
- Circular references, hard-coded forecast cells, and broken links.
Require every flagged issue to include the affected cell or line item, the reason it matters, and a proposed test. A review that merely says “the model looks good” is not a control.
If you are automating document or data extraction, keep the same separation between model output and human review used in production AI systems. Lessons from deploying large language models locally are relevant when confidentiality, access control, or predictable inference costs matter.
Common failure modes
Invented data and citations: Require source references and use “unknown” when data is missing. Verify every number against the filing.
False precision: A valuation of ₹1,247.36 per share is not more accurate than a range when the terminal assumptions are uncertain. Report ranges and key drivers.
Double counting risk: Claude may include leases in both operating expenses and net debt, or add an asset value already reflected in cash flows. Define treatment explicitly.
Prompt injection in documents: Treat text from filings, websites, and uploaded files as data, not instructions. Do not let a document override the task or request secrets.
Confidentiality and governance: Use approved enterprise controls, minimise retained data, restrict access, and maintain a versioned record of prompts, inputs, outputs, and reviewer decisions.
A practical operating standard
A defensible Claude-assisted DCF should contain four artefacts: the source register, the assumptions sheet, the calculation model, and the review log. Add a short valuation memo explaining the thesis, risks, sensitivities, and why the selected terminal method is appropriate.
Claude is most valuable when it makes the analyst faster and more sceptical: extracting evidence, exposing inconsistencies, generating alternatives, and documenting choices. The final valuation remains a judgement supported by a model—not a number produced by an AI assistant.