Data Room Check: Opportunity Analysis

Date: 2026-10-04 Owner: Alexandre Cela Status: Go (approved by Alex 2026-10-04)

Problem Statement

A family office (a company that invests one family's money) must decide on a deal in 2 to 3 weeks. A wrong number in the seller's slides, or a risk in an unopened file, can change that decision. Today, nobody may notice either one.

Current Situation

The seller shares a folder of documents, called the data room. An analyst reads it by hand and notes findings in a spreadsheet. They often rely on the lead investor's summary, and that investor wants the deal to close.

A concrete case (made up for the demo). Fintech A's slides say its largest customer is 12% of revenue. The customer list in the data room shows 31%. If that customer leaves, almost a third of revenue goes. Today, the analyst catches this only if they open that file.

Proposed Solution

An AI checker for the data room. A language model (an AI that reads and writes text) does the reading. Code checks its work. 1. Verifies every slide claim. The AI finds the matching figure in the data room. It marks each claim confirmed, contradicted or not found. 2. Hunts for every listed red flag in every file. The list is fixed: change of owner clauses, legal claims, contracts ending soon and more. The AI also notes other unusual facts. 3. Shows the lineage of every result. Each result carries its file, page and an exact quote. Code confirms the quote is on that page, and the numbers match, before the result is shown. Results that fail are labelled unverified.

The bar it must meet: 100% of results with verified lineage, and at least 95% of flags real. Below either, it doesn't ship.

It states facts and never recommends. The analyst and the associate (who signs off the analyst's work) decide.

Expected Impact: money

The family office makes money by paying the right price and avoiding losses it could have seen. Every number below is made up for this example.

• Paying the right price. The slides say revenue of $38.0M. The audited accounts say $36.2M. At a price of 4 times revenue, that gap is $7.2M of value paid for but not there. On a 10% stake, the family overpays $0.72M.
• Asking for protection before signing. A $2.1M legal claim, and a loan that must be repaid if owners change, surface before the deal closes. The family can ask for a lower price or for the seller to cover those costs.
• Fewer hours per deal. Analyst and associate hours fall, so the same team can review more deals. The hours saved are unobserved.

Problems, hypotheses and success

One row per atomic problem: one job, one hypothesis, one metric. "Today" values and impacts are assumed, not measured.

#	Problem	Today	Hypothesis	Success metric	Expected impact
1	Wrong numbers in the slides reach the decision	Only numbers the analyst has time for get checked	If every slide number is checked against the data room, then wrong ones get caught, because each mismatch shows both pages	Planted wrong slide numbers flagged: 100%	Fewer deal decisions made on a wrong number. Analyst hours re-checking slides: down (unobserved)
2	Risks sit in files nobody opened	Only opened files get read	If one warning-sign list runs over every file, then fewer risks are missed, because no file waits for the analyst's time	Planted risks found: 100%, including 2 or 3 not on the list	Fewer risks discovered after the deal closes. Files skipped for lack of time: 0
3	The associate can't see what was checked	Spreadsheet notes, no page links	If every result shows its page and a quote that code matched to that page, then the associate trusts it without re-reading, because checking is one click	Results whose quote is found on the cited page: 100%	Associate sign-off becomes spot-checks, not a re-read. Sign-off time: down (unobserved)

Must not get worse: claims shown as fact that failed the check: 0.

Not tackled here: - Retyping numbers for the financial model: needs a set list of fields. - Trusting the lead investor's summary: needs a sample summary to check. - Asking any free question: a search can't prove a fact is absent, so it can't support "nothing was missed." - Turning findings into a lower price: this is the investor's judgment.

How It Could Be Delivered

• Data: made up. Fintech A as an older company with loans: 10 slides, 13 files, 9 known issues in an answer key.
• Finding evidence for a claim, three ways compared: one model call with the whole data room; RAG (search passages by meaning, then answer); PageIndex (the AI walks each file's table of contents). The red flag hunt reads every page, with no search.
• Results are computed once and stored, so the demo runs free. Details are in the PRD.

Value Equation (Hormozi)

Factor	Analyst (uses it)	Associate (signs off)
Dream outcome	"I checked everything and I can prove it."	"I can sign off without redoing the work."
Perceived likelihood	Low to medium. Investors distrust AI answers they can't check.	Low. They'll trust it only after spot-checking pages.
Time delay	Results minutes after the folder loads (unobserved).	Same day as the analyst's run.
Effort & sacrifice	Upload the folder, review flagged rows, learn one new screen.	Click through flagged rows instead of reading files.

Verdict. Belief is the weak factor for both. The cheapest lever is to show the source page and quote beside every result. Another is to say "not found" instead of guessing. The outcome doesn't need improving.

Recommendation

Go, as a portfolio prototype. The problem is plausible, the demo needs only fake data, and it adds a trust story Deck Rank and BDC Footnote Tape don't show. A user test with 2 analysts would make it a product case, not just a demo.

Sources: conversation with Alex 2026-10-04 (brief confirmed), ~/Documents/Company Tracker/company_scorer/demo/make_sample_decks.py, outputs/portfolio/deck-rank/linkedin.md, outputs/portfolio/bdc-footnotes/linkedin.md

Data Room Check

Stage: Team Kickoff (lightweight AI PRD) Last Updated: 2026-10-04 Owner: Alexandre Cela Status: Draft Opportunity analysis: outputs/opportunity-analyses/data-room-check-opportunity.md (Go, 2026-10-04)

Disclaimer. Fintech A, its data room, its people and every number in this document are fictional, created for this prototype.

---

Hypothesis

A family office (a company that invests one family's money) has 2 to 3 weeks to decide on a deal. A wrong number in the seller's slides, or a risk in an unopened file, can change that decision. Today, nobody may notice either one.

Atomic problems (carried from the opportunity analysis). "Today" values and impacts are assumed, not measured.

#	Problem	Today	Hypothesis	Success metric	Expected impact
1	Wrong numbers in the slides reach the decision	Only numbers the analyst has time for get checked	If every slide number is checked against the data room, then wrong ones get caught, because each mismatch shows both pages	Known wrong slide numbers flagged: 4 of 4	Fewer deal decisions made on a wrong number. Analyst hours re-checking slides: down (unobserved)
2	Risks sit in files nobody opened	Only opened files get read	If one warning-sign list runs over every file, then fewer risks are missed, because no file waits for the analyst's time	Known risks found: 6 of 6, including 2 not on the list	Fewer risks discovered after the deal closes. Files skipped for lack of time: 0
3	The associate can't see what was checked	Spreadsheet notes, no page links	If every result shows its page and a quote that code matched to that page, then the associate trusts it without re-reading, because checking is one click	Results whose quote is found on the cited page: 100%	Associate sign-off becomes spot-checks, not a re-read. Sign-off time: down (unobserved)

Must not get worse: claims shown as fact that failed the check: 0.

Supporting evidence: none from users. The problem comes from how this kind of investing works. No analyst has confirmed it.

---

Strategic Fit

Why this. Every result is checked by code before it's shown. Deck Rank ranks decks. BDC Footnote Tape turns lender reports into a spreadsheet. Neither one checks a seller's claims against evidence.

Why now. It extends the private capital AI offer to a new buyer: the family office, through its analyst and associate.

Impact sizing. Not done yet: no deal volume or time-per-deal data exists. Alternatives considered are in Appendix B.

---

Expected Impact

If this ships, the analyst gets every slide number marked confirmed, contradicted or not found, and every file checked for the same warning signs, each with a page. The associate signs off from that record. If it doesn't ship, coverage still depends on which files the analyst had time to open. Time saved is unknown. A test with 2 analysts on one data room would measure it.

---

Value Equation (carried from the opportunity analysis)

Factor	Analyst (uses it)	Associate (signs off)
Dream outcome	"I checked everything and I can prove it."	"I can sign off without redoing the work."
Perceived likelihood	Low to medium. Investors distrust AI answers they can't check.	Low. They'll trust it only after spot-checking pages.
Time delay	Results minutes after the folder loads (unobserved).	Same day as the analyst's run.
Effort & sacrifice	Upload the folder, review flagged rows, learn one new screen.	Click through flagged rows instead of reading files.

Verdict. Belief is the weak factor. The cheapest lever is to show the source page and quote beside every result. Another is to say "not found" instead of guessing.

---

Non-Goals

• Free questions about anything. A search can't prove a fact is absent.
• Advice or a lower price. The decision is the investor's.
• Hundreds of documents. This data room has 13 files. Search is built and measured (see trade-offs), but behaviour at scale isn't tested.
• Privacy controls and logins. Out of scope for the prototype.

Trade-offs made: - Three ways to find evidence for a slide claim, compared on the same answer key. - A. One model call. The whole data room goes to the model at once. No search. It can't skip a page, and it stops working at roughly a few hundred pages. This is the baseline. - B. RAG (search, then answer). Pages are cut into passages. Each passage is turned into numbers that capture its meaning (embeddings). For each claim, the closest passages by meaning and by keyword go to the model. - C. PageIndex (follow the table of contents). Each file becomes a tree of section headings. The model reads the headings and walks down to the pages that matter. No embeddings. - B and C are more than 13 files need. They're built to learn both methods and to measure what each misses against A. - The red flag hunt never uses search. It reads every page of every file, because search can't prove a file was covered. - Checked once, stored. Each data room is checked when it loads. Reopening shows stored results and costs nothing.

---

Success Metrics

Targets are in the atomic problems table. Two extra metrics: - Correct slide numbers wrongly flagged: 0 of 6. - Evidence pages found, per method: the share of answer-key evidence pages that A, B and C each find, shown side by side. Each method also reports how many of the 4 known wrong slide numbers it flags.

The answer key lists every issue in the data room, from a full manual read before the first run. It's stored apart and never shown to the model.

Kill criteria: after 2 attempts to improve the AI's instructions, we stop and don't publish if either is true: - Lineage below 100%. Some results lack a file, a page and a quote that code finds on that page. - Precision below 90%. Of everything the tool flags (wrong slide numbers and red flags), fewer than 90% are real issues in the answer key. (Lowered from 95% by Alex on 2026-10-05.) - Recall below 100%. The tool misses any of the 4 wrong slide numbers or 6 red flags in the answer key. (Added 2026-10-05.)

Why: an unverified result sends the associate back to the files, false flags teach the analyst to ignore flags, and a missed issue is the problem this tool exists to solve.

Test set caveat: the prompts were tuned on this one data room (3 revisions, all logged in Langfuse). A pass here shows the method works on what it was tuned on. A second data room the prompts have never seen is needed to show it generalises.

---

Rollout Plan

1. Data room and answer key. Passing: 13 files load, and every known issue sits on a known page.
2. Checks and scoring. Passing: the metrics table is filled from a logged run, for A, B and C.
3. Interface and case study. Passing: the interface loads in a browser and /portfolio-case-study goals pass.

The PM reviews each step before the next one starts.

---

AI Behavior Contract

Principle: the model reads and drafts, code checks and computes, the person decides.

Dimension	Specification
Primary tasks	Extract slide numbers. Find the matching evidence. Run the warning-sign list. Note other unusual facts.
Inputs	Every page tagged with file name, page number and section heading. Slide claims: the pages found by method A, B or C. Red flags: every page of every file
Model output	A list of claims. Each has a value, unit, period, source page and an exact quote.
Code checks	The quote must appear on the cited page (close match). The number must appear in the quote with the same unit and period. Totals and shares are computed in code from cited rows.
Labels	Confirmed, contradicted, not found or unverified. Each fact is marked stated, computed or inferred.
Never	Recommend, rate the deal or guess. A claim that fails a check is shown as unverified, never as fact.
Logged	Every run: input files, model name, raw model output, check results

Behavior examples:

Case	Input	Expected output
Good	Slide 6: "Largest customer is 12% of revenue"	Contradicted. Computed 31% from the customer revenue list, p.2. Both pages linked.
Good	Slide 5: "Gross margin 68%"	Confirmed. Audited accounts FY2025, p.4, quote shown.
Edge	Slide 4: "Revenue FY2025 $38.0M". Monthly accounts say $38.0M. Audited accounts say $36.2M.	Conflict. Both sources shown with dates. No pick between them.
Edge	The model cites p.7 but the quote is on p.8	Unverified. Shown apart from the facts.
Reject	"Should we invest?"	The relevant findings, then: "The decision is yours." No recommendation words.

---

Open Questions

• [ ] PDF reader library: decide at build step 2, with trade-offs shown. @Alex
• [ ] Embedding model for B. Use the open-source PageIndex library for C, or build a small version. Does the demo show A, B, C or all three? @Alex
• [ ] Public demo: stored results only, or live runs with the visitor's own model key? @Alex
• [ ] Final product name before the case study. @Alex

---

Appendix A: Data room

Fintech A: payments software for mid-size retailers, founded 2014. A buyout fund leads the deal. The family office puts in money alongside it.

Files (13): pitch deck (10 slides) · audited accounts FY2024 · audited accounts FY2025 · monthly accounts FY2025 · customer revenue list FY2024 to FY2025 · top 3 customer contracts · loan agreement · ownership table · board minutes (last 4 meetings) · law firm letter · office lease.

Answer key: known issues (9):

#	Type	Slide says	Data room shows	On the list?
V1	Wrong number	Largest customer 12% of revenue	31%, customer list	n/a
V2	Wrong number	Revenue kept from last year's customers 115%	104%, computed from customer list	n/a
V3	Wrong number	410 customers	372, customer list	n/a
V4	Conflict	Revenue FY2025 $38.0M	Audited $36.2M; monthly accounts $38.0M	n/a
D1	Risk		Loan must be repaid if owners change	Yes
D2	Risk		$2.1M claim from a former partner	Yes
D3	Risk		Largest customer's contract ends in 5 months	Yes
H1	Risk		CFO resigned, in board minutes	No
H2	Risk		Office rented from the CEO's company above market rent	No

Correct slide numbers (6): gross margin, staff count, founding year, offices, loan size, products. They test that the tool doesn't flag correct numbers.

Warning-sign list (7 checks): change of owner clauses, legal claims, contracts ending within 12 months, unpaid taxes, loan covenant breaches (loan conditions broken), customer concentration over 20%, plus "other unusual facts."

---

Appendix B: Alternatives and risks

Alternatives considered: - Open question box with search (RAG). Not chosen: a search can't prove a fact is absent, so it can't support "nothing was missed." - Typed questions turned into database queries (text-to-SQL). Not chosen: it adds a second way to be wrong. - Keep Fintech A as an early-stage startup. Not chosen: older companies with loans are closer to typical family office deals.

Risks and recovery:

Risk	Detection	Fallback
The AI misses the 2 issues the red flag list doesn't name	Neither found after 2 attempts to improve its instructions	Remove that check; state the tool only finds what the list names
B or C misses evidence pages, so a claim reads "not found" when the data room has it	Fewer evidence pages found than A on the answer key	Show A's result; report the gap in the case study
The known issues are too easy, so the result proves little	All 9 found on the first run with no prompt work	Add 2 harder ones and report both runs
The tool marks itself correct	Scoring reads the answer key only	Answer key is written first, stored apart, and never passed to the model
The model invents a quote	Quote not found on the cited page	Unverified label; counted in the guardrail