Waiting on AI won't get you fired. Or will it?

The conversation that goes nowhere

I have had a version of this conversation dozens of times this year. It usually starts at a conference, and it starts well. The person across the table is an executive at a portfolio company, or the operating partner who looks after it. They are energetic and specific. Portfolio monitoring is done by hand. The data is incomplete. Reconciliation takes days. Everyone in their network is "doing something with AI" and they know they should be too.

The problem is real and clearly defined. There is no direct line to a solution, but that is what a second conversation is for. Then comes the sentence that ends it: "We'll discuss it at the next board meeting." Or: "We need more time to align internally." Months pass. Nothing happens. Not a no, not a yes. Just a slow fade.

This is not bad luck or a few indecisive people. It is the norm. At the SuperReturn AI Summit in Berlin this June, one panelist described a workshop where only 8% of 50 investors had actually worked out and prioritised their AI use cases. His summary of the state of most portfolios: "We have more pilots than Lufthansa."

So how can decisions take months in a market that changes weekly? I have been trying to answer that for myself, because the answer decides what I should say in that second conversation. I think there are three reasons. Short-sightedness plays a part in each, of a very understandable kind, but none of them is stupidity.

Three reasons the decision never gets made

1. Nobody knows what the right decision is

Everyone can use ChatGPT. Very few people have understood the layer underneath it, or seen what the best engineers actually do with these tools: agents that run on a schedule, company knowledge bases the model can read, connectors into live systems, loops that keep working until the job is done. Compared with that, a chat window is a toy. None of this is a question of intelligence. Understanding what changed in the last three years takes fundamental, unhurried learning, and most executives have no time for it. They rarely admit this, to others or to themselves.

The result is that the practical questions have no owner. Who should we hire, and how would we even evaluate that person? Should we buy or build? Does building make any sense when the frontier labs release something new every week? What should change in our processes, and what must not be touched? You cannot make a decision when you cannot name the options.

Meanwhile, the other side of the market is making the problem worse. Silicon Valley is now building tools for problems that traditional businesses do not have yet: frameworks to evaluate AI output, routers to pick the right model per task, dashboards to monitor token spend per use case, systems that trace the decisions of autonomous agents. Every one of these layers solves a real problem for teams that already run AI in production, and every one of them adds another term a traditional executive has to decode before joining the conversation. So the vocabulary keeps moving away from the people who most need it.

2. Waiting looks free

Big bets need courage, and once reputation is on the line, waiting seems like the rational strategy. Let the market sort itself out. A leader will emerge among the tools, and once it is popular enough we will adopt it. No wasted budget, no security incident, no chaos, no being the early adopter who got it wrong.

The flaw in this reasoning is that nobody gets fired for waiting. That part is true. What happens instead is that the decision gets made over your head. The board decides. The fund decides. A peer company in the portfolio proves the case and your operating model is templated from theirs. At SuperReturn, James Stevens of Bain Capital named the number one predictor of a top-performing portfolio company: a CEO who is curious and technically literate. Aaron Rudberg of S2G Ventures put it more bluntly: "If a CEO isn't talking about AI regularly, you have the wrong CEO." The asymmetry of risk is an illusion. The decision is being made either way; the only variable is whether you are in the room.

3. It might all be a bubble

And then there is the belief that the whole thing is overhyped. Language models are not as transformational as advertised, AGI is speculation, the frontier labs raise billions and spend more, mostly on compute, heavily subsidising a product with no unit economics to speak of. Why move my workflows onto something whose price might triple next year?

I want to take this seriously rather than dismiss it, because parts of it are correct. There probably is a bubble in the capital markets around AI. Usage probably is subsidised, and prices may rise. But both things were also true in 1999, and the crash did not rescue the retailers and publishers who had waited for clarity. A bubble and a transformation are compatible. The 2000 crash killed Pets.com and left the internet standing.

There is also a practical hedge that has quietly emerged. The model is becoming a commodity. Nils Rode, CIO of Schroders Capital, said on stage that "LLMs like OpenAI or Claude will quickly become a commodity; we regularly swap them." Permira, CVC and Bridgepoint said versions of the same thing. If the cost risk sits with whoever bet the company on a single vendor, then the way to manage it is to build workflows that can swap models, not to build nothing.

Add these three together and the stall makes sense. After the initial shock of ChatGPT, most traditional businesses did not experience an earthquake. They kept their clients, users and patients. Nothing forced a decision, so no decision was made.

Meanwhile, on the other side

For me the question of whether AI is transformational is settled, because I see it at three levels every week, and none of them requires believing a lab's press release.

Individually. I am not an engineer. I came to this company as a designer and I run it as a CEO. My own productivity has roughly quadrupled, and I am not an unusually advanced user. The engineers I work with report more, and the good ones report a lot more.

As a company. Vecten is a 40-person firm. Over the last year we realised three to four times as many of our OKRs as in any previous year, with the same headcount. We did not become smarter. We changed how work gets done, and the numbers followed.

For a client. We ran a multi-year platform engagement for a PE-backed healthcare provider with a traditional six-person team on time and materials. When the client needed a significant budget cut, we did not shrink the team and accept less. We rebuilt the delivery model around AI: meetings transcribed and turned into tasks automatically, engineers reviewing AI-drafted solutions instead of writing from scratch, releases and reports generated by pipelines. The six-person team became two. The budget fell by 57%. Output rose by 50%. Per dollar spent, that is roughly 3.5 times more software than before, and the full product roadmap through September 2026 was kept intact.

Across portfolios. The organisational evidence is now public, and it was the theme of SuperReturn's AI track this year. Google walked through a property and casualty insurer whose auto-claims process took six to seven weeks across eight to ten manual steps; after automation, roughly 94% of the people came out of the process and claims-margin EBITDA rose by 250 basis points, proven with A/B tests across markets. EY-Parthenon turned an automotive client's procurement data into something operators could simply ask questions of, and surfaced enough savings on multi-billion-euro spend to cut about €400M. Aurelius put a coding agent that writes SQL queries onto one company's inventory data and found €2.5M of stock value in about two weeks. Accordion layered sentiment analysis on top of a facility-services company's existing churn model, routed the scores to reps, and cut churn by around 10%, worth $18M of revenue on a $150M business.

And one anecdote that stayed with me. Lixia Zhu of 76Columbus Family Office built a fully working mobile app in five weeks, using Claude and a no-code tool, that would previously have taken a 20-developer team six months. She now asks every fund manager she meets which of their portfolio companies a language model could simply wipe out.

Then there is the number that describes the moment we are in. Ewa Treitz of AWS reported that of the roughly 18,000 AI projects AWS has run with PE-backed companies since 2022, 94% were single-task automations: a copilot here, a chatbot there. Only 6% re-engineered a whole workflow end to end. But almost every new project started in the last six months falls into that 6%. The adopters have just changed gear.

Polarization

Put the two halves together. One group is standing roughly where it was in 2023, having survived the initial shock and concluded that not much needs to change. The other group has moved from copilots to rebuilding workflows, and its gains compound: every automated process frees people to automate the next one. A flat line and an exponential curve look similar for a while. Then they do not.

This is what I mean by polarization. It is not that some companies use AI and some do not. It is that the distance between them is growing faster than the laggards can perceive, because the laggards are measuring themselves against last year, and the adopters are measuring themselves against last month. The same panelist noted that every organisation splits into thirds: one waiting eagerly for these tools, one willing to learn, one resistant. The market is splitting the same way.

What makes it dangerous is that the erosion is silent. Take law firms. More and more people tell me they no longer bring routine work to their lawyers. The hourly rates are high, the day-to-day retainer tasks are manageable with AI, and the lawyer is called for final verification or the hard cases. The firm does not lose the client. The client just sends less, and less, and never mentions why. I think of this as silent resignation, and it will happen to any business whose moat is a relationship its customers can partly serve themselves.

As with every cycle, a large share of businesses will not make it through this one. Some will barely catch up. The ones that get more time are those that made a deliberate bet, in isolation from the core, or were built from day one on first principles. If you are reading this and recognising your company in the first group, the rest of this piece is for you.

What to do if you are on the wrong side

This is for the executive who is actively exploring, reading, testing tools, evaluating vendors and talking to people, but still has no plan they trust. Six moves, in roughly this order.

Set the guardrails first

Decide what can and cannot go into which tools, where your data is allowed to sit, and who signs off. For EU companies this matters twice: it is a compliance requirement and, increasingly, a selling point. Helena Malikova of the European Commission noted at SuperReturn that institutional investors now actively want their data stored in Europe; sovereignty has flipped from cost to advantage.

Guardrails are what let you move, not what stop you. Devika Shanker-Grandpierre described an energy company where five promising pilots died at the General Counsel's desk. The fix was to sit legal, compliance and safety down to define an acceptable level of risk together, then sort use cases: low-risk ones get a fast lane, high-risk ones get red-teaming. Her mantra: "say yes, but safely."

Appoint someone

You need one person who understands this and believes in it: an internal owner, a consultant, or a new hire. Without that person you will always be behind, because you will always lack a reliable account of what is happening. This applies to you as well. You do not need to become an engineer, but you need enough curiosity to look under the hood and a mid-level understanding of the trends. Most of all you need a direct line of communication with the expert. This cannot be fully outsourced.

Often this means hiring people who get it rather than converting people who do not want to. BCG's framework, echoed by several practitioners on stage, puts the technology at 20–30% of the value of an AI transformation, and people and process at 70–80%. Firms consistently overspend on the 30% and underinvest in the 70%.

Get the data sorted

One speaker on the data panel summarised it as "IA before AI": fix the information architecture first. Point a strong model at a messy SharePoint and you get confident wrong answers and a failed project. Centralising, cleaning and maintaining company data in one reliable place is the unglamorous 80% of the work, and it is where AI projects are won or lost. This is the business we are in, so take the recommendation with that in mind, but I have not met anyone on either side of the polarization who disagrees with it.

Make it a routine, not a project

I know it is a cliché to call AI a mindset shift, but it is one, and the reason matters. AI is not something that happens once. It is like a mature decision to work on your health: a diet and a training plan lead nowhere if all you do is one heroic month. You change your lifestyle. In a company that means a recurring meeting, an exploration budget, and protected time for your team to try things, with no assumption that any single solution will close the topic for good. The conference had the same metaphor: Cornelia Andersson of With Intelligence compared data hygiene to a gym membership. Everyone signs up in January; by May most have quit.

Build alongside, do not rewire

Do not start by changing the culture, the systems, or the processes that are deeply embedded in your team and your customers. Those are hard to change, they work well enough, and they are the wrong place to experiment. Instead, build something new: a service line, a portal on top of the existing applications, a process run by a new team. Do it independently and, as far as possible, without permission, so you step on nobody's toes and change nobody's responsibilities.

Then prove it. Dominic Gallello of Bridgepoint put the rule well: "One success is an experiment, two is a playbook." Once you have two, roll it out across the company with proof in hand rather than a bet. That was exactly the path in the healthcare case above: we tested the new delivery model on ourselves first, then on one engagement, and only then made it our standard offer.

Ship something

Get out of the vacuum of ideas and hesitation. Better or worse, start building and experimenting. If your company will not let you, because there is no approved tool, play in your own time with no company data and bring the ideas back. And decide before you start how a pilot ends. The most common reason pilots stall, according to the SuperReturn panel on execution risk, is that nobody agreed on a kill switch: the conditions under which the pilot either scales or is cleanly stopped.

Stay close to the people who pay you

One last thing, and it is the one I would keep if I could keep only one. Your existing relationships, subscriptions, contracts and processes are your biggest asset right now. They are the reason the earthquake has not reached you yet. Appreciate them, and do not assume they last.

Stay close to your customers, users, patients, whoever they are. Measure the relationship. Ask them directly how they are using AI and what they would rather do themselves. Do not let them out of your sight, because when they leave, they will leave the way clients are leaving law firms: quietly, a little less each quarter, without a conversation.

The decision about AI in your business will be made. The board will make it, or the fund will, or your customers will make it for you by drifting away. The only question still open is whether you are the one making it.