AI automation for venture capital: sourcing is not judgement

A fund sees far more opportunities than it can examine, which makes filtering the obvious automation target and the one with a specific hazard attached. A model built from a fund's past investments will reproduce that fund's past pattern — including whatever was accidental about it — and present the result as a signal.

The work that automates cleanly is everything else: keeping track of what you have seen, running the logistics of diligence, chasing portfolio reporting, and knowing what you already know about a company or a person before a meeting.

What we'll cover

• Sourcing volume and the filtering hazard
• What a fund actually loses track of
• The relationship graph
• Building it with CodeWords
• Diligence logistics
• Portfolio reporting, which is chasing
• Fund operations and LP reporting
• What must stay judgement
• Helping the portfolio rather than reporting on it
• Making it survive
• Limits worth knowing about
• What to build first
• Frequently asked questions

Sourcing volume and the filtering hazard

Volume exceeds capacity by a wide margin at any fund with visible deal flow.

A model trained on past investments learns the fund's historic pattern, which includes its biases as faithfully as its insights.

Base rates are brutal. The outcomes being predicted are rare, which makes any model's precision poor and its confident-looking output misleading.

Signals are gameable once known, and the people most able to optimise for them are the best-networked founders rather than the best ones.

Screening people has legal dimensions in some contexts and reputational ones in all of them.

Use it to order the queue, not to close it. A ranking that changes what gets read first is defensible; an automated rejection is not.

What a fund actually loses track of

Companies seen too early, which is the largest and most recoverable miss — a good company met eighteen months before it was ready.

Why something was passed on, which is essential context when it returns and is almost never recorded usefully.

Who introduced whom, which is the fund's actual network and lives in individual inboxes.

Follow-ups promised in meetings, which are dropped at a rate nobody measures.

Competitive intelligence encountered incidentally and never shared internally.

Record the pass reason in structured form, since a company met early and passed for timing reasons is a different thing from one passed on the team.

The relationship graph

The fund's network is its asset and it is distributed across individual mailboxes and memories.

Know who at the fund knows this person before a meeting, which is both useful and frequently unknown.

Track warm paths to companies you want to meet, which is the sourcing work that actually converts.

Record who referred what and what happened, since referrer quality varies enormously and nobody measures it.

Respect the privacy boundary. Mining personal inboxes to build a graph is a decision that needs to be made explicitly and with the people whose inboxes they are.

Keep it current, because people change roles constantly and a stale graph produces embarrassing introductions.

Building it with CodeWords

CodeWords connects to more than 3,000 integrations, and the connection is made once and reused.

1. Open CodeWords and start a new automation.
2. Describe what should happen in plain language to Cody, the automation builder: what is tracked, what is assembled before meetings, and what is chased.
3. Authorize the connections — the CRM or deal system, the mail and calendar, the data room and your portfolio reporting.
4. Describe the exceptions: a company seen before, a conflict with an existing investment, a founder already known to the fund.
5. Run it over last quarter's pipeline and check what it would have surfaced.

You describe the outcome; Cody builds it, connects it, and deploys it. The free plan covers light use, with Pro at $39 per month and Business at $100 per month as usage grows; details are on the pricing page.

Diligence logistics

Most of diligence is coordination — requests, documents, references, calls — rather than analysis.

Track the request list against what has arrived, which is otherwise a spreadsheet somebody maintains badly under time pressure.

Chase automatically, since the deal timeline is short and the chasing is what delays it.

Assemble the reference calls: scheduling, questions, notes in a consistent shape so they can be compared.

Check the obvious things automatically — corporate records, litigation, sanctions, domain and product checks — which is a repeatable list.

Not the assessment. Whether the numbers make sense and whether the team can execute are the two questions diligence exists to answer, and neither is mechanical.

Portfolio reporting, which is chasing

Late reporting is the norm, and it is a coordination problem rather than a reluctance one.

Send the request with the template, on a schedule, so the founder is not reconstructing what you asked for.

Chase escalating by age, which works and is unpleasant to do by hand.

Normalise what arrives, since every company reports differently and comparison requires a consistent shape.

Flag what changed materially, so a partner reads the exceptions rather than forty updates.

Make it worth the founder's time. A portfolio report that produces nothing back is the reason it arrives late, and returning a benchmark or an introduction changes that.

Fund operations and LP reporting

Capital calls and distributions have deadlines and defined notice requirements.

Valuation updates on a cycle, assembled from portfolio data rather than reconstructed.

LP reporting is repetitive, deadline-bound and scrutinised, which is the profile of good automation.

Compliance obligations vary by jurisdiction and fund structure and have their own calendars.

Keep the audit trail, since fund operations are examined and the evidence needs to exist.

Do not automate the narrative. The commentary in an LP letter is a judgement about the portfolio and is read as one.

What must stay judgement

The investment decision, entirely.

The assessment of a team, which is the primary thing most funds claim to evaluate and is not observable in data.

Whether a market is real, which requires knowing things that are not written down.

The relationship with a founder, before and after investing.

Any decision about a person made from an automated score, which carries obligations in several jurisdictions and reputational risk everywhere.

Saying no. A pass delivered well preserves a relationship that frequently matters later, and an automated rejection is remembered.

Helping the portfolio rather than reporting on it

Funds automate the reporting they receive and rarely the support they provide, which is the half founders actually value and the half that differentiates one investor from another.

Answer the recurring question once. Every portfolio has the same questions about hiring, pricing, fundraising and process, and the answers are written repeatedly in individual emails.

Make introductions findable, since the fund's network is the most requested and least systematically delivered form of help.

Share benchmarks back, anonymised, which is the thing founders most want from reporting and almost never receive.

Surface the company that has gone quiet, since a founder who stops communicating is the signal that matters most and produces no alert.

Connect founders to each other on specific problems, which is frequently more useful than connecting them to the fund.

Respect confidentiality throughout, because the fastest way to lose a portfolio's trust is for something shared with an investor to appear somewhere else.

Making it survive

Handle the company that appears twice under different names, which is common and which fragments history.

Keep the pass reasons permanently, since their value increases with time.

Respect confidentiality. Material shared in diligence is confidential and should not leak across a portfolio.

Handle conflicts explicitly, since an automation surfacing one portfolio company's data to a competitor is a serious failure.

Report the outcome: companies seen, passed with reasons, in diligence, reports received and outstanding.

Limits worth knowing about

Outcome data is sparse and slow. A fund learns whether it was right over a decade, which is not a feedback loop any model can use.

Survivorship bias distorts any analysis of what worked.

Private company data is poor, inconsistent and frequently self-reported.

Confidentiality constrains what can be aggregated or shared, even internally in some structures.

Automation cannot source proprietary deals, which come from relationships and reputation rather than from coverage.

What to build first

The pass-reason record: every company seen, with a structured reason for the decision and a date to revisit where the reason was timing. Companies met too early are the largest recoverable miss a fund has, and nothing else in a fund's systems brings them back at the right moment.

Two habits make the difference. Use any scoring to order the queue rather than to close it, since a model trained on past investments reproduces the fund's past pattern including whatever was accidental about it. And make portfolio reporting worth the founder's time by returning something — a benchmark, an introduction, an answer — because a report that produces nothing back is precisely why it arrives late.

Frequently asked questions

Can deal screening be automated?

To order the queue, yes. To close it, no. A model trained on a fund's past investments learns its historic pattern including its biases, the outcomes being predicted are rare enough to make precision poor, and the signals become gameable once known.

What does a fund lose track of most?

Companies seen too early. It is the largest recoverable miss, and bringing them back at the right moment requires a structured pass reason and a revisit date, which almost no fund records usefully.

Why record structured pass reasons?

Because a company passed for timing is a completely different thing from one passed on the team, and the distinction is what makes the record actionable when the company returns eighteen months later.

What part of diligence automates?

The coordination — request tracking, chasing, reference call scheduling and note structure, and the repeatable checks on corporate records and sanctions. Not the assessment of whether the numbers make sense or the team can execute.

Why does portfolio reporting arrive late?

Because it produces nothing back for the founder. Sending the template with the request, chasing by age, and returning a benchmark or an introduction changes the incentive rather than the process.

Is mining personal inboxes to build a relationship graph acceptable?

It is a decision that needs to be made explicitly and with the people whose inboxes they are, rather than as a technical default. The graph is valuable and the privacy boundary is real.

What must stay entirely human?

The investment decision, the assessment of a team, whether a market is real, the founder relationship, and saying no — which delivered well preserves a relationship that frequently matters later.

Why can models not learn from a fund's outcomes?

Because the feedback arrives over a decade, the successes are rare, and survivorship bias distorts any retrospective analysis. It is not a loop a model can usefully learn from.

What portfolio support automates well?

Answering recurring questions once rather than repeatedly by email, making the fund's network searchable, returning anonymised benchmarks from reporting, and surfacing companies that have gone quiet — which is the signal that matters most and produces no alert anywhere.

What is the fastest way to lose a portfolio's trust?

Something shared in confidence with an investor appearing somewhere else. Any automation spanning a portfolio needs explicit conflict and confidentiality handling rather than assuming good intent.

Related reading

• Automating proposal generation
• Automating meeting scheduling
• Workflow automation for finance teams