A quick note for those readers just joining: Carried Away is my Substack on how artificial intelligence is upending venture capital itself. The eight previous posts are here, starting with “The Disruptors’ Blind Spot,” which lays out the premise.
An extensive report on AI-driven venture capital came out last month, and buried on page forty is the most honest sentence anyone in our business has published about this AI shift, followed in the very next sentence by the least honest one.
The survey is Andre Retterath’s Data Driven VC Landscape 2026, the fourth edition of the industry’s best map of its own AI adoption: 345 firms, the most AI-forward cohort venture has, from workflow builders up to firms running in-house engineering teams. Retterath has done the field a real service with it and should be commended.
The honest part first. Asked what is holding VC firms back, nearly half of respondents named time and bandwidth, and 41 percent named data quality. Firms that are falling behind, the report says, already have the right tools. And then the sentence: “They’re failing to find the time to embed them into real workflows on top of clean data.”
The constraint on becoming an AI-driven firm, in other words, is the attention of the partners who would have to remake it.
Then, the very next sentence: “This is exactly the gap that agentic tooling is designed to close.”
He found a time bottleneck and prescribed a tool.
I don’t think Retterath is being dishonest. He believes that second sentence, and so does nearly everyone else in our business, which is what makes it the industry‘s least honest sentence rather than his. We keep writing it ourselves, in budgets if not in prose: for the third straight year, most firms plan to spend more on their tech and AI stack (pg. 21), budgeting as if the whole fix were on a price list because that’s what management fees buy.
I made a version of this argument a few weeks ago, in my post “A New Engine, Not a New Paint Job“:
“You can buy the software. You can license the data. You can hire a few engineers and seat them in the corner. None of it, on its own, produces a single better investment.”
The report’s own data agrees, and then it reaches for the tool anyway.
The trouble is arithmetic. Nothing you buy arrives installed. A tool has to be pointed at clean data and wired into the way your process moves through your week, and yes, some of that work you genuinely can pay for: the data plumbing, the engineer who owns the stack.
Yet the last step is different, and it is not a single task. Teaching the system what your partnership counts as an exceptional company can only be done by the people who know what the firm believes, and those people are the senior partners. Someone has to decide which calls it is allowed to make on its own, and which still come to a partner. Someone has to sit with what it surfaces, deal after deal, and tell it where it got a company read wrong, until it stops getting them wrong. None of this is work you can seat an engineer in the corner to do. Bought off the shelf or built in-house, it is yours either way. So the final stretch of embedding is labor, the labor is priced in partner attention, and partner attention is the thing that was scarce before the tool showed up.
And this bottleneck reaches past the investment call. The report maps AI across the whole firm, and inside each function sits a layer of judgment no vendor can fill: what makes an intro worth acting on, when to step into a struggling portfolio company, what to put in an LP update and what to leave out. You can buy the scaffolding around each layer. The layer itself is the firm’s taste, and only the people who hold it can install it.
Retterath made this case himself, months before the report. In a February essay on the venture firm of 2030, he argued that a firm’s real moat is its taste, its “new core IP,” the one thing a rival cannot buy because it is distilled from that firm’s own accumulated decisions. I think he is exactly right. His essay pictures that taste encoded into an algorithm that keeps learning. It says less about who does the encoding, and what it costs them.
The fair objection is that this is the entire meaning of the word agentic. Yesterday’s software needed wiring; agents wire themselves. Maybe, eventually. But this is the edition in which the report itself declares that agentic VC has “moved from concept to category,” a third of the surveyed firms now describe themselves as AI-led, and time is still the constraint they name first. The gap the agents were designed to close is the gap the survey still shows. Anyone who has delegated real work to one can guess why. An agent front-loads the specification: a tool needs configuring, an agent needs to be told what you believe.
I know what that judgment layer costs, because I recently paid for it, with time, in one aspect of my own personal life. My digital health data lives in at least a dozen places. There is the fitness tracker on my wrist, my Strava app, a vitamin supplement tracker, three different running apps I have rotated between, and a virtual personal trainer workout app. There are blood panels from an online lab company and another set from my own doctor, the medical history I have re-typed into every one of them and into ChatGPT and Claude over the past few years, and finally the rehab app my physical therapist put me on after a minor injury sidelined me this winter. Looking for a way out of the pain, I finally sat down with the most capable AI model I could get my hands on and pulled all of it into one picture. The machine did the part I had been avoiding for years: aggregating and reconciling it all into one central repository.
The picture still cost me half a Sunday afternoon I had to guard against my calendar, because the part that mattered was work only I myself could orchestrate to set up future progress. Which recovery score to trust on a morning my body said otherwise. What my injury history means for what is safe to push this week. Which of the training plans I believe and will follow. The machine gave me the numbers but not what they were for. It had all the intelligence and none of the answers.
Now scale that up to a partnership. The history a firm would hand its machine is deal notes, IC debates, the memory of every pass on a company that went on to become a unicorn. The report calls that kind of proprietary data “the only non-replicable input” and says outright that “no vendor can sell what your own firm has accumulated” (pg. 42). That is true, and it is the easy part. Yet a pile of deal data records what the firm decided, not the reasoning it would carry into a company it has never seen. And the loop that is supposed to learn the rest, deal after deal, still runs on a person: it improves when someone tells it a read was wrong, or when an outcome finally lands, and in venture that outcome lands a decade later, if it lands at all. Encoding what the partnership actually believes is that same weekend afternoon of mine at firm scale, and every hour of it competes with a live term sheet or a portfolio founder who needs calling back today.
Last week I argued that the winning firms will be the ones that codify their judgment fastest. This is what that costs.
AI does not remove venture’s bottleneck; it relocates it.
The scarce partner-hours used to go into making the calls, one at a time. In a firm that genuinely rebuilds, those same hours go into encoding the calls once, so the machine can make the routine ones at scale.
Some firms will spend the attention now, deliberately, at the cost of deals they could have chased, and stop requiring it later. Everyone else will keep doing what the survey shows us all doing: adding line items, and waiting for the AI tools to install themselves.
This is a really interesting piece. Appreciate you writing it. In the context problem you’re describing in the investment firm context, if every meeting and conversation were recorded (why we’re investing, why we’re passing, etc.), would that solve the problem? Recordings would be more passive than the active time you describe needing to spend on that “Sunday afternoon.” Because of course you have the data on how the investment performs from a quantitative private perspective... so long as you have the inputs from the initial decision making process, that should be enough to draw an analysis...unless I’m missing something. I get that this all will take years since we can’t go back and create recordings from year’s prior :)