AI Engineer World's Fair 2026

Build for the Memo, Not the Demo — Notes from 200 Investment Committees

Shawn Chan24:23

Loading slides…

Read the talk

Build for the Memo, Not the Demo

Financial AI earns trust when a skeptical reader can trace claims, reconcile numbers, challenge assumptions and identify the person accountable for the decision.

From a talk by Shawn Chan

Would you trust this paper with $100 million?

Before his company commits $100 million, Shawn Chan has to answer a deceptively simple question: does he believe the paper in front of him? For fifteen years, a confident person handed him that recommendation. Now, sometimes, a chatbot supplies it. The grammar is better, the formatting is cleaner, and follow-up questions never provoke defensiveness. None of those qualities establishes that the recommendation is right.

Slide compares a confident friend at dinner with an AI system at 2am, joined by “I'm 100% sure” and “It was wrong.”Open full source frame
Confidence and correctness are not the same skill.

A financial AI product must survive scrutiny, not merely create a good first impression. A demonstration can impress a room for five minutes; an investment committee is a room whose job is to resist being impressed. The same evidence discipline that makes a product credible also helps its founders raise money. That distinction becomes urgent when a CEO sees a compelling demo, a major customer brings in compliance, or the next funding round is three months away.

0:150:30
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:15 · section reference included

Trust in the number

Chan’s perspective comes from cross-border mergers, IPOs and strategic investments spanning Hong Kong, Mainland China, the UK and the US. He reports attending about 200 investment committee meetings and reading hundreds of founder and banker decks—including 47-slide seed decks. The recurring lesson is that a number does not get a deal approved on its own. Trust in that number does.

Slide titled “Fifteen years on the other side of the table” shows three experience statistics and the takeaway “Money doesn't follow intelligence. Money follows trust.”Open full source frame
Fifteen years, about 200 investment committees, and hundreds of pitch decks.

Early in his career, Chan watched a polished, expensive banker present beautiful slides containing confidently wrong numbers. A senior person asked, “Where does this number come from?” The banker stalled. The presentation had no usable answer to the question that mattered most. Chan approaches AI from that side of the table: as the buyer and evaluator deciding whether to trust a system and the founder selling it, rather than as its builder.

2:493:10
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:49 · section reference included

Two different machines

A demo and a memo operate under different conditions:

DemoInvestment memo
InputOne clean documentFilings, transcripts, broker notes, spreadsheets, call notes
EvidenceA convenient exampleHundreds of pages, often in conflict
OutputA fluent answerA defensible recommendation
SuccessImpress the audienceSurvive an argument

The memo is the document a real committee reads before real money moves. Its inputs include rushed notes from last Tuesday’s call alongside formal filings. Disagreement is part of the working material, not an exceptional condition that can be removed from the demonstration.

The everyday comparison is a phone summarizing a long email versus that same phone standing before a bank to defend your mortgage application. A plausible summary may satisfy the first task. The second requires the system to be right and to prove why its answer deserves belief.

5:466:03
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:46 · section reference included

When the demo becomes a financial event

In Chan’s February 2023 example, a large technology company promoted an AI assistant with an incorrect answer about a space telescope. Chan reports an approximately 8% one-day share-price decline and roughly $100 billion in lost market value. He connects the selloff to the unchecked sentence; contemporaneous reporting on Alphabet’s decline also identifies competitive and advertising concerns, so the sentence should not be treated as the sole established cause.

The missing discipline is familiar to a junior analyst: establish where the claim came from and whether anyone checked it before publication. Once money is watching, even a promotional sentence can face the memo’s standard of scrutiny.

Slide displays approximately $100,000,000,000, describes market value gone in one day, and asks where the claim came from and whether anyone checked it.Open full source frame
“The most expensive typo in history”: an unchecked sentence and roughly $100 billion in market value.

Inside a product, that scrutiny becomes six questions: which source deserves belief, whether the numbers agree, whether contradictions remain visible, whether a statement is a fact or a guess, how quickly it can be substantiated, and whose name is attached to the decision.

7:197:39
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:19 · section reference included

Relevance is not source authority

The first failure happens before an answer is written: retrieval selects evidence without adequately distinguishing its authority. Chan uses three comparisons:

  • Audited filing: an accountant speaking under oath.
  • Analyst note: a confident friend at a party.
  • Internal email: something overheard in an elevator.

A passage can be close to the question without deserving the same trust as an audited figure. A retrieval system that treats those properties as interchangeable can turn a persuasive assertion into apparent evidence.

In an anonymized example, Chan watched an expensive AI tool quote a rough estimate from a group chat six months earlier. The actual audited number was nearby in the filings—three rows away, in his account—but the system favored the confidently phrased chat message. The resulting answer looked authoritative because the source distinction had disappeared. Source trust must travel with the retrieved claim.

9:4510:04
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:45 · section reference included

Numbers must agree with the document and reality

Chan’s failed-memo example gives revenue growth as 18% on page one and 17.4% in a table on page eleven. The difference is 0.6 percentage points, but its significance is larger than its size. If nobody checked the easy arithmetic, a committee has reason to doubt whether anyone checked the harder assumptions. In his account, the memo failed because of what the discrepancy revealed about the process.

The larger-scale example is an American real-estate business that used an algorithm to buy houses. Chan describes around half a billion dollars written off, closure of the unit and a quarter of staff let go. Those rounded figures require care: the matching Zillow Offers episode involved forecasts as well as realized charges; Zillow’s subsequent shareholder letter reports $405 million in combined third- and fourth-quarter inventory write-downs, a category distinct from total business losses.

Chan characterizes the failure as insufficient supervision rather than a lack of model intelligence. His engineering requirement is continuing reconciliation with reality after launch. Internal agreement between two pages is one control; checking whether the system’s numerical expectations remain aligned with actual outcomes is another. Fluency and confidence provide neither.

11:2211:28
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:22 · section reference included

Put the disagreement in front of a human

A CEO’s growth figure on an earnings call can disagree with the official filing. That gap is valuable diligence material: it creates a question worth investigating. A system optimized to produce smooth prose may instead select whichever version reads better, leaving the reader unaware that a conflict existed.

Chan recounts a discrepancy caught only because someone happened to have both documents open. “Luck is not a control.” The builder’s responsibility is to preserve the disagreement and put it before a human. Automatically choosing a preferred answer is not equivalent to resolving the underlying issue; it can simply erase the evidence that a review is needed.

13:1913:45
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:19 · section reference included

Keep estimates from becoming facts

Consider the sentence: “The company will likely receive approval next quarter.” It describes an estimate, not a completed event. Committees need to evaluate such judgments separately from the facts supporting them. When a system blends both into seamless prose, readers lose the boundary between what is established and what they are being asked to believe. A $100 million decision can then become an assessment of the document’s overall tone.

Chan watched an expectation of approval soon become a statement that approval had been received across three memo drafts. Nobody needed to deliberately falsify the document: successive rewrites removed the uncertainty until the estimate looked like a fact. The approval did not arrive on schedule.

The control is a durable uncertainty label. A tag or color must remain recognizable when the content is copied into someone else’s slides three weeks later. For the approval example, the pending status has to remain attached to the claim through editing and reuse; better wording alone cannot turn an expected approval into a received one.

14:5915:08
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:59 · section reference included

The click-through is the product

A correct claim still has limited practical value if nobody can find its source. Chan illustrates the danger with the New York legal brief containing six nonexistent court cases: convincing citations gave invented authorities the appearance of verifiability. The incident corresponds to Mata v. Avianca.

The lawyer also asked the chatbot whether the cases were real and received reassurance. Chan places that exchange before filing; the court’s opinion records conflicting accounts of its timing, with the later declaration placing it after the first show-cause order. The verification failure is the same: asking the generator to endorse its own output does not independently establish that the cited authority exists. The court imposed sanctions on the attorneys and their firm. Realistic names, page numbers and formatting had made the unsupported material easier to believe.

The verification path is part of the product. Chan proposes a 30-second test: when someone challenges a sentence, one click should reach the exact source paragraph. The alternative is opening seven browser tabs and scrolling while a room waits. He has been that person in a real meeting. A citation list that looks complete but leaves the reviewer searching has failed the practical test.

16:4716:53
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:47 · section reference included

Someone has to sign

An airline chatbot told a grieving customer to buy a full-price ticket and claim a bereavement discount afterward. The error concerned retroactive eligibility, not whether bereavement fares existed. The customer took the airline to a tribunal. The corresponding Air Canada case shows how confidently delivered advice can become an organizational liability.

Chan recounts the airline’s rejected position as treating the chatbot as a separate legal entity responsible for its own actions. The tribunal awarded damages; responsibility did not disappear because the information came through software.

For a financial decision, the architectural consequence is a findable human signer. The system should support that person’s review and responsibility, not make it harder to determine who owns the result. Build around the accountable person.

19:1119:21
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

19:11 · section reference included

Five requirements before a live deal

These failure stories become concrete acceptance requirements. Before allowing a vendor’s system near a live deal, Chan would demand:

  • Receipts attached to claims. Each sentence links directly to its source paragraph and carries the source’s trust level. A citation tab at the end is insufficient.
  • Visible facts and estimates. A reader can immediately distinguish established information from someone’s best judgment.
  • Automatic numerical agreement. The system refuses to release a memo with inconsistent figures, rather than depending on a person catching them at two in the morning.
  • Explicit contradictions. When sources disagree, the system raises the issue instead of silently selecting the friendlier answer.
  • Logged human approval. The record identifies who reviewed the memo, what changed and when the reviewer signed. That record is the audit trail.

None of these requirements asks for a smarter model. Chan treats them as problems of evidence handling, controls and honesty. The competitive test is whether a tired, skeptical finance professional can trust the output at eleven at night without opening seven tabs—not whether the underlying model gained a few benchmark points.

Slide titled “The fix: a buyer's spec” lists five numbered requirements, ending with a logged human approval gate.Open full source frame
A buyer's spec: receipts, separate facts and guesses, consistent numbers, surfaced contradictions, and logged approval.
20:5621:10
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

20:56 · section reference included

Your pitch becomes somebody else’s memo

After the fundraising pitch ends, the deck becomes an internal investment memo. Someone on the other side of the table writes down the case for investing and checks the numbers the founder said aloud against the data room. The evidence discipline expected of a financial AI product therefore applies to the company selling it. Your presentation has to remain credible when you are no longer in the room to explain it.

23:1323:24
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

23:13 · section reference included

Resources

Transcript download
Full transcript

Loading transcript…