∞Direct answer
Direct answer
Write a brief for AI research. Name a person to check each result. Set the goal, data limits, proof, stop rules and a review note. NIST has four linked parts: govern, map, measure and manage. [1]
Start with a narrow job
State one job. For example, find seed firms in climate tech. They must fit your plan. Do not ask AI to pick who gets a meeting. Do not ask it who gets an investment.
Name the market, stage, place, cheque range and thesis facts. State what sits outside the brief. A clear boundary cuts weak results and makes review faster.
List the data AI may use. Use sources you can check. Do not take a profile, post or score as proof. Keep out deal notes and sensitive data. Only use them if your team has a lawful need.
- Job: find research leads, not final decisions.
- Scope: stage, market, place and cheque range.
- Inputs: named public sources and approved internal records.
- Out of scope: sensitive facts and unverified claims.
Give one person real authority
Name the reviewer before the search starts. State what that person can change, reject or pause. A reviewer who can only click approve is not in control.
For some high-risk AI systems, the EU AI Act needs human oversight. It must help people read outputs, act, or stop the system. It also calls for skill, training and power. These help test any sourcing flow. [2]
Name an escalation owner too. Review results when sources conflict. Review them when a portfolio conflict shows up. Also review claims with no source. Do not let a rank hide these things.
- Decision owner: named investor or team member.
- Power: edit, reject, pause or stop the workflow.
- Escalation: conflict, weak source, sensitive data or unclear use.
- Record: reviewer, date and reason.
Require evidence, not a verdict
Make a short research card for each lead. Show matched facts. Give source links and the check date. List facts not found. Tell the tool to say "not known" when proof is absent.
Do not use one total score as the decision. Keep each factor separate. A firm may fit the market but miss the stage. A strong claim may still have a weak source.
NIST calls for clear jobs, written rules, checks and review. It says notes can help people check work and take charge. [1]
- Match factor and plain reason.
- Source link and date checked.
- Confidence or missing fact.
- Known conflict or reason to pause.
Test the brief before wide use
Run the brief on a small set first. Include known good leads, poor fits and borderline cases. Compare the output with a human review. Save what it missed and what it overstated.
Check a sample of leads before outreach. Look for stale facts, false rivals, duplicate firms and weak links. Change the brief when a test shows a repeat problem.
NIST says risk work must keep going. It calls for a plan to watch and review work. Roles need to be clear. [1]
- Test set: good, poor and borderline cases.
- Check: facts, links, fit and conflicts.
- Measure: false leads, missed leads and reviewer changes.
- Update: version the brief and note why it changed.
Protect people and keep the final call human
A sourcing result can affect who gets attention. Do not use it as an auto pass list. A person must read the proof before outreach. Do this before a meeting, diligence or a decision.
If you use profiles of people, data rules may apply. The ICO says UK GDPR limits some fully automatic decisions. This is so when they have a legal or like major effect. Where rules apply, fix errors, cut bias and let people get human help. [3]
Get good privacy and legal advice for your case and place. An AI brief is a work note. It is not legal advice. It is not due diligence. It is not an investment decision.
- No automatic exclusion from a human review.
- No outreach based on a score alone.
- Keep a route to question or correct a material result.
- Review local privacy and data duties.
FAQ
Frequently asked questions
What is an AI sourcing brief?
It is a short rule set for AI research. It says what to do, what data to use, and who checks it. It must not give the investment call to AI.
Who should be the human reviewer?
Pick someone who knows your thesis. That person must change results. Give them time and power. A name with no power adds no real control.
Should the brief use a score?
A score can sort research leads. Keep the factors and proof clear. Do not let one score replace a check of stage, source, conflicts and missing facts.
When should the workflow stop?
Stop or escalate if sources conflict. Do so if there is no proof. Do so if a close portfolio conflict shows up. Stop if sensitive data enters the task. Note why. The owner then picks the next step.
REF
Sources
- 1
National Institute of Standards and Technology
5 AI RMF Core ↗ - 2
- 3
Information Commissioner's Office
Rights related to automated decision-making including profiling ↗