COUNCILS BLOGPOST - JUNE 2026

The Bottleneck Moved.

AI is everywhere in the portfolio now. Value isn't. Across three councils, a summit, and a survey of close to fifty companies, here's the pattern underneath and why the hard part was never the model.

Over the past three months, we've convened more than eighty people — founders and operators across five councils, covering tech, marketing, finance, product and people, and with founders and LPs at our recent AI summit in San Francisco. Alongside the rooms, we ran a survey across nearly fifty companies in our own portfolio — early and growth stage, spanning eleven countries and fourteen industries. The survey gave us the numbers. The councils told us why the numbers look the way they do.

Here's the short version. AI is everywhere now. Value isn't. And the reason isn't the one most people reach for.

The instinct is to blame the model — not smart enough, not reliable enough, not ready. That's not what operators are seeing. The model is doing its job. What keeps happening is subtler and more useful to understand: every time AI removes a constraint, the constraint reappears one step downstream, somewhere the team isn't measuring yet. The bottleneck doesn't disappear. It moves. And most companies are still pointing their dashboards at where they used to be.

That one pattern explains almost everything else we heard.

Everyone's running it. Almost no one can prove it's working.

Start with the number that hasn't budged. A year into serious use, 58% of growth-stage companies (surveyed across our portfolio & reinforced by operators in the wider ecosystem) still can't say whether AI is delivering measurable value — the same share as a year ago. Not "no." Just: “too early to tell”. It’s what we’re referring to as the value-capture gap, and it's the defining finding in the data is tha usage is universal but proof is not.

The councils put faces on it. A product leader described shipping 80–90% of his code with AI, with every product manager equipped with agents — for roughly 20% more value reaching customers. "Where are the missing 60%?" A finance leader, asked whether AI shows up in the P&L, gave the honest version: you can't isolate the AI from everything else that changed in the same six months. A people leader said she can watch her team use AI every day and still not tell whether it's making them better. Different functions, identical hole.

“The bottleneck didn't disappear, it simply moved.”

So where did the value go? It didn't evaporate. It went to whatever the team wasn't watching.

In engineering, the gains are real and measured — 68% of growth-stage companies say engineering is where AI lands hardest, and heavy users report saving five to ten hours a week. But clear the constraint on writing code and a new one appears at reviewing it, trusting it, and getting it into a market that still moves at its old pace. One product team got a request no one in the room had heard before in their careers: the go-to-market side asked them to slow down. They were shipping features twice a day and the people meant to sell them couldn't keep up. The bottleneck had walked from engineering to distribution, and nobody had moved the dashboard with it.

In finance, it moved to verification. Operators described checking AI "like a junior" except a junior tells you when they're unsure, and the model never does, so you check everything and hand back the time you saved. In people, it moved to the humans themselves: the company can suddenly attempt far more, and the real limit becomes how much change the team can absorb at once.

None of this is new, either. Our March councils with marketing and engineering leaders already found it in a single line — "coding is not the bottleneck anymore; humans are." Three months on, that's true of every function, not just engineering.

You're paying twice and the old benchmarks are breaking.

While value stays hard to see, cost is easy to see, and it's climbing. AI budgets are roughly doubling; the share of growth-stage companies spending more than €100k a year on AI jumped from about a fifth to over forty percent in a single year. And almost nobody is taking anything out to pay for it: only 16% have clearly reduced their legacy software spend. This conundrum is what we’re calling the dual-cost trap — AI bolted on top of everything you already run.

The finance council showed exactly where the line falls. The systems of record stay — the ERP, the accounting tool — because the data lives there and someone has to own whatever might replace them. What's actually getting switched off is the reporting layer: licenses for the old dashboards cut and replaced by something AI builds on demand. (As an aside that tells its own story: in 2025 OpenAI led the portfolio; in 2026 Claude is the most-used tool across it, and Cursor has become the default for writing code. The tools are turning over fast.)

Underneath the spend sits a harder number nobody has settled. Counted honestly, the 80% software gross margin that has anchored how these companies are valued for fifteen years looks more like 50–60% once AI cost is in the figure. No one in the room knew what "good" is supposed to mean any more — and that question reaches well beyond the finance team.

The org chart is being redrawn and the data and the discussions don't agree.

Headcount is where the change is most visible, and where the numbers and the conversations start to diverge. In the survey, 45% of growth-stage companies are reducing headcount, and the level most affected is entry and junior — the roles whose repetitive work AI does well. The finance room confirmed it from the ground: the junior, data-entry work is the first to go.

But the people leaders weren't sure that's the right call, and said so. The junior role is at once the most exposed to automation and the most native to the tools — the youngest people are often the most fluent in them. Cut juniors to save cost, and you may be trading away the talent that adapts fastest. So the verdict split: the survey says juniors go first, and the people closest to the decision aren't convinced they should. Nobody resolved it — and that unresolved gap is more honest than a clean answer would be. (Early-stage companies, for their part, are reshaping roles rather than cutting — same technology, different directions by stage.)

Higher up, senior people are running smaller teams and spending their days reviewing an agent's output instead of doing the craft they loved. The most honest moment of the quarter came from a people leader on the human weight of all this: "I don't think we handle it." By default, the cost of the transition lands on the people team — without a mandate, and without anyone having decided it should.

The agent era is mostly still a copilot era.

One more thing worth saying plainly, because the hype around it is loud. Only 13% of companies are running anything genuinely end-to-end; most of what gets called an agent is still a scripted workflow with a better name. The gap between the market's story and the operational reality is wide. Pretending otherwise is how you end up buying the demo instead of the result.

What the operators who are ahead actually do.

None of this is an argument to slow down. It's an argument to look in the right place. The companies pulling ahead aren't the ones with the best model access — everyone rents the same models. They're the ones who find the new bottleneck before it costs them, and who stay honest about what AI is and isn't doing.

In practice it came down to two moves. Know where AI is actually creating value, not just where it's busy. And be willing to take something out to pay for it — switching off the tool it replaced rather than running both. Underneath both sits the unglamorous work almost nobody enjoys: building the measurement and the feedback loop that let you prove any of it. The defensible position was the same in every room we sat in — not the model, but what you build around it: the proprietary data and the loop that gets better the more your product is used, the things a competitor with the same tools can't clone in a weekend. The model was never going to be the hard part; the hard part is everything around it, and that's the part you can still win.

What struck us across all of it wasn't pessimism. It was operators getting specific, being done with the hype, and asking better questions than the click-bait headlines would suggest.

That's what these councils are for, and we're running more of them. If you're an operator with a view on any of this, or you want a seat in one of our next rooms, drop us a line!