ChatGPT and Claude for Indian Stock Analysis: A Safe Workflow

ChatGPT and Claude can be excellent assistants for Indian stock analysis. They can also produce a polished report containing a number that never appeared in any filing.

The difference is the workflow.

General language models are strongest when you give them the evidence, define the task and keep financial calculations outside the prose engine. They are weakest when asked to recall current company facts or deliver a ready-made investment verdict.

What to provide before asking for analysis

For a serious company review, collect:

• the latest annual report
• the latest quarterly result and notes
• recent investor presentations
• three to five concall transcripts
• a source-linked financial-history table
• relevant shareholding and credit-rating disclosures
• your existing thesis or the questions you want answered

Label every table with company, period, unit and consolidated or standalone basis. An unlabeled spreadsheet invites the model to infer the wrong meaning.

If the task is historical, include only documents that were public by the relevant date.

A safe five-pass workflow

Pass 1: Build the business map

Ask the model to explain how the company makes money using only the supplied documents.

Prompt:

Using only the attached annual report and investor presentation, map the company’s operating segments, customers, revenue drivers, major costs, capital needs and stated risks. Cite the page for every material claim. Separate reported facts from your interpretation.

Review the result against the segment note and management discussion.

Pass 2: Identify the important drivers

Ask for a driver tree rather than a generic SWOT analysis.

Prompt:

Convert the business model into a driver tree. For each segment, show the most likely volume, price, cost and capital variables. Identify where each driver is disclosed and mark any driver that is not directly observable.

This produces a structure that can feed a financial model.

Pass 3: Compare management language

Provide several concalls and ask what changed.

Prompt:

Compare these calls chronologically. Show changes in demand, pricing, margins, capex and guidance. Quote or cite the relevant passage. Distinguish a genuine change from repeated wording, and flag questions the latest call leaves unanswered.

Do not ask only for sentiment. “Positive” and “negative” compress too much nuance.

Pass 4: Audit the financial story

Give the model a verified financial table rather than asking it to invent one.

Prompt:

Using the attached sourced table, identify which movements deserve investigation: revenue versus receivables, profit versus operating cash flow, debt versus cash generation, other income, margins and share count. Do not recalculate or fill missing cells. For every observation, name the exact rows and periods used.

The model should generate questions, not allegations.

Pass 5: Challenge the thesis

The most valuable prompt asks the model to disagree.

Prompt:

Here is my investment thesis. List the three assumptions that carry most of the conclusion. For each, find evidence in the supplied documents that supports it, evidence that contradicts it, and the next measurable disclosure that could falsify it. Do not recommend buying or selling.

This moves the model from summary toward useful research assistance.

Keep maths in the spreadsheet

ChatGPT and Claude can help write formulas, check a model structure and explain why statements do not balance. The spreadsheet should remain the calculation engine.

A robust division of labour:

Job	Best owner
Source financial history	Filing or trusted data platform
Calculate ratios and forecasts	Code or spreadsheet formulas
Explain documents and movements	Language model
Choose assumptions	Analyst
Approve the conclusion	Analyst or investment process

If the model suggests an assumption, place it in a clearly labelled input cell. Do not let a generated figure become an invisible constant inside a formula.

Common failure modes

Confident number recall

Ask a chatbot for a company’s latest revenue without a source and it may retrieve a stale value, mix periods or simply generate a plausible answer. Require a document citation or supply the number yourself.

Consolidated and standalone mixing

The same company can report both. A model reading multiple documents may combine them unless the basis is explicit.

Quarter and annual mixing

One PDF may show a quarter, nine months and the previous full year on the same page. Verify period duration before comparing growth.

Secondary-source drift

Web search can surface aggregator pages with different definitions and update dates. For core financial facts, prefer company and exchange documents.

Prompted certainty

“Tell me whether this is a good stock” invites a verdict unsupported by the investor’s objectives, valuation method and portfolio. Ask for evidence, scenarios and failure conditions instead.

False completeness

A well-structured memo may still omit the one disclosure that matters. Keep an evidence checklist and inspect source coverage before trusting the summary.

When a general chatbot is enough

ChatGPT or Claude works well when:

• you already possess the documents
• the task is explanation, comparison or drafting
• you can verify the answer
• the company set is small
• you do not need continuous monitoring

When you need a research platform around it

A dedicated system becomes useful when you need:

• reliable Indian financial data across many companies
• source links for every figure
• point-in-time history
• consistent calculations and sector-aware metrics
• document search across filings and concalls
• screening and backtesting
• portfolio context and monitoring
• retained research memory

The intelligence of the model is only one layer. The rest of the stack determines whether the output is dependable.

Our broader guide to finance AI in India explains that stack, while AI stock analysis in India covers the full company workflow.

How Altys works with the same principle

Altys supplies the Indian research context around the language model: filings, concalls, guidance, shareholding, financials, factors, macro data, screening, modelling, forecasting, backtesting and monitoring.

Code performs the calculations and preserves point-in-time history. AI reads the documents, compares evidence and cites the source. The analyst controls assumptions and capital allocation.

That does not make the model unimportant. It puts it in the role where it is most powerful.

The short answer

Use ChatGPT and Claude as fast research associates, not as databases or fund managers.

Give them primary evidence. Constrain the task. Demand citations. Keep maths deterministic. Ask for contradictions. Verify every decision-critical number.

Used that way, a general model can save hours without quietly taking ownership of the judgement.