# Build for the Memo, Not the Demo — Shawn Chan, China Resources Holdings

## Executive summary

This talk contrasts 'demos' (polished, impressive marketing presentations) with 'memos' (deeply scrutinized documents that survive intense financial review). The speaker argues that most AI products are built for the demo—designed to impress for five minutes. However, for real-world applications involving significant capital, the product must pass the 'memo test,' which requires absolute verifiability and accountability. Key architectural requirements include ensuring every claim has a traceable source (provenance), reconciling conflicting data points, and maintaining clear separation between established facts and speculative guesses.

## Key takeaways

- The Demo vs. Memo Test: A demo aims for fluency and confidence; a memo must survive an argument and prove its accuracy under scrutiny. The moment real money is watching, every sentence becomes a memo sentence, meaning there is no safe demo anymore.
- Source Trust Hierarchy: AI systems must differentiate between sources of varying trust levels (e.g., an audited filing vs. a group chat rumor). Treating all sources equally leads to unreliable outputs.
- Data Reconciliation is Mandatory: The system must automatically check that figures agree across all sections of the document (e.g., page one vs. table on page eleven). Failure to reconcile numbers signals a critical flaw.
- Contradictions are Signals, Not Bugs: Instead of smoothing over conflicts (e.g., CEO's number vs. official filing), the AI must surface contradictions. The gap between conflicting numbers is often the most important piece of information.
- Accountability and Provenance: Every claim must be linked directly to its source paragraph (provenance), not just a citation tab. Furthermore, the final decision requires an auditable human sign-off gate.

## Technical details

- Source Provenance: The system must provide a 'receipt' for every claim, linking it directly to the source paragraph and attaching an associated trust level. This is superior to using a simple citation tab.
- Fact/Guess Separation: Guesses (estimates) must be explicitly labeled and kept separate from verified facts within the document structure, preventing the 'guess' from adopting the appearance of a fact.
- Data Integrity Checks: The system should refuse to generate or process a memo if internal figures do not mathematically reconcile across different sections, forcing human review rather than masking the error.

## Practical implications

- Implement a mandatory 'receipt' feature where every generated claim links to its exact source paragraph and includes a trust score.
- Design the architecture to visually separate labeled facts from speculative estimates (guesses).
- Build automated reconciliation checks that flag any mathematical or numerical discrepancy between different data points in the document.
- Ensure the AI is trained not just to answer, but to surface conflicts when sources disagree, rather than smoothing over them.
- Integrate a final, auditable human approval gate (a 'fundable human') into the workflow before outputting any critical decision.

## Topics

AI Product Design, Data Governance, Financial Due Diligence, LLM Reliability, Information Architecture, LinkedIn, Hong Kong, Mainland China, UK, US Markets

Source: https://www.youtube.com/watch?v=tJFjeMBKbIY