AI Is a 200 km/h Train. Most Companies Still Run 40 km/h Tracks. Five Questions to Fix That.

hero-image
calendar
Nov 17, 2026

On November 12th, right after a day when Web Summit was canceled due to a storm in Portugal, we attended several thought-provoking sessions on enterprise AI.

Halfway through the “Ready or not here AI comes”, Martin Schroeter, CEO of Kyndryl, shared the image that stayed with us longest: “You piloted this brand new 200-kilometer-an-hour bullet train called AI, but now you want to take that bullet train and put it on the train tracks that have been built for 40 kilometers an hour.”

If your company ran an AI pilot last year, you probably know the feeling. The demo worked, but  production didn’t happen. Speakers on different stages kept discussing the same matter from different perspectives. It’s not the model that stops the train. It’s the tracks: what the company wants, who decides, who controls the data, and whether people are ready to work differently.

Pilots Prove the Technology. Production Tests the Company

Kyndryl runs infrastructure for banks, airlines and insurers, with about 70,000 people in 60 countries. Schroeter’s summary of why pilots stall was blunt: “The pitfalls from pilot to production are not the technology. You’ve proven the technology works.”

What remains is everything around it. Old systems that were built for a slower train. People who were never asked to change how they work, governance that nobody designed. And then there are regulators whose questions, in Schroeter’s words, are fairly simple: “where is your data right now and who can see it.”

Christos Rizos, Group Business Development Director at AustriaCard, opened his masterclass on “AI sovereignty: Data control, governance and predictable AI costs” with the same split. Last year, everyone was experimenting. This year, everyone is putting AI into production, and production comes with questions a pilot never had to answer.

Rizos then showed what that shift involves. In a demo, you upload a folder of HR policies, connect a model, and get answers. In production, the same tool depends on several things: the model’s API, the embedding service that searches your documents, sign-in, and the connectors to your databases. If one link fails, the tool stops. Rizos’s advice was to start with measurable tasks and define success and failure criteria before building, with control counted among them: who can see which documents, alongside accuracy and speed. Schroeter explained why this step pays. In Kyndryl’s latest readiness survey, about two-thirds of customers already see a positive return from AI, and he was clear that the return comes only once a pilot reaches production.

Notion image

We’d add one point from our own work. AI speeds up whatever it touches. A clean process gets faster. A process with a broken hand-off between sales and accounting gets faster too, and so do its mistakes. A pilot can hide this because it runs on a tidy sample. Production runs on your real Tuesday.

Question for your next planning meeting: which process would you be comfortable seeing run ten times faster exactly as it runs today?

AI Works Only If You Know What You Want

During the “Beyond the AI hype” panel discussion, Tariq Krim, Founder of Human Intention, said: “AI in general only works if you know what you want.” The two required skills are knowing what you want and evaluating what you got. The work in between gets simpler and faster.

What does that look like in software teams? Engineers used to build brick by brick and put the door where the door goes. Now there is “this stream of tokens shooting with high pressure out of the model”, and the job becomes building gates, tests, and evaluations that mold the stream into the shape you need. Once a team masters that, in the speaker’s words, “you are now 100 times faster.” At a medical company, shipping a major product went from a year to a week.

The same logic holds outside engineering. An order-to-cash process, a claims queue or a month-end close can be automated quickly once someone has written down what “done right” looks like  (which fields, which exceptions, which approvals, what counts as an error). A pilot that starts with a tool and a hope rarely has that definition.

The panel also addressed the noise around this. Georg Zoeller, Co-founder of the Centre for AI Leadership (C4AIL), pointed out that every company now needs an AI narrative, so a toothbrush with a pressure sensor that has sat on shelves for 15 years is suddenly “AI” and costs $200 instead of $45. If the pitch you’re hearing can’t say what the process will produce and how you’ll measure it, you’re looking at the toothbrush. It’s just an old thing with an AI sticker, a higher price, and no real change in how work gets done.

Question: for the process you’d automate first, can you describe a correct outcome in one paragraph that your CFO and your operations lead would both agree on?

The Weekly Report Test

The “AI’s Octopus Era” panel took the idea further, looking at how companies are organized and comparing them to an octopus. It keeps its brain for guardrails and strategy, while the arms decide locally. AI could push decisions down to the people closest to the data.

Notion image

Max Schireson, Partner at Battery Ventures, argued that the leader’s job becomes creating an organization where the people closest to the work have the context to decide. Much of what executives decide in a day, he said, could be decided better and faster by the people who brought them the question, if those people knew the priorities. The example was concrete: tell everyone that this year the company is shifting “from a pure growth focus to a mix of growth and profitability”, and they can make their own calls in line with it. The moderator, Jim Edwards, Executive Editor of Fortune, added that “everyone on the planet has had those days of work where when your boss leaves to go on vacation, all of a sudden it’s a lot easier to get the work done.” It got a laugh because most people who’ve had a manager know the feeling.

Then came the uncomfortable part. Asked what happens to middle roles, Schireson said that people who use the new tools to add value will matter more than ever. “If they don’t, and their job is just production of a weekly report, then you probably don’t need them anymore.”

For a mid-market company, this cuts in an interesting direction. It has fewer layers than an enterprise, so the octopus is already closer to reality. But it probably also has a few roles, or a few hours in many roles, that exist to move information from one system or one person to another. Those hours are where automation pays back first, and they’re also where the company’s leadership needs to decide what those people do next.

Question: how many hours a week does your team spend producing reports that someone else reads to make a decision you could have pushed down?

Control Is a Cost Line

The “AI sovereignty” masterclass by Christos Rizos sounded at first like a European policy debate. It turned out to be a session about money and risk that any CFO would recognize.

Rizos split the topic into three parts. Sovereignty is who controls your AI, governance is how you exercise that control, and predictability is mostly about cost. On cost, Rizos shared a personal example: coding experiments that consumed half a billion tokens in a single day. On a $100-a-month personal subscription, that was covered. At business pricing, he estimated the same day at about $2,000, with 90% of the tokens being input. Agents work around the clock, and they consume tokens while you sleep. That’s why a pilot priced on a flat subscription says little about the production bill. There should be a line for usage.

Notion image

Two other points deserve a place in any AI business case:

  • Your model can change under you. When a provider releases a new version, a workflow you tested may behave differently. Rizos’s team saw newer models get better at general topics and worse at the narrow task a company needed.
  • “Local” data can still leak. In a typical setup, an HR policy library stays on your servers, but each question sends the most relevant passages to an external model. Question by question, the document goes out in pieces. Keeping data in your country, Rizos added, isn’t the same as controlling it, since a US provider can be compelled under the US CLOUD Act to hand over data wherever its data center sits.

Schroeter approached control from the security side. Agents, he said, are “relentless goal-seeking bits of software code”, and they need to be controlled like any other code with access to your systems. Some US banks have decided to deny third-party agents in their infrastructure rather than wait for regulation. The companies Schroeter sees getting it right do three things: they know exactly what they have, they build governance into the design, and they can isolate or shut an agent down if something goes wrong.

Before you scale, you should know who can ask the system what, which documents it can reach, which model answers, whether a human reviews the result, and whether every step is logged. Rizos listed exactly those five checkpoints for any workflow (ask, retrieve, answer, review, trace).

Question: if your AI provider changed its model next week, who in your company would notice, and how?

Typewriters Next to the PC

The last idea came up a few times. It is about people.

“Our customers tell us by and large their people are not ready for AI,” said Martin Schroeter. The fear is logical, because people ask what happens to their jobs. His fix inside Kyndryl was education built per role: a designer, a coder, and an engineer use AI differently, so a generic 20-minute training doesn’t change much. He also reminded the room that when the PC arrived, typewriters stayed in offices for years before people changed how they worked.

The “Beyond the AI hype” panel didn’t agree on where this ends. The moderator, Jeff Wilser, host of the AI-Curious podcast, would rather see four people plus agents do the work of eight than see a team of four cut to two. The opposite view came up too: “I think there will be fewer humans in many of these jobs.” The reasoning was that companies treat agents as humans to swap in, when the technology works more like a process and an operating system. Between the two views sits a practical rule for any company: automate what you’re not good at, and keep what makes the work yours.

For a founder, this is a design decision more than a philosophical one. Before an automation goes live, decide who gets the hours back and what they’ll do with them. Teams adopt what makes their day better. They probably quietly route around what they suspect is there to replace them.

Question: for your first automated process, have you named who will own it day to day, and what they’ll stop doing?

Where to start

Wilser closed “Beyond the AI hype” with a prompt he gives CEOs, which we think beats any tool list as a starting point: imagine you could double without spending a nickel. What would you do?

The answer usually points to one process, the one where growth currently means hiring more people to move the same data between the same systems.

What is that one process in your company? Let’s start with it.

Insights
code-unsplash
calendar
Nov 18, 2026
ISO Certifications Explained: What They Mean and the Five Cody Holds
ISO certification means an accredited body audited a company’s management system against an ISO standard. Here’s which types matter for a software vendor, and how to check a certificate.
avatar
clock
9 min
code-unsplash
calendar
Nov 10, 2026
Why Most AI Pilots Never Reach Production, and How to Get a Working Prototype in Two Weeks
Most AI pilots stall on data, integrations, security reviews and ownership. What changes between a demo and production, and how a forward-deployed engineer builds a working prototype inside your systems within two weeks
avatar
clock
12 min
code-unsplash
calendar
Nov 5, 2026
AI Act Certification: Which Standard Fits Your Company
What the AI Act requires and when, and how three certification routes compare. Notes from the ECCP webinar with a regulator, a certification body and a legal researcher.
avatar
clock
5 min
code-unsplash
calendar
Nov 3, 2026
FDE, Consultancy or In-House Team? How to Staff Your Next AI Attempt After a Pilot Didn't Deliver
How to staff your next AI attempt after a pilot didn’t deliver.
avatar
clock
6 min
code-unsplash
calendar
Nov 2, 2026
AI Hasn't Paid Off Outside Tech, Says Apollo. Uber's Two-Week "AI Pods" Show What Changes That
AI spending went up. Costs didn’t come down. What Uber’s two-week AI pods show mid-market CEOs and CFOs about getting AI out of the pilot.
avatar
clock
4 min
code-unsplash
calendar
Oct 28, 2026
Kairos Launch at Oslo Innovation Week 2026
AI in Finance, Oslo: what works, what doesn't, and why we launched Kairos there
avatar
clock
4 min

WANT TO DISCUSS A PROJECT?

Let’s get in touch

SERVICES

PERSONAL DETAILS

Let`s talk
Follow us on social media