🔥How to Build an AI Harness for… | Data Driven VC Summary

1. Key Themes

The AI Advantage Gap Is About Infrastructure, Not Model Access

The central thesis of the article is that AI differentiation among investment firms is not determined by which model they use, but by what institutional context and structure has been built around it.

"Most funds have access to the same AI as their competitors and a very different amount to show for it. The gap isn't the model. It's everything the firm has built around it."

The problem is that firm knowledge remains siloed: "It sits in people's heads, scattered documents, and individual chat histories. So every new conversation starts largely from zero."

The "AI Harness" as a Scalable Knowledge System for Investment Firms

The article introduces a concrete framework — an "AI harness" — that systematizes firm-level knowledge, rules, and judgment so AI can execute tasks consistently across team members, not just for the individual who prompted it.

"An AI harness fixes that. It turns what the firm knows and how it works into something AI can consistently use across tasks and team members."

The method is deliberately incremental: "You don't build that harness all at once. You build it workflow by workflow: codifying the knowledge, rules, and judgment behind each task until AI can execute it consistently across the firm."

Task-Specific Knowledge Inputs Matter More Than General Prompting

The article stresses that effective AI use in VC requires mapping each workflow to the specific knowledge it requires — not applying generic prompts. Different tasks require different context.

"Deck screening lives and dies on the firm's current investment thesis... A portfolio update runs on KPI history and board notes. LP commentary runs on real fund performance and the tone of past letters."

It also flags the gap between marketing language and operational reality in firm theses: "The LP-facing thesis is written to persuade... The version the model needs is shorter and current: what's actually getting a deal past the first call this quarter."

Leadership Ownership Is the Bottleneck to AI Adoption

The article identifies governance and accountability — not technology — as the primary failure point in AI adoption at VC firms.

"Signing off can't be a partner glancing at a demo and nodding along. It has to mean someone is now accountable for this workflow the same way they'd be accountable for anything else that touches the firm's process."

"If nobody at leadership level is willing to own it, the workflow stays a good pilot, however well it did in testing."

2. Contrarian Perspectives

The bottleneck to AI adoption at investment firms isn't access to better models — it's organizational discipline.

Most AI discourse in VC centers on which model to use or which tools to buy. The article pushes back hard: firms have identical model access and wildly different outcomes. The differentiator is whether a firm has done the structural work to encode its judgment and make it machine-readable.

"The gap isn't the model. It's everything the firm has built around it."

This reframes AI investment from a "pick the right tool" problem to an organizational design problem.

The public-facing investment thesis is operationally useless for AI.

Firms typically assume their thesis documentation is sufficient context for AI tasks. The article argues the opposite — the LP-facing thesis is a marketing document that fails as an operational input.

"The LP-facing thesis is written to persuade, updated once a year if that, full of language that reads well in a deck without doing much operational work... which sectors quietly stopped being a yes six months ago even though nobody updated the site."

This means firms need a separate, living internal thesis document — one that reflects actual, current decision criteria — to make AI-assisted screening useful.

The most technically sophisticated functional owner of the AI stack drives firm-wide adoption.

Conventional wisdom might assume that AI tools are easy enough for anyone to deploy. The article's data from the DDVC Landscape Report 2026 contradicts this.

"One pattern stood out: the more technical the functional owner of the stack, the higher the AI adoption across the firm."

This implies firms should prioritize technical competence in whoever manages their AI infrastructure, not just enthusiasm or seniority.

3. Companies Identified

Exa

Description: AI-native web search and data infrastructure company
Why mentioned: Newsletter sponsor; positioned as the enabling data layer for AI-driven investment workflows
Quote: "Exa turns the web into live, structured data that your agents and pipelines can act on. AI-forward firms use Exa to source companies by thesis, enrich target businesses programmatically, and monitor signals across the market in real time."

vcskills.com (in partnership with OverDrive)

Description: A curated library of VC-specific AI "skills" — pre-built workflow rules for investment tasks
Why mentioned: Referenced as a resource where firms can find pre-built versions of task-specific AI rules rather than building from scratch
Quote: "We've been curating VC-specific skills with our partner OverDrive at vcskills.com, and a decent version of this task's rules may already be sitting there."

4. People Identified

Andre Retterath

Description: Author of Data Driven VC newsletter; investor and practitioner focused on AI and data applications in venture capital
Why mentioned: Author and practitioner behind the five-step AI harness framework; references internal research from the DDVC Landscape Report 2026
Quote: "Hi, I'm Andre and welcome to my newsletter Data Driven VC which is all about becoming a better investor with data and AI."

5. Operating Insights

Decompose each workflow before writing a single prompt. Before building AI tooling for any task, map its anatomy: what triggers it, what inputs a human currently assembles, what a good output looks like, and what the model specifically needs to see. This diagnostic step prevents misaligned outputs caused by incomplete context.

"Before touching a prompt, map out what that task consists of... This step is mostly diagnostic. You're mapping the task's real anatomy before deciding what the AI gets fed."

Always test workflows with someone who didn't build them — this is the step most firms skip.

"Hand the workflow to a different associate. A different team. Someone who wasn't in the room when the rules got written... This is where you find out whether the harness generalizes past the one person who built it. That's the entire point of building it in the first place."

The test should catch not just bad AI outputs but also rules that made sense to the author and confused the next person — a sign the rule is not yet finished.

Embed hard guardrails for sensitive task types. The article distinguishes between soft rules the model is expected to follow (e.g., flagging unverifiable claims) and hard stops that don't depend on model performance at all.

"Nothing going to a founder or an LP without a person actually reading it first... If the task touches anything MNPI-adjacent or LP-facing, get compliance looking at this list before it goes any further."

6. Overlooked Insights

Knowledge documents have a decay problem that most firms aren't managing. The article briefly but sharply notes that an outdated context document is actively harmful — the model will make calls based on stale information every time it is invoked.

"Whatever gets written down, review it on an actual cadence. A page that's accurate this quarter beats a document from eighteen months ago every time the model uses it to make a call."

This implies firms need a maintenance calendar for their AI knowledge base, treating it as a living operational asset rather than a one-time documentation project.

The "pilot trap" is a governance failure, not a technology failure. The article notes that many promising AI workflows never graduate from pilot status — and the reason is almost never technical. It is the absence of someone at the leadership level willing to formally own the workflow's outcomes.

"If nobody at leadership level is willing to own it, the workflow stays a good pilot, however well it did in testing."

This is a subtle but important organizational design signal: AI adoption in professional services firms may stall not from tool limitations but from accountability gaps at the top.