At Rise of AI 2026 in Berlin, on the Applied AI Stage, Christian Guttmann argues that the governance models most organizations use for AI were built for an older, predictive-AI paradigm and are now too slow for agentic systems that make decisions in real time. Drawing on work across telecom, finance, healthcare and software, he says the recurring failure modes he observes when AI adoption stalls have nothing to do with the technology itself: misaligned incentives, organizational silos, and threats to the identity of experts whose judgment AI could replace. He illustrates this with a hospital scheduling project in Australia that saved money but was blocked because the hospital's budget rules punished underspending, and a retail bank where product managers were paid to push individual products, undermining an AI built to recommend across the whole portfolio. His central proposal is to move governance out of monthly committee reviews and bolt-on policy documents and build it directly into the AI platform, through evaluation frameworks, gateways, agent identities and observability, so that velocity and trust grow together as systems move toward multi-agent and eventually embodied and reasoning-based AI.
Key takeaways
- Guttmann says the failure modes he sees in enterprise AI deployments are incentives, structure and identity, not the underlying technology.
- An AI project that would have saved a hospital money was blocked because its budget rules punished underspending, not because the AI failed.
- A bank had to change how product managers were paid before an AI that recommended across the whole portfolio could be adopted.
- He expects multi-agent systems using A2A protocols within 12 to 24 months, followed by embodied and causal/reasoning AI.
- He recommends leadership teams ask upfront who loses status, what gets rewarded that AI will break, and what to tell the board when it fails.
Detailed notes
Why current governance models are outdated
Guttmann says most AI governance in companies was built for a predictive-AI paradigm of periodic model reviews, which cannot keep pace with agentic systems that now make decisions, such as approving expenses, closing support tickets or placing trades, in the millions every day. Having worked across telecom (predictive churn, customer operations), finance (next-best-action for banking customers), healthcare (an IBM project in Australia optimizing hospital schedules) and software development, he says the mismatch between slow, monthly governance cadences and fast agent decision-making is consistent across industries.
Three failure modes that have nothing to do with the AI
He identifies three recurring reasons AI projects fail to be adopted, none related to the technology: incentives (KPIs still reward the pre-AI way of working), structure (optimizing individual silos instead of the whole business), and identity (professionals fear their judgment being replaced). In the Australian hospital case, an AI solution that would shorten waiting queues and save money was blocked because the hospital's budget structure would have punished it for underspending the following year. In a retail bank's next-best-action rollout, product managers paid on the volume of their own product (credit cards, loans) resisted an AI that recommended across the whole portfolio according to each customer's needs; the bank first had to change how the team was paid before the AI could work as intended.
From bolted-on to built-in governance
Guttmann describes most existing governance as 'bolted on': owned by risk and legal, expressed as policy documents engineers are meant to follow but often do not, reviewed in monthly committee meetings that slow deployment without changing system design. He advocates moving governance inside the platform itself, something he says his current company in Sweden, EQT, has recently released, so that it is owned by engineers and enforced through evaluations, gateways, and agent identities that track individual agent performance. He places companies into four categories along two axes of governance strength and shipping speed: most organizations today sit in the 'tactical' quadrant, shipping fast but reactive on governance, citing a McKinsey case of agents run loose causing a security incident, and, as a separate example he gives, an incident at another company where both the main and backup databases were deleted within about nine seconds. Only about 1% operate in the top-right quadrant where governance is part of the platform, enabling both speed and trust.
What is coming: multi-agent, embodied and causal AI
Looking 12 to 24 months ahead, Guttmann expects multi-agent systems in which individual agents, including ones belonging to different companies or vendors, communicate through MCP servers and A2A protocols and eventually negotiate autonomously toward new actions. Further out he expects embodied AI, physical robots already advancing in Asia and Japan, to require a different governance model, and causal or reasoning AI, since most current large language models only predict the next word and construct a reasoning explanation after the fact rather than reasoning as they act. He places quantum computing as a governance-changing factor more than ten years away.
Five questions for leadership rooms
Guttmann recommends leadership and CEO rooms ask five questions before deploying AI at scale: who in this room loses status if this works; what gets rewarded today that the AI is about to break; which silo did the AI just make invincible; where does the agent stop and the human start (with a concrete metric, such as a euro threshold, for when a human must be involved); and what will be told to the board the first time the system fails. He argues these questions should be asked up front, not after an incident, and that protecting experts' identity means giving them new responsibilities, such as judging agents and owning exception handling, rather than asking them to validate their own replacement.
Q&A highlights
Asked how to get honest answers in boardrooms about who loses status, Guttmann says concrete use cases, discussed openly and mapped to specific business-process changes, tend to produce honest responses; he describes a Swedish offsite where a team mapped AI's effect on different business lines and found the work would not disappear but shift, referencing another speaker's figure of 500% more work alongside more automation. Asked why Germany is slow on digitalization and AI adoption, he points to the smaller size and tighter networks of countries like Estonia and Sweden, a divergence he traces to the 1970s and 1980s when Germany, unlike the Nordic countries, resisted digitizing personal data, and to a cultural tendency to lead with risks and costs rather than upside. On a question about organizational versus technical controls when companies buy AI systems rather than build them, he says the goal is still to concentrate as much control as possible in a platform, referencing a point he says another speaker, Chris Boos, made about building a platform rather than a framework, and comparing this to how Google and Amazon operate their own platforms rather than making individual employees faster.
Quotes
In 10 years, no one will ask how good is your AI, but they will ask, can you still trust it.
Christian Guttmann
Who in this room loses status if this works?
Christian Guttmann
Most governance models that we are employing today are outdated.
Christian Guttmann
Questions answered in this talk
Why does AI adoption fail in large organizations even when the technology works?
According to Christian Guttmann at Rise of AI 2026, the recurring failure modes are misaligned incentives, organizational silos, and threats to the identity of experts, not problems with the AI or models themselves.
How should companies embed governance into AI systems?
According to Christian Guttmann at Rise of AI 2026, governance should be built into the platform itself, through evaluation frameworks, gateways, agent identities and observability, rather than handled through separate monthly committee reviews and policy documents.
What AI developments does Christian Guttmann expect in the next 12 to 24 months?
According to Christian Guttmann at Rise of AI 2026, he expects multi-agent systems in which agents from different companies communicate via MCP servers and A2A protocols, followed later by embodied AI (robots) and causal or reasoning-based AI.