Where AI Helps and Where a Person Signs: A Human-in-the-Loop Workflow for Sourcing, Meeting Notes and Venture Screening
Updated: Oct 21

SMART IMPACT CONSULT | INSIGHTS | INVESTMENT INTELLIGENCE
AI can draft the first pass of your sourcing list, your meeting notes and your screening scorecard. It cannot be accountable for any of them. This is a three-stage workflow that keeps a named person in charge of every decision, and the record that proves it.
By Smart Impact Consult
The Speed Is Real. So Is the Gap.
Most investment teams now use AI somewhere: to widen a list of companies, to transcribe a founder call, to summarise a pitch deck. It saves hours. The risk is not that the tools are used. It is that nobody can say, afterwards, which parts a person checked.
Regulators have started to say this plainly. In May 2024, Europe's securities regulator (ESMA) warned firms about over-reliance on AI tools and about "neglecting the importance of human judgment", and said responsibility stays with the firm's management body whether decisions are made by people or by AI tools. ESMA's statement is aimed at investment firms in the EU, but the principle travels: whoever signs the decision owns it.
THE SIC TEST For every AI output that reaches a decision, can you name the person who checked it? |
The Workflow at a Glance
The workflow has three stages: sourcing, meeting notes and venture screening. At each one, AI does the volume and a person does the judgement. A short record shows that it happened.
Stage | AI drafts | A person decides | Record kept |
1. Sourcing | Widens the long list, removes duplicates, groups by sector | Confirms each company exists, trades, and who owns it | A source link for every company |
2. Meeting notes | Drafts the transcript, summary and action items | Gets consent first, corrects the notes, approves what is filed | Approved notes and a consent record |
3. Venture screening | Scores against written criteria and flags gaps | Tests claims against evidence, then decides | The decision and a written reason |
Stage 1. Sourcing: Widen the List, Then Confirm It Is Real
AI is good at breadth. Give it a sector, a country and a stage and it will return candidates in minutes. It is also capable of returning companies that have closed, merged, been acquired or never existed in the form described. So the rule is simple: AI finds, a person confirms.
Open every source. A person confirms each company against its registry entry, its own website, or independent news or filings.
Check ownership and trading status. A company that is wholly owned by a larger group may not be a fit for an independent-company fund, whatever the summary says.
Remove what has no source. A company with no link behind it comes off the list.
WHAT FAILS A company is on your list because a model summarised it, and nobody opened the source. |
What to keep: a source link for every company on the list, and the name of the person who opened it.
Stage 2. Meeting Notes: Ask Before You Record, Review Before You File
AI note-takers are now built into video-call tools, which makes it easy to record a founder call without anyone deciding to. Two things need a person: consent before the call, and review before the notes are filed.
In Kenya, the Data Protection Act, 2019 requires personal data to be collected for a specified purpose, kept to what is necessary, and not retained longer than needed (section 25), and recognises consent as one lawful basis for processing (section 30). In practice that means telling everyone how the call is recorded and why, capturing only what the purpose needs, and deleting recordings on a set date. Check the law that applies to you, since founders and investors on one call may sit in different jurisdictions.
Tell people, and get consent, before recording or transcribing.
Correct the notes. A person fixes names, numbers and commitments before anything is filed.
Set a retention date. Delete recordings when the purpose is served.
WHAT FAILS An AI note-taker joins a founder call, and the founder was never told it was recording. |
What to keep: the consent record and the approved version of the notes, not the raw transcript alone.
Stage 3. Venture Screening: Score with AI, Decide with a Person
AI can score a pipeline against your criteria quickly and consistently. What it cannot do is know whether the inputs are true. Founders' figures can be inaccurate or deliberately inflated, and generated text can invent them: TechCabal's report on StartupFuel's DiligenceGPT notes that founders' market-size figures can be inaccurate or manipulated, including LLM-generated ones.
Write the criteria first. Before the AI scores anything.
Trace every claim that moves a score. If a number drives the outcome, a person finds the evidence behind it: data-room documents, bank or customer evidence, independent market data.
Treat the score as a prompt, not a verdict. A named person decides, and writes the reason.
WHAT FAILS A founder's market-size figure was generated by an AI tool, and nobody checked it. |
What to keep: the criteria, the evidence behind each claim that moved the score, and the decision with its reason.
Five Checkpoints That Keep the Loop Closed
Gate | What must be true |
1. Source | Every company traces to a link a person has opened. |
2. Consent | Everyone in the room knows how the call is recorded, and why. |
3. Evidence | Every claim that moves a score has proof behind it. |
4. Decision | A named person decides, and writes the reason. |
5. Record | The AI output, the edits and the decision are kept together. |
Decide Which Tools May Touch Which Data
A workflow is only as safe as the tools underneath it. Law-firm guidance for fund managers, such as Cooley's AI Policy for Fund Managers (30 April 2026), points to a few basics that apply to any small team:
List every tool in use. AI is often built into tools a firm already uses.
Separate enterprise from consumer versions. Consumer tools often lack the contractual protections enterprise versions carry.
Check where data is stored. Kenya's Act restricts moving personal data abroad without safeguards or consent.
Keep a person in the decision. The Act gives people a right not to be subject to a decision based solely on automated processing that significantly affects them (section 35).
For teams that want a wider governance frame, the US NIST AI Risk Management Framework organises the work under four functions: govern, map, measure and manage.
The Decision Record: Six Fields
The record is what turns good practice into something you can defend. One short entry per decision is enough.
Field | What it holds |
When | Date of the decision |
Who | The named decision-maker |
AI used | Tool, and what it produced |
Changed | What the person corrected |
Evidence | What was checked, and where from |
Reason | Why advance, hold or decline |
THE PRACTICAL IMPLICATION Can you produce the record when an investor, auditor or founder asks? If not, the workflow exists only in people's heads, and that is not a control. |
AI does the volume. A person does the judgement. The record proves it.
Map This Workflow to Your Team
Smart Impact Consult runs workflow-mapping sessions for investment teams: we map where AI is already in your sourcing, notes and screening, mark where a person must decide, and draft the checkpoints and decision record to fit how you work.
Sources and further reading
Editorial verification note
The three-stage workflow, five checkpoints and six-field decision record are Smart Impact Consult's own illustrative framework and are not legal advice. Statements about ESMA, Cooley, the Kenya Data Protection Act, NIST and TechCabal are drawn from the linked sources; laws and guidance change, so check current versions and the rules that apply to you before relying on them.




Comments