Use this before a score, built or bought, is allowed to rank companies in the firm's pipeline, and again at each annual review of the score. Fill it from the firm's own records and the vendor's written answers, not from a demonstration. The questions follow Bonelli, "Data-Driven Investors" (The Review of Financial Studies 39, no. 7, 2026, pages 1909 to 1969; figures from the author's September 2024 manuscript); the lookback is the firm's own evidence.
Firm / scope (funds, stages, sectors the score will cover):
Prepared by / date:
Score owner (who can explain any value it shows):
Source: built in-house (by whom) / bought from a vendor (name) / AI tool reading firm records (name):
Definition
- [ ] The outcome the score predicts is written down in one sentence (for example, "raised a next round within 24 months," or "exited for more than it raised").
- [ ] The score is described to the team as a triage tool if it predicts a common outcome such as a next round.
- [ ] The history it learns from is named: the firm's own records, a vendor's database, or both, with the years covered.
- [ ] Each record shows a similarity measure next to the score: how closely the company resembles companies the firm or the market already funded.
- [ ] A company with no close comparables reads "cannot place," never a low score.
- [ ] The score orders the queue. It does not remove companies from it.
Outcome predicted:
Training history and years:
How similarity is measured:
Vendor or builder questions
| Question | Answer | Evidence seen | Acceptable? |
|---|---|---|---|
| What outcome is the score trained to predict? | |||
| What history is it trained on, and does it include companies that never raised? | |||
| Can the firm see why a company scored as it did? | |||
| How does it treat a company with no close comparables? | |||
| What firm data does it read, and where does that data go? |
Lookback
Run on companies the firm met in the past three years, scored as the model would have scored them at the time of the first meeting.
Companies in the lookback / period:
Proposed cutoff (companies below it get no first call):
| Company | Score at first meeting | Similarity (high / middle / cannot place) | Above or below cutoff | What the firm now thinks of it |
|---|---|---|---|---|
- [ ] The firm's most valued companies from the period are listed, and their position against the cutoff is recorded.
- [ ] If more than one of them falls below the cutoff, the cutoff is lowered or the score is not adopted.
Most valued companies that would have fallen below the cutoff:
Placement and monitoring
- [ ] A share of first meetings is reserved for companies the score cannot place or ranks low: percent, with an owner:
- [ ] The pipeline report counts first meetings by similarity band each quarter.
- [ ] A falling count of low-similarity first meetings is reviewed at the next pipeline meeting, not left for the annual review.
- [ ] The score, similarity, date, and data used appear on each record.
| Quarter | First meetings | From low-similarity companies | From the reserved lane | Notes |
|---|---|---|---|---|
Decision: adopt as triage / adopt with a wider reserved lane / run the lookback again later / do not adopt
Approved by / date:
Next review (at least annually, and after any change in model, vendor, or training data):
Common failure modes
- The score predicts a next round and is read as predicting a great company.
- A company with nothing like it in the history scores low and never gets a call.
- Nobody tracks what share of first meetings come from unfamiliar companies, so the drift is invisible until the portfolio shows it.
- The lookback is skipped because the demonstration looked convincing.
- The reserved lane has no owner and quietly closes.
This review decides whether a score belongs in the firm's pipeline and how it is used. It does not evaluate the vendor's security or data handling; use the vendor review and acceptance brief for that.