Why Investment AI Needs Your Firm's Own Research

Ask the same general-purpose AI to analyse the same large Indian company twice and it can produce two polished answers. Ask two investment firms whether that company belongs in their portfolios and the correct answers may be completely different.

One firm may own a concentrated quality-growth portfolio. Another may run a market-neutral factor strategy. A family office may have an existing operating-business exposure to the same sector. Their mandates, time horizons, constraints and original assumptions are different. Public information alone cannot resolve those differences.

This is why the next useful step in investment AI is not simply a better chatbot. It is an AI-assisted research system that can work with the firm’s own evidence and still show the humans exactly what it did.

Public intelligence is becoming abundant

Annual reports, exchange filings, investor presentations and earnings calls are available to every serious analyst. Foundation models can summarise much of this material quickly. That is a meaningful productivity improvement, but it is not automatically an investment edge.

If ten firms ask a model for the reasons revenue grew, they may all receive similar public facts. The harder questions are local to each firm:

• Did the result confirm the assumption in our approved model?
• Did management deliver the milestone it gave us two quarters ago?
• Does this change the downside case or merely the next-quarter estimate?
• Is the position already adding to an exposure we promised to limit?
• Did we previously allow an exception to our scorecard, and is that reason still valid?

An AI system cannot answer those questions well unless it has access to the correct firm context and a reliable way to connect that context with new evidence.

What “your firm’s own research” actually means

It does not mean pouring an unstructured drive into a model and hoping for a proprietary answer. A useful context layer has several distinct parts.

1. The investment mandate

The mandate defines what the team is trying to achieve and what it is not allowed to do. Universe, liquidity, concentration, sector exposure, valuation discipline and decision rights all belong here.

Without the mandate, “Is this a good stock?” is underspecified. A ₹10,000 crore company can be investible for one portfolio and structurally ineligible for another before anybody debates its moat.

2. The firm’s evidence

Internal notes, channel checks, sector work, meeting notes and annotated filings can contain insights that public-document search does not. But provenance still matters. A dated management statement is not the same as an analyst’s interpretation, and neither is the same as an approved forecast assumption.

The system should preserve who wrote an item, when it was created, which company and period it concerns, and which primary evidence supports it.

3. The assumptions in the model

A forecast is a compact statement of belief. Volume growth, pricing, margins, credit cost, reinvestment, capital intensity and valuation all contain assumptions. They should be individually visible, dated and owned.

When new evidence arrives, AI can help locate the relevant assumption. It should not silently replace it. A proposed change and an approved change are different states.

4. The decision record

The original thesis, variant view, risks, committee challenges and exceptions explain why capital was committed. If these live only in an old presentation or in one portfolio manager’s memory, the monitoring system starts every quarter without the most important context.

A decision record turns “what happened?” into the more valuable question: “what happened relative to what we believed?”

5. Portfolio context

Company research is not a portfolio decision. A business can remain attractive while the position becomes too large, too illiquid or too correlated with the rest of the book. The research system needs the portfolio’s actual exposures and rules before it can frame a material change properly.

6. Feedback from outcomes

Over time, the firm accumulates a distinctive dataset about itself:

• forecasts versus actual outcomes;
• management promises versus delivery;
• scorecard overrides that worked or failed;
• sectors where the team’s assumptions were repeatedly optimistic;
• risks that appeared in the record but were not acted upon.

This is research memory. It is difficult to copy because it is created by the firm’s own decisions across time.

A context layer is not permission for opaque automation

More context can make an AI answer more relevant. It can also make a confident error more persuasive.

That risk is especially important in finance. A 2026 CFA Institute Research and Policy Center study tested 900 prompts across nine models and ten investment scenarios. Identical financial information produced different evaluations when it was framed positively or negatively. Simple instructions to avoid bias did little; structured human balancing was more effective. The practical lesson is not that AI is unusable. It is that context, prompts and human review must be designed deliberately. Read the CFA Institute study on managing LLM bias in investing.

A governed system should therefore preserve four boundaries:

Item	What the system should show
Reported fact	Source, period, basis and publication date
Calculation	Formula, inputs and applicability
Assumption	Owner, rationale, effective date and version
Decision	Approver, date, exception and review condition

When those categories blur, the prose may improve while the research becomes less auditable.

The practical workflow

A useful firm-context workflow looks less like a conversation and more like a controlled research loop:

new filing, result or transcript
              ↓
retrieve the relevant approved thesis and model version
              ↓
identify facts that changed and assumptions they may affect
              ↓
attach citations and propose—not silently apply—updates
              ↓
route material changes to a named analyst
              ↓
record the decision, rationale and next watch condition

The model helps with reading, comparison and retrieval. The team continues to own interpretation and capital allocation.

Questions to ask before connecting private research to AI

An investment firm should test the operating system around the model, not just the quality of a demo answer.

1. Selection: Can a user control which documents and versions are used?
2. Provenance: Does every material claim link back to its evidence?
3. Separation: Are reported facts, calculations, assumptions and decisions visibly different?
4. Permissions: Are private materials separated by firm, team and role?
5. Review: Can proposed changes wait for a named human to approve or reject them?
6. History: Can the team reconstruct what it knew and believed on a prior date?
7. Portability: Can core outputs be exported and independently checked?
8. Disclosure: Is it clear when selected material is processed by an external model provider?
9. Failure states: Does missing evidence appear as unavailable, rather than as a plausible guess?
10. Portfolio link: Can company-level evidence be interpreted against the real book and mandate?

Download the firm-context review checklist to score a current or prospective research system.

Where Altys fits

Altys is built around the idea that AI is an interface to a governed research process, not the process itself.

Its broader research layer connects point-in-time Indian financials, filings, concalls, management guidance, ownership, factors, scorecards, strategies, research grids and monitoring. Core analytical outputs across surfaces such as strategy work, scorecards, GenGrid and alerts can be exported to Excel where applicable, so the team can inspect and verify the work outside the application.

For eligible PMS and family-office pilot users, Altys Vault provides tenant-separated private research storage and search. Company Thesis can draft from explicitly linked Vault versions, preserve citations, check source numbers and move through revise, approve or reject states. These are controlled pilot surfaces, not a claim that private material should be sent indiscriminately to any model.

Altys’s Thesis Model is currently an internal pilot for selected modelling workflows. It keeps assumptions and revisions visible and supports value export for review; it should not be interpreted as universal, unattended model generation.

The aim is straightforward: new public evidence should meet the firm’s existing research, not overwrite it. Analysts should be able to see what changed, which assumption is affected, who decided, and what the portfolio should watch next.

The durable advantage is the learning system

Models will improve. Public-document summaries will get cheaper. Interfaces will converge.

The harder asset to reproduce is a firm’s connected history of evidence, assumptions, decisions and outcomes. That history is not valuable because it is private by itself. It is valuable because it helps the team ask better questions, challenge repeated mistakes and maintain continuity when people or market regimes change.

If everyone has access to capable AI, the winning investment team is unlikely to be the one that generates the longest memo. It will be the one that can turn its own accumulated judgement into a process that remains sourced, testable and accountable.