A score is useful only when it helps an investment team make a specific decision. In twenty minutes, you can make a first deal-scoring agent reviewable. You cannot make it production-ready, prove that its weights are correct, or delegate investment judgment to it.

The goal of this walkthrough is smaller and more useful: create an evidence-backed first version that a partner can challenge, correct, and improve.

Minutes 0-3: define the decision before the score

Choose one decision with a clear review horizon. For example: which new inbound opportunities deserve partner review this week?

Do not ask the agent to identify the best companies. That hides the fund's constraints and invites a generic ranking. State the stage, geography, sector, ownership, portfolio, and timing boundaries that matter to this decision.

Write down who reviews the result and what the score is allowed to change. In this version, it may propose an internal queue. It may not reject a company, contact a founder, move a CRM stage, or make an investment recommendation.

Minutes 3-7: create an illustrative 100-point rubric

Use the fund's thesis, not a generic startup template. The following weights are only an example to make the first calibration visible:

Thesis fit: 25 points. Stage, geography, sector, business model, and ownership match.
Team-market evidence: 20 points. Relevant founder experience and credible evidence connecting the team to the problem.
Customer or traction evidence: 20 points. Quality and freshness of customer, usage, revenue, or other fund-approved evidence.
Market structure and timing: 15 points. Why the opportunity can matter now, including credible counterevidence.
Differentiation and defensibility: 10 points. Evidence that the company is meaningfully different and that the advantage may persist.
Portfolio and strategic constraints: 10 points. Fit with portfolio exposure, reserves, ownership goals, and the fund's current priorities.

A missing fact is not zero. Mark it unknown, reduce confidence, and show which source or founder question could resolve it. Otherwise the score rewards companies with more public information rather than companies with stronger investment evidence.

Keep hard exclusions outside the score

Some conditions should stop or redirect the review instead of quietly subtracting points. Examples include uncertain company identity, a plausible portfolio conflict, an excluded geography, a prohibited business model, or a stage outside the fund's mandate.

Encode each condition with its evidence requirement and route. An identity ambiguity may stop research for clarification. A possible portfolio conflict may go directly to a partner. An out-of-scope stage may be held rather than scored.

This prevents a company from receiving an attractive total while violating a condition the fund considers decisive. It also makes exceptions visible: a partner can override a rule, but the system records why.

Minutes 7-12: calibrate five fictional cases

Do not begin with live dealflow. Use five fictional cases that expose different failure modes.

Case A: clear fit. Strong thesis match, current evidence, no obvious conflict, and few unknowns.
Case B: clear non-fit. Interesting company, but outside a hard stage or geography constraint.
Case C: high potential, weak evidence. Strong narrative and team signal, but customer claims are old or unverified.
Case D: thesis exception. The company misses one written preference, yet demonstrates a pattern partners have advanced before.
Case E: possible portfolio conflict. Strong score on most dimensions, but a strategic overlap requires human judgment.

Ask the agent to show the evidence behind every criterion, what remains unknown, the strongest countercase, and which rule determined the proposed route. If a reviewer cannot reconstruct the score, it is not ready for real opportunities.

Minutes 12-16: define three internal routes

Use three queues rather than a false yes-or-no verdict.

Prioritize: evidence supports timely partner review. This is not an advance decision.
Review: uncertainty, an exception, or a conflict requires explicit judgment.
Hold: current evidence does not justify attention now, but the case and revisit condition remain recorded.

Thresholds are calibration aids, not investment policy. A fund might initially test 75 and above for Prioritize, 55-74 for Review, and below 55 for Hold, then change those bands after examining false positives and missed opportunities. Hard exclusions still route separately.

Every routed case should include its score, confidence, source dates, missing evidence, countercase, and proposed next question. The partner sees the reasoning before the number.

Minutes 16-20: add correction and authority

A scoring agent becomes specific to the fund only when corrections survive the meeting. Record which criterion changed, the old and new assessment, who corrected it, why, and whether the rubric itself needs a new version.

Do not silently rewrite history. A company assessed under version 1 should remain reproducible after version 2 exists. If partners repeatedly advance thesis exceptions, that is evidence to revisit the operating rubric rather than train the agent to hide the exception.

Keep the next action separate from the score. The agent may draft founder questions or a CRM update, but outreach, stage changes, pass decisions, and other consequential actions require their own authority. Approval for a score is not approval to send.

The review packet

The first usable output is compact:

decision and review horizon;
rubric version and illustrative weights;
hard exclusions and ambiguity rules;
evidence, source links, and freshness dates;
score, confidence, and proposed internal route;
missing evidence and strongest countercase;
reviewer correction and reason;
separately controlled next action.

After twenty minutes, you have not automated investment judgment. You have made part of the screening process explicit enough to test. That is the right starting point: the agent makes the evidence and rules consistent; the investment team decides when the model is wrong and what happens next.

Building a Thesis-Aware Deal-Scoring Agent: A 20-Minute Walkthrough