In this Rise of AI Virtual Chat episode 17, recorded in September 2020, host Fabian Westerheide talks with Christian Guttmann, Chief AI Officer at TietoEVRY, founder of the Nordic AI Institute and professor of AI, about how companies build a realistic AI roadmap. Guttmann, who has a psychology background and a PhD in distributed multi-agent systems, reflects on a decade of AI hype and reality: funding and political attention have grown sharply, but adoption inside established companies remains slow outside fast movers like Google, Amazon and TikTok. He explains that not every AI approach needs large datasets, pointing to few-shot learning, deep reinforcement learning and knowledge graphs, and argues that common sense and causal reasoning, not just pattern matching, are the field's next real challenge. He then lays out a step-by-step roadmap: assess the company's identity and ambition, distinguish real autonomous AI from analytics or automation, start with meaningful iterative use cases, centralize standards and research while decentralizing customer work, and size internal AI teams to company scale. The conversation closes on automation versus augmentation, emotion AI and Guttmann's optimism about AI's role in healthcare and drug discovery over the next decade.
Key takeaways
- Christian Guttmann says AI integration into established corporations has been much slower than expected, outside fast movers like Google, Amazon and TikTok.
- Not all AI needs big data: deep reinforcement learning can improve from single feedback signals, and few-shot learning aims to generalize from one example.
- Guttmann argues common sense requires causal reasoning, citing Judea Pearl's point that machine learning fits curves but does not explain why events happen.
- For structuring AI teams, he recommends centralizing standards and research while decentralizing customer relationships and daily project work.
- Asked what matters more for successful AI adoption, Guttmann says clearly it is people and organizational acceptance, not the technology itself.
Detailed notes
Guest background and a decade of AI hype versus reality
Christian Guttmann is introduced as an "ecosystem builder": originally from Germany, raised in Australia, now based in Sweden, with a psychology background and a PhD in distributed multi-agent systems completed about 22 years earlier. He is a professor of AI in Australia, founder of the Nordic AI Institute, and Chief AI Officer at TietoEVRY, a roughly 24,000-person infrastructure and consulting company, while also advising and investing in AI startups. Asked how the field has changed over 10 years, he says the biggest surprise was how fast companies realized the value of pattern matching once large e-commerce firms in the US and Asia used it to personalize ads and feeds; before that, similar technology was often relabeled "intelligent systems" or "analytics" because "AI" sounded too esoteric. Westerheide adds that Asgard had four or five AI portfolio companies in Berlin in 2014 and over 200 by 2020, and that Angela Merkel once called the internet "Neuland" but now discusses AI.
What worked, what stalled: ethics, integration and hype cycles
Guttmann says he did not expect how much debate around bias, moral machines, robust AI and explainability would emerge, calling it a topic absent from his roadmap six years earlier and now a major field. What surprised him more, though, was how slowly AI has been integrated into established businesses: Google, Amazon and TikTok move fast, Siemens surprised him by moving well, but the rest of the corporate world, especially small and mid-sized companies, remains very slow. Both agree the field is not in an "AI winter" but past peak hype, calling it "autumn" and heading toward more realistic, proof-driven expectations rather than science-fiction comparisons. Guttmann also cautions against over-regulating speculative future risks, arguing that AI regulation, a reflection of a society's ethics and values, should stay innovation-friendly and be tested in sandboxes before being written into law.
AI without big data: few-shot learning, common sense and knowledge graphs
Guttmann pushes back on the idea that all AI needs massive datasets: deep reinforcement learning can improve from a single right-or-wrong feedback signal rather than millions of labeled examples, and one-shot or few-shot learning approaches aim to recognize a new object after seeing it only once, closer to how humans learn. He points to knowledge representation and knowledge graphs, used by companies like LinkedIn and Facebook to connect people and by Google Search, as an underrated area that retrieves and connects information faster than databases. On the field's next challenge, common sense and transfer knowledge, he cites Judea Pearl's work on causal and probabilistic networks: machine learning is good at fitting curves but does not explain why something happens, and building genuine common sense requires understanding causal relationships, not just correlations. He also describes an Asgard portfolio company that built swarm software controlling hundreds of drones with a single command, about ten years ahead of its market since industry and defense did not yet operate enough drones to need it.
Building a company's AI roadmap
Guttmann lays out a four-part approach for making a business "AI ready." First, assess identity and ambition: is the company product or service based, B2B or B2C, and what it wants to become in two to five years. Second, get clarity on what AI actually is, distinguishing autonomous decision-making systems, such as self-driving cars or automated ad placement, from dashboards, analytics or RPA that only display information without acting. Third, choose meaningful, iterative use cases that create visible value and win over skeptical stakeholders rather than trivial pilots or a single high-risk "big leap"; he references a Dutch company that agreed with its board and investors, years earlier, to accept lower dividends while pivoting its business around AI. Fourth, structure the organization: full centralization fails because a small central team lacks the knowledge held at the edges of the company, while full decentralization causes teams to reinvent the same solutions; Guttmann recommends centralizing standards and research, the approach he says Siemens and SAP also use with a central research lab plus local machine learning teams, while decentralizing customer relationships and daily project work.
Q&A highlights
Asked whether companies should build an internal AI consulting team, Guttmann says yes, especially to keep skills in-house, and gives a rough sizing rule: around 10 employees, buy AI capability through APIs; around 250, consider building some capability; at 5,000 or more, maintain an internal AI team. Asked about the relative importance of technology versus people for successful AI adoption, he answers clearly that it is the people, citing organizational acceptance, changing workflows and trust, such as a salesperson trusting daily call recommendations, as more decisive than the technology, and frames the whole shift as digitalization, of which AI is one term. On automation versus AI, he says automation and RPA execute simple, repetitive scripts that create value quickly but cannot learn or self-improve, while true AI systems have some awareness of their environment and can adapt from feedback; he prefers the word "augmentation" to "automation" for human-AI collaboration, since automation sounds like it simply removes jobs. On emotion AI, prompted by an audience question, he distinguishes needing emotions to be intelligent, wanting AI to have emotions, and being able to build them, and says recognizing emotion, for example to flag hate speech, differs from a system having emotions itself; he sees an underestimated market for emotionally aware systems such as virtual companions, more visible during the coronavirus period. Closing on the next decade, he expects growth in natural language processing, machine translation, virtual companions, healthcare robots and AI-assisted drug discovery, mentioning recent discussions on COVID-related drug research at the Royal Society of Medicine, and hopes AI will help address problems such as cancer treatment and transport safety.
Quotes
AI has always been around and has developed all the time.
Christian Guttmann
AI is not just machine learning.
Christian Guttmann
It's the people. The people are a major factor in this.
Christian Guttmann
Questions answered in this talk
Does every AI solution need a lot of training data?
According to Christian Guttmann at this Rise of AI Virtual Chat, no: deep reinforcement learning can improve from a single feedback signal and few-shot learning aims to recognize something after seeing it only once, so large datasets are mainly needed for pattern-matching approaches like deep neural networks.
Should a company centralize or decentralize its AI team?
Christian Guttmann recommends, in this Rise of AI Virtual Chat, centralizing shared standards, practices and research while decentralizing customer relationships and daily project work, since full centralization loses edge knowledge and full decentralization leads teams to reinvent the same solutions.
What matters more for successful AI adoption, technology or people?
Christian Guttmann says clearly, in this Rise of AI Virtual Chat, that people matter more than technology: organizational acceptance, changing workflows and trust in AI recommendations determine whether adoption succeeds.