What belongs behind an AI assisted investment memo

A well-written memo can hide a weak research process. A market estimate appears precise, a competitor summary looks complete and customer demand seems supported by several sources. But the sources may repeat the same announcement, use different definitions or describe plans rather than completed activity.

AI can help organise this material. It cannot make the underlying evidence stronger merely by summarising it. For venture investors assessing APAC and emerging-market companies, GAT recommends treating provenance—the history of a claim—as part of the investment work.

Record where each consequential claim came from

The W3C PROV Data Model describes provenance through the entities, activities and agents involved in producing information. A venture firm does not need to implement that technical specification to use its central idea: a claim should retain a traceable relationship to its source and the work performed on it.

For a material memo claim, record the source, its date, the relevant passage or data field, who reviewed it and any transformation applied. A currency conversion, aggregation or estimate should be distinguishable from a number directly reported by a company.

This is especially important when terms such as revenue, bookings, annual recurring revenue and transaction value appear together. A model may present them fluently without resolving the differences. The analyst should.

Separate source quality from source quantity

Five articles citing one press release provide a different evidential base from five independently conducted interviews. Count the underlying evidence, not just the links.

An official filing may establish a reported financial figure. A customer interview may illuminate implementation quality. A founder presentation may explain strategy while remaining a management assertion. Each can be useful if the memo makes its role clear.

For private-company diligence, some of the strongest evidence may be confidential. Preserve access controls and authorised use. A claim ledger does not require placing sensitive records into a public AI service or giving every team member unrestricted access.

Give uncertainty a visible location

A memo should show what is verified, what is reported by management and what remains an analyst judgment. Those distinctions can sit beside the relevant conclusion rather than disappear into a disclaimer at the end.

Consider an illustrative startup selling into three countries. Signed customer contracts, active deployments and a planned distributor launch should not all become “regional traction.” The memo can describe each milestone accurately and explain what remains unproven.

Record contradictory evidence too. If a customer praises the product but reports heavy implementation effort, the investment discussion needs both observations. Removing the difficult detail makes the summary cleaner and the decision weaker.

Make human accountability specific

In GAT’s interview with Global Ventures founder Noor Sweid, she described AI-assisted sourcing and portfolio workflows while keeping investment decisions and accountability with people. Those comments are her account of her firm’s practices, not proof that every investor operates the same way.

For a fund adopting AI, “human reviewed” should identify a responsibility. The analyst checks claims and definitions. A reviewer challenges conclusions and missing evidence. The investment committee decides whether the remaining uncertainty fits the proposed investment.

Keeping the prompt and model version can assist internal review, but a model name is not a source for a business fact. The evidence must still lead to a document, record or identifiable human source.

Preserve the decision record after approval

Information changes after a memo is circulated. Keep versions so the firm can distinguish what it knew at approval from what it learned later. Link corrections to the affected claim and conclusion.

The aim is not to eliminate judgment from investing. It is to make the basis of judgment inspectable. AI-assisted research becomes more useful when the firm can explain which evidence supports a decision, which assumptions remain and who accepted them.