Part 16 of 19 in The Patient Owner: Charlie Munger’s Ideas, as Rakesh Jhunjhunwala Might Have Taught Them to an Indian Investor← Part 15Part 17 →
This chapter explores how Rakesh Jhunjhunwala might have used modern AI tools in investment analysis. It considers the potential and limitations of machine learning models in financial decision-making.
- Rakesh would use AI tools like a junior analyst, not for conviction.
- AI should focus on detailed financial data, not broad strategy.
- He would distrust models that always agree with him.
- AI is useful for summarizing data, not making decisions.
- Scores and predictions from AI should be viewed skeptically.
Aniruddha. This is the chapter you cannot answer, so I will ask it anyway. You died in 2022. The machine that reads, summarises, flatters and drafts arrived in force just after. Charlie died in 2023, having warned for decades about misjudgment and about people who want a formula. If you were sitting here now, with a model that can digest ten years of annual reports before the coffee cools, what would you allow it to do?
Rakesh. I would hire it the way I hired a junior analyst, and I would fire it faster. A junior can read. A junior can list related parties. A junior can put the pledge data in a table and the auditor’s resignation in a sentence. A junior cannot be your conviction. The moment the junior becomes the conviction, you are borrowing knowledge, and I have already told you what borrowed knowledge pays.
What he would have switched on
The imagined answer, built from the rules he actually used, is specific.
He would have pointed the machine at the footnotes first, not at the strategy page. Related-party transactions over ten years, changes in auditor, promoter pledges, contingent liabilities, the gap between profit and cash from operations, the words management stopped using. These are inversion engines. A human gets bored on page forty. Boredom is how fraud and frailty survive. A machine does not get bored. Use that. Do not waste it on a poem about the total addressable market.
He would have asked it to argue the other side, in his own harsh register, every time he was tempted. ‘Write the short note. Name the incentive. Assume the promoter is tired or vain, not evil. What breaks?’ He would have distrusted a model that could not be made to disagree with him. Agreement is cheap. A tool that only agrees is a mirror, and mirrors are how people walk into traffic.
He would have used it to expand the circle slowly, not to abolish the fence. A doctor can ask a model to summarise a new regulation, compare a clinical paper with its methods, or turn a conference-call transcript into the three claims that were new. That is apprenticeship. Asking it ‘what should I buy tomorrow?’ is the abolition of the fence, and the fee will arrive.
He would have used it on startup pitch decks the same way. Extract the four sentences: job, payer, reason not to leave, cause of death. Flag every number that is a hope rather than a history. Compare the hiring plan with the revenue plan. List the questions a rude CA would ask. Then switch it off and talk to a customer, because a customer is not in the weights of the model.
What he would have switched off
Aniruddha. The vendors will offer a score. A Munger score. A Jhunjhunwala score. A confidence interval on next quarter’s earnings, dressed as science.
Rakesh. Switch it off. A score is an authority tendency with a progress bar. Charlie’s whole warning about formulas in soft systems applies twice to a formula that speaks in complete sentences. Markets are not physics. A language model is a gifted mimic of the last plausible paragraph. Plausible is not true. In India, plausible has raised more money than true.
He would have been specially wary of the lollapalooza the machine makes easy. Social proof, because the model has read what the crowd says. Availability, because it recalls the vivid case. Over-optimism, because it is trained to be helpful, and helpful often means encouraging. Authority, because it answers in a tone no junior analyst has earned. Stress removal, because it will give you a decision before Friday. Add a notification that praises you for acting, and you have built a casino that speaks like a research department.
Aniruddha. There is a clinical cousin. A model that reads a scan can be a second reader. A model that tells the family what they hoped to hear is a harm. I use machines as a second reader in my own work — to draft, to check, to find what I skipped — and I do not let them sit in the surgeon’s chair. The chair is where responsibility lives. Responsibility that can be delegated to a subscription is not responsibility.
Rakesh. Say that to your angels. The machine may prepare the note. The human signs it. If you cannot explain the note without the machine in the room, you did not use a tool. You used a borrowed brain, and borrowed brains do not refund the loss.
A small practice, if he were reviewing the week
Monday: machine extracts cash versus profit, pledges, related parties, auditor changes, for every holding. Tuesday: machine writes the bear note in two hundred words, forbidden from using the words ‘optionality’ and ‘strategic’. Wednesday: a human calls a customer, a dealer, a former employee, or reads the one primary document the machine summarised too smoothly. Thursday: nothing. Thursday is sit tight, and a machine that needs a task on Thursday will invent one. Friday: only if the obituary facts have moved, a decision. Not a feeling. A fact that was written down last quarter.
Aniruddha. He would have loved Thursday. The vendors will not.
Rakesh. The vendors are not your partners. A tool is a junior. A junior who wants to trade every day is a junior you do not hire.
Evening question. Ask whatever model you use to write the strongest case against your favourite idea, using only primary documents, and to list what it does not know. If you feel attacked, the tool is working. If you feel praised, change the prompt or change the tool.
Frequently asked questions
How would Rakesh Jhunjhunwala use AI in investment analysis?
He would use AI to analyze detailed financial data, like related-party transactions and auditor changes, but not for making final decisions.
What are the limitations of AI in financial decision-making?
AI models can mimic plausible arguments but may not reflect true insights, leading to over-reliance and potential misjudgment.
Why would Rakesh distrust AI models that agree with him?
Agreement is cheap and can create a false sense of security, which is why Rakesh valued tools that could argue against his ideas.
What tasks would Rakesh assign to AI tools?
He would use AI for extracting and summarizing financial data, writing bear notes, and identifying discrepancies, but not for making investment decisions.