My day at the summit
Seven sessions on the main stage. I skipped the security and start-up panels, so there are no notes from those. All sessions were streamed on Yle Areena: the morning, the economy panel and the rest of the afternoon.
- 10:15Opening words
- 10:35ELLIS keynote
- 10:55AI & Research
- 11:35AI & Enterprises
- 13:15AI & Economy
- 15:00AI & SecuritySkipped
- 16:00AI & StartupsSkipped
ELLIS Panel, 10:55–11:35
Research: the next leap also needs rebels
Moderated by Eric Malmi, with Danica Kragic, Matthias Bethge, Kuan Yen Tan and Erkki Oja.
On stage
- Before the panel, Samuel Kaski’s ELLIS keynote: the institute has recruited researchers from around the world (I noted 19), and health is its first big research focus.
- So far, models have been trained on human-made data. We are now entering a time when AI-generated data has to carry more of the load.
- Robots do backflips, but they break a lot and are expensive. Materials still need real innovation, and human senses remain far superior.
- Europe is not ahead. Quantum computing could add speed.
- Erkki Oja, the inventor of Oja’s rule, was a rebel in the 1980s who did not believe in rule-based expert systems. With Teuvo Kohonen, he was one of Finland’s early neural network pioneers.
- LLMs are the dominant trend with no strong challenger, and we no longer really know how they work. They run on text, which is a rigid code. A four-year-old knows things without reading a word.
- Backpropagation (Hinton et al., 1986) took 40 years to go from idea to working systems. We need out-of-the-box thinking alongside LLMs.
- Jobs are going to change and we need to adapt very quickly. Oja’s view: the intelligence of LLMs is overstated.
ELLIS Panel, 11:35–12:15
Enterprises: stop piloting the edges, redesign the core
On stage
- Pekka Lundmark: small companies have an adoption gap, large companies a scaling gap.
- Value comes when you redesign the work. That happens more in start-ups, which is why this is not a question for the IT department.
- Lundmark: redesign your core processes using AI, not 30 pilots. It has to be on the CEO’s and the board’s agenda, with a dedicated team, freedom to act and the CEO’s backing.
- Stage 1 was creating and distributing models. Stage 2, real use of AI, is only beginning.
- The data at the heart of industry is the next important training material, as the internet’s data has already been used. The bottleneck is building transmission grids; new energy can be brought online quickly.
- Running this transformation the traditional way, through middle management, may be too slow. That layer may also be concerned about its own future.
Using a tool is not the same as redesigning the work. The 38% is from Statistics Finland, Use of information technology in enterprises 2025. The 11% was cited on the panel.
Also on stage
- In Europe, IT is still seen as a back-office function, and AI too often lands there. The business should be the driver.
- Changing processes has been hard partly because systems have chained us: change is hard and expensive. That constraint may no longer exist.
- A company’s edge is coded into its ERP, but there is no manual for how it got there. Company knowledge hides in Oracle and SAP silos.
- Accenture’s view: a new AI layer will sit on top of the systems of record.
- Real success comes from co-creation, not from procuring process automation.
Will AI sit on top of the ERP, or change it?
I’ve written about the possible scenarios at more length. Here is the short version of how I see the paths.
Long readCode became cheap. What happens to systems of record like ERP?A new layer on top
AI sits above SAP and Oracle, and the ERP becomes a back office of interfaces and business logic.
My view: that alone doesn’t justify what these systems cost, and changing them stays expensive and complex.
AI inside the ERP
Embedded AI arrives in the core systems themselves.
My view: it helps, but it can’t fix processes that nobody fully understands. This transformation will also be slow, and will it ever really match your processes?
Rebuilt beside the ERP
The process that gives you an edge is rebuilt next to the ERP, one process or module at a time.
My view: this is the one I’m watching. It starts from the company and from AI, not from the system.
Why have companies shaped themselves around the processes their systems offer, when it should be the other way round?
The question this panel left me with
Panel 1, 13:15–14:30
Economy: companies run on the same manual data work
On stage
- New technology lowers the cost of processes.
- Björn Wahlroos: the biggest change will come in the public sector. A large number of people handle simple tasks there, and there is no price component. With AI, the welfare state could become affordable.
- Edward Jung on what AI should do for people:
Does it make you feel better, or does it make you better? Heroin makes you feel better.
- Jussi Westergren: the tools keep changing. If you think Claude is the last thing we’ll see, you’re wrong. Europe needs more ambition.
- Trust in the leaders of the leading AI companies is low, and that is why people are scared.
- Big tech, not carmakers, leads in autonomous driving. How did that happen?
- Westergren and Peter Sarlin: Europe has huge potential, but the fight for the brightest minds is on, and many have already gone to the US.
- Edward Jung: Europe is the one asking what the human role is in all of this. He sees that as one of the biggest differences from the US and Asia.
- The most entertaining panel of the day, and it didn’t feel overly scripted.
The same flow of talent and capital from Europe to the US shows up in The Draghi Report on EU Competitiveness (European Commission, 2024) and in studies by the European Commission Joint Research Centre (JRC) and the EIB (2026).
What I take home
The same pattern keeps repeating. The biggest gains come from rethinking a company’s core processes.
InteractiveWhat is a business? A process viewThat is also the hardest part. Pekka Lundmark, who has led large organisations, said it plainly: changing processes is very difficult. And I agree with Accenture that the rigidity and cost of systems has held it back too.
So are we entering a time when this starts to change, one process or module at a time? When the process that gives you a competitive edge is rebuilt beside the ERP, starting from the company and from AI?
A thread running through the day
Where is the productivity leap?
On stage
- Speakers kept coming back to the same point: the productivity leap from AI does not show in the statistics yet.
Steam took decades to show up in the numbers
Steam’s contribution to labour productivity growth in Britain, in percentage points per year.
Steam only matched water as a power source in Britain around 1830, and it realised its potential with high-pressure steam after 1850. Source: Nicholas Crafts, “Steam as a General Purpose Technology: A Growth Accounting Perspective”, The Economic Journal, 2004.
Meanwhile, AI runs on a different clock
Steam needed six decades before it even matched water power. Between last year’s AI Summit and this one, twelve months were enough to shake whole markets.
- November 2025Three frontier models in twelve daysOpenAI’s GPT-5.1 (12 Nov), Google’s Gemini 3 (18 Nov) and Anthropic’s Claude Opus 4.5 (24 Nov). Opus 4.5 passed 80% on SWE-bench Verified, a test of fixing real software issues.
- January 2026Agents move into office workAnthropic introduced Claude Cowork on 12 January, and on 30 January added plugins for legal, finance, sales and marketing work.
- February 2026The “SaaSpocalypse”Software stocks fell hard. Thomson Reuters dropped 16%, and a software stock index lost about a quarter of its value between 12 January and 23 February. Investors began questioning per-seat licences and billable hours.
- February–March 2026The release pace holdsGoogle’s Gemini 3.1 Pro and OpenAI’s GPT-5.4 followed within weeks of each other.
- April 2026An AI that finds zero-daysThrough Project Glasswing, Anthropic gave partners such as AWS, Apple, Microsoft and CrowdStrike access to Claude Mythos Preview, which had found thousands of vulnerabilities in every major operating system and browser. Cybersecurity had to rethink how flaws are found.
- April 2026Another wave in one weekClaude Opus 4.7 (16 Apr), OpenAI’s GPT-5.5 (23 Apr), pitched for real work and agents, and DeepSeek’s open-weight V4 preview (24 Apr).
- 16 November 2026AI Summit 2026The productivity leap still doesn’t show in the statistics.
My take: the pace is almost incomprehensible, and the market has already reshaped itself around what these models can do. The real effects on productivity will show a little later.
Sources: model release dates as reported by the companies and tech press; software sell-off from The Batch; Project Glasswing from Fortune; GPT-5.5 from Fortune.
How software development has been priced
time × hourly rate= ?
Would I recommend it, and go again?
Absolutely.
You will hardly see a line-up of speakers this strong at any other Finnish event, on a topic this current. It almost annoys me that I have to wait a year for the next one.
Maybe a new model for events could be smaller gatherings, more often, because AI moves like a bullet train.
Then again, many panelists referred back to what they had said a year ago. Some of the fundamentals and challenges have stayed much the same.
Have you seen a company reshape a core process instead of running another pilot?
I’d like to hear how you see it.