👋 Hi, I’m Andre and welcome to my newsletter Data Driven VC which is all about becoming a better investor with data and AI.
👉 Join the DDVC x Slush Breakfast in Helsinki Nov 18 with Goodwin & Harmonic
Brought to you by Affinity – The AI-first private capital CRM
What the most data-driven funds got right (and wrong) about AI
AI has made investment workflows cheap to automate. The hard part is organizational: who owns scoring, who sees which deals, and which IC memo sections need a human.
Join Andre Retterath of Data Driven VC and Lindsay Bayuk, Affinity’s CMO, on Nov 17 at 1 PM ET to learn why messy-data scoring models lose trust, why 80%+ of IC papers are recycled, and how permission-aware MCPs protect sensitive deals.
Around Christmas last year, Claude Code got good enough that anyone could build seemingly reliable software.
By Q1, everyone in VC was suddenly a builder. Investment professionals, finance teams, ops teams.
By Q3, many of them were pulling back. The weekend build turned into a maintenance job nobody had signed up for.
Most funds will never hire a dedicated engineering team. In our 2026 Landscape Report, we call them workflow builders: investment, finance and ops people stringing bought tools together into end-to-end workflows, often without a single engineer.
This episode is for them. Five steps, in the order I'd tackle them.
Step 1: Get the Data Foundation Right
Every workflow you build later reads from the same data, so its quality sets the ceiling for everything else.
Without engineers, the risk is quieter. Nobody is assigned to notice when the CRM, the data provider and the shared drive start disagreeing.
For a workflow builder, the data foundation comes down to three decisions:
Pick one system of record. For us that's Affinity. Everything else connects to it.
Bring in your first-party data. Pitch decks, IC memos, board decks. At Earlybird that's three decades of material, and most of it lived in Google Drive.
Add third-party data on top. Pick the provider that covers your stage and region best.
Then test it against ground truth. Take a random sample from your portfolio and recent deal flow, and check whether the system has them, and whether the data is right.
The system of record is the baseline for what becomes your system of action. Skip this step and everything you build later inherits the gaps.
Step 2: Map the Work Before You Automate It
The biggest mistake I see: firms start doing something before they've structured the problem.
Think about your workflows in three layers:
Data. Identifying, enriching and deduplicating companies and people.
Workflows and scoring. Hard criteria like region and stage first, then the softer judgment calls.
Productivity. Drafting memos, transcribing calls, pushing notes into the CRM.
The productivity layer is where most workflow builders should start. It's the easiest to verify and the cheapest to get wrong.
To find the right tasks, one of the smartest hacks I've heard came from one of our community breakfasts. Get a young AI-native student. Have them shadow a partner for 2-3 weeks, every single hour of the day. Write down every task, how often it comes up, and how much time goes into it.
Then automate 80/20. Two months in, you'll be surprised how much is automated, and you'll have the ROI to justify a real budget.
Step 3: Buy the Parts, Own the Glue
Almost every part of the stack can be bought today, and a workflow builder should buy almost all of it.
What you own is the glue: how the tools connect, what triggers each workflow, and what the output looks like.
In practice, that glue has two parts:
An automation layer that moves data between tools when something happens. n8n and Zapier are a good place to start.
An agent layer that reasons over the data. For us that's Claude, connected to our system of record through an MCP.
Your Claude or ChatGPT is only as smart as the connectors you give it.
The part nobody plans for is maintenance. Workflows break when a tool changes its API or someone renames a field. So for every workflow you wire up, write down which tools it touches, what triggers it, and who gets the alert when it fails.
For a fund without engineers, that one-page register is the closest thing you'll have to a codebase.
Step 4: Standardize the Output
Give the same input and the same desired output to five investment professionals, and you'll get five different outcomes.
Give it to an agent with a standardized skill, and for the first time ever, you get consistency.
The most practical way to get there: collect 10 of your own past memos and give them to the model as the target. Every fund's memo looks different, so show it yours.
I went through how to build skills task by task in "AI Harness for VC Firms." Before writing one from scratch, check vcskills.com for a version someone in the community has already shared.
Step 5: Give It a Senior Owner
In our Landscape Report, the more technical the owner of the stack, the higher the AI adoption across the firm. Most workflow builders don't have that person yet. In 80% of them, investors and partners own the stack.
Without engineers, nobody owns the glue by default. Someone has to own it on purpose.
My recommendation: give ownership of the strategy and the budget to someone at least at partner level. Ideally a GP, someone who knows the P&L of the management company.
Once the workflows prove their ROI and you have the conviction, find someone technical to own it full time. Because we all know what happens to investors when a portfolio company calls. That's where our focus goes, and the next automation waits.
Find new LPs in our database of 30k+ active family offices, fund of funds, pensions & more via Final Close
Where It Breaks
The failure mode I worry about most is AI slop.
Once the workflows run, everyone can produce an 80-90% good output. Even interns. People feel incredibly powerful, and they quickly stop thinking.
Our rule at Earlybird: before anyone throws a case into Claude, they answer one question. What are the 2-3 key things that need to go right to make this a huge outlier?
I think this will become way more of a problem over the next 12 months.
Bottom Line
You don't need an engineering team to automate a fund.
The lesson from the Q3 pullback is to build right: start from clean data, automate the work you've mapped, buy what someone else maintains, and put a senior name next to the budget.
Running a fund without engineers? Reply and tell me which workflow you automated first, and what broke along the way.
Stay driven,
Andre
