Nothing that you want to build is impossible, the only real limit holding you back (assuming you are a Claude user) is your token limit. A second lesson is that it gives you a closer look into the everyday functions of your job and tests how well you really understand what it is you do or are wanting to do with AI.

When you try to systematise something, whether that’s a workflow, a process, or a way of evaluating a company, you have to be able to describe it. You have to know what you're doing and why. Through interrogation, you start to see the structural problems in your own data, your own processes and potentially your own thinking (although open to much wider philosophical debate). We did a firm-wide problem analysis and discovered that while we had goals, in several cases we hadn’t defined a real why beneath them. Finding the why became the first piece of actual work, and it was the most important thing we did.

This matches what broader research is beginning to surface. A 2026 study by Baillie Gifford's investment team noted that AI, when used passively, can bypass the effortful processing that builds understanding, but when used actively, as a tool you interrogate rather than one that answers for you, it hones your thinking rather than blunts it. The MIT research they cited found that 83% of people who used AI to write essays couldn't quote their own work minutes later, because the effort that builds durable understanding had been skipped entirely. Venture, whilst about speed, is also a constant game of learning and understanding, the key skills that should not become a shortcut.

The easiest mistake to make when you have access to powerful tools is to reach for perfection before you've earned it. Our first ambition for AI-assisted dealflow was a sophisticated system that could read the whole market, a comprehensive signal engine that would surface every relevant opportunity at scale, built properly, beautifully and all at once. What we ended up with was something that looked impressive but did almost none of what it was actually intended to do, because we'd constructed it from the top down rather than from the ground up.

The thing about building all in one go is that you don't find out it doesn't work until you're already deep into the build. An imperfect MVP, something scrappy, specific and live, would have told us the right information within weeks - information the beautiful version took months to get wrong. A start that you iterate on is always going to be more useful than a finished thing that misses the point. It took longer than it should have to realise the real, smaller, more specific problem was simply whether we had a clear view of the companies and signals we needed right now, given our current thesis and fund stage. Don't build the castle, solve the problem, and be willing to do it badly at first, because that is where the learning lives.

AI makes things faster, which does not always translate to being good.

There is an important distinction between speed of output and depth of understanding, and the two can move in opposite directions. Thomas Germain, writing for BBC Future in May 2026, described it clearly: "AI is giving us, for the first time, an easy way to trade process for product." The essay sounds better, the output is cleaner, but the mental work, the struggle and the challenges, are exactly what builds real capability. We felt this most acutely in how we were approaching research, because AI could pull together a summary of a market faster than any analyst, but if the person reading it hadn't done any prior thinking, the summary is effectively meaningless.

We experienced certain scenarios where forming a view first and using AI to challenge it, for instance a fake IC, gave us more depth. Related to this, we found that the bottleneck in deal evaluation has shifted. It has gone from data access, to now knowing what matters and when. Moving away from static company snapshots toward continuously refreshed signals i.e. what's changed in the last 30 days, what's the "why now" for this company specifically.

To make AI a useful sparring partner for investment decisions, we built a resource library, essentially a structured record of how we think, how we evaluate companies, and how our IC process works. The more context it had, the better its pushback became. However in doing so, we have now debated another question which is: if the more you train a system on your own thinking and the better it reflects you, does that also make it more likely it is to miss an outlier. An investment thesis is, by definition, a set of priors, and the most interesting companies sometimes break those prior entirely. Take Stripe, WhatsApp and Uber as examples.

This problem runs deeper than it first appears. For example, what signals are good signals? Headcount growth is a good example, because for years it was a reasonable proxy for business health and momentum, now it can just as easily signal inefficiency. A team of five running on AI-native infrastructure can do what fifty people did three years ago, so a company that hasn't scaled its headcount isn't stalling, it may be the most capital-efficient company in the cohort. If you feed AI a labelling framework built on old assumptions, it will confidently encode those assumptions and apply them at speed, so the bias doesn't disappear; it accelerates. Our response has been to build in limits on how far AI goes in the judgment chain: it scores, it signals, it challenges assumptions, but it does not conclude. How we maintain separate filters specifically designed not to deprioritise companies that don't match existing patterns is the real goal, because the best investments are often the ones that don't.

With information now readily available, you can use external sources like academic papers to challenge your own thinking. Furthermore, as we see a convergence towards the same systems, same trained data, same way of thinking with AI i.e. AI warning signs of writing, our differentiator will be our original thoughts. AI could potentially lead us to stray away from the normal by getting us to challenge it, to challenge ourselves to think differently. The road is never straightforward - challenging our beliefs on whether or not we are looking at something from the right angle may help train ourselves too.

Meetings became higher quality conversations. We use AI-assisted note-taking through Wispr Flow/Granola to generate detailed records from every meeting, and those records become the backbone of pre-reads circulated before the next discussion. When everyone has already absorbed the context, the meeting itself can focus entirely on debate, judgment, and the kind of human interaction that can't be systematised, which as a VC, is where most of the real value lives. We bought back meeting time and reinvested it in thinking time. The administrative layer, deck screening, CRM ingestion, pipeline tracking, cold outreach, used to absorb hours that should have gone to genuine deal evaluation and relationship building. Automating those tasks didn't just save time; it freed attention and time for human-led debates, discussion and judgement.

Accountability became structural rather than personal, with action items, follow-ups, and deal-tracking that used to live across scattered notes and memory now sitting in automated workflows. The smaller and more specific the system, the more reliably it holds the firm to the commitments we actually make. It also helps us aim to perform best-in-class SLAs with the small lean team we have. The unexpected win was becoming more retrospective. The process of building workflows forces you to examine how you work, because you can't automate a process you can't describe. In describing our processes, we found ourselves asking regularly, as a practice, whether this is actually the best way to do it.

If we take a step back, the running thread across these three lessons is critical thinking - plan, understand, think critically. There is some irony in this as those are the very things education should aim to teach people, ultimately they are the three pillars of a high-functioning mind.

Here, I think, the science matters as much as the philosophy. The human body is naturally akin to conserving energy such as taking the easy route or the shortcut as is the nature of survival and staying comfortable. In that sense, we are wired make our life easier and everything else seem harder. However, hard is relative and if we don't make ourselves do hard things, nothing gets easier either. AI makes it increasingly attractive to outsource cognitive thinking, precisely because it offers speed amongst many other things. This has raised two questions for me, the first being whether this is a problem of how much time we spend with AI, or of how intentionally we use it? The second however is slightly deeper; AI has enabled us to produce more and at at a quicker rate. If these time expectations have been carried forward, have we then prioritised greater output over honing our craft and development?

We use AI as a sparring partner, not a decider. Its job is to challenge our thinking using data, surface angles we might not have considered, and apply pressure to assumptions we've made quickly. Howeber what it cannot do, and what we don't ask it to do, is make the call. Tom Slater, writing about this dynamic for Baillie Gifford, put it plainly: "The people who will thrive are not those who use AI the most, but those who can still think without it." A six-month longitudinal study he cited found that as participants used AI more frequently, their actual performance declined even as their confidence grew, with the gap between how well they thought they were doing and how well they were actually doing reaching nearly 35 percentage points. That finding shapes how we use these tools.

When it comes to decisions and interactions, we've been deliberate about where the human layer sits. The expertise, the read on a founder/investor, the pattern recognition built from years in the market, none of that is outsourced. What AI does is make the surrounding infrastructure sharper: better prepared, well organised and pressure-tested before we join a meeting room.

We invest in early-stage companies, many of which are building with or on top of AI, and our own experience of building gives us a different vantage point. The firms and founders who get the most from AI are not the ones who adopt it fastest or deploy it most broadly, they're the ones who are clearest about the problem they're solving, disciplined about where human judgment is irreplaceable, and willing to do the harder work of mapping their architecture before they automate anything.

The architecture is stronger, the workflows are more coherent, the meetings are more succinct, and we have a much clearer sense of where AI sits in our laundry list of tasks as well as where it doesn't. A system trained on your own thinking will help reduce the initial hurdle however will also surface gaps of your own. The temptation to shortcut understanding rather than hone it is always there. The distance between A and D can still swallow the B that was the initial problem.

What we went looking for was efficiency and what we found was a much clearer understanding of how we work, what we're for, and what only we can do. That seems like a reasonable return on the investment.