From the INMA World Congress: four publishers' agentic AI playbooks
Inside four publishers running agents in production, and why most are still talking about it.
INMA World Congress • Agentic AI Seminar • Hotel Adlon Kempinski, Berlin • 8 May 2026
Michael Smith, CTO of Darwin CX, opened his session with a line he had borrowed from Brian Alvey, CTO of WordPress VIP: publishers, Alvey had told him, are focused on “doing just enough AI to stay behind.” It is what Smith hears from publishers around the world.
It was a useful provocation to open the seminar at the Hotel Adlon, part of the INMA World Congress in Berlin. The four speakers who followed session chair Jodie Hopperton, INMA’s Product and Tech Initiative Lead, represented a spectrum from careful strategic positioning to radical operational deployment. What connected them was a lack of patience for abstraction. The technology is no longer theoretical. The questions are organisational.
Holtzbrinck: getting into the room
The first question for any publisher serious about AI is where to learn from. Holtzbrinck’s answer was to open an office in San Francisco. Katharina Neubert and Claire Furino presented Holtzbrinck’s AI hub in San Francisco, now nine months old, staffed by two people and funded at less than 1% of group revenues. The point Neubert was explicit about: European publishers typically encounter AI companies through their sales functions. The people who build the products and the models are based in the Bay Area, and they are not the same people.
The hub operates across three layers:
Signal-gathering and synthesis for group leadership;
Knowledge transfer through an internal residency programme;
Direct investment, with six companies backed since launch. The residency is deliberately non-technical.
Any Holtzbrinck employee can apply for a 3-month residency in San Francisco to develop a project from idea to MVP, typically in 11 weeks. Twenty-four to thirty projects have gone through the programme, ranging from a neuro-symbolic control layer sitting between MCP outputs and model results, to a book club app designed to surface audience signals that inform commissioning decisions. The latter was built by a staff member who, by her own account, had not worked with AI before joining the residency. Furino, who leads the technical side of the hub, described building the entire management infrastructure (application tracking, C-suite reporting across all seven business units, and an agentic layer for processing inbound requests) in 30 days with two people.
“Within 30 days, Katharina and I, with the help of artificial intelligence, have been able to build infrastructure that five years ago would have required a whole team of software engineers, if not a vendor contract for three or four months.”
Claire Furino, Director of AI Hub Programmes
Key takeaways from the Holtzbrinck session:
Physical presence in San Francisco yields access to the people building AI products, not just selling them.
A residency open to non-technical staff accelerates cultural change alongside technical capability; it also changes how individuals think about what is possible.
Shared tracking infrastructure converts individual project outcomes into institutional knowledge across the whole group.
Investment positions the hub to generate returns that offset its own cost.
Darwin CX: the operational case
Smith broke AI adoption into three waves and argued publishers needed to read each one differently. The first wave (generative AI for content) was largely paused for reputational risk. The second (transcription, source detection, and editorial review tools) is already in use across major publishers, whether it has been officially sanctioned. The third, agents running business operations, is where the commercial case is clearest, and where most publishers have barely started.
His demonstration centred on marketing workflow automation. The traditional model requires a marketer to raise tickets with the engineering, data, legal, and channel teams before a campaign deploys. An agentic workflow replaces this with a natural-language prompt: describe the campaign goal, and specialist agents execute in parallel across segmentation, compliance checking, channel selection, and deployment. The marketer reviews a summary screen and approves. Darwin CX reports the following from a client deployment it described as live and in production:
• 100% reduction in cross-system navigation.
• 97.5% fewer clicks.
• 14–15x faster campaign deployment.
• Marketer’s time split shifted from 8 hours strategy / 32 hours execution to 30 hours strategy / 10 hours execution per week.
“The masthead is your product. Everything else is plumbing. It’s necessary. But on the business side, AI agents can run and do work for you.”
Michael Smith, CTO of Darwin CX
The implication for publishers is that the right place to start is not editorial. Segmentation, paywall configuration, campaign deployment, lifecycle marketing: these are operational workflows where agents can move now, and the risk to editorial credibility is zero.
Hindustan Times: What full deployment looks like
Before the case study, Verma laid out why standing still is no longer an option. Audience trust is deteriorating: 40% of global readers now avoid news sometimes or often, up from 9% a few years ago. Print and digital revenues are declining across the board. And AI is redistributing traffic in ways publishers have limited ability to influence. When Google rolled out changes to Discover in September 2024, Hindustan Times was among those hit hard. Publishers dependent on the platform are reporting referral losses of between 20% and 70%.
The logic follows directly. If publishers are losing traffic to AI-driven search, the answer cannot be to run more pilots. The speed at which AI can compound internal capability is the only lever publishers directly control.
Hindustan Times Digital has moved past experimentation and is measuring outcomes:
97.4% of code is now generated by agents.
Monthly code output has risen from 300,000 lines to 1.7 million with the same or smaller developer headcount.
Engineering delivery timelines have improved by 60–70%.
Sales cycle length has been cut by 40% through AI-generated pitch decks, produced in hours rather than five to seven days.
Repeat client rates in advertising have improved by 20% through automated post-campaign insight reports sent directly to clients.
12 AI tools are currently in production; 10 more from a recent hackathon are being deployed this month.
The organisational model behind these numbers is as instructive as the technology itself. Hindustan Times identified AI evangelists in every department: not technical staff, but functional domain experts who act as translators between capability and application. AI adoption is now linked directly to performance appraisals: staff who are not using AI will not meet expectations. A public leaderboard gamifies adoption, with the most active users recognised at monthly all-hands meetings.
“The question is not if we adopt it. It’s how fast we can execute.”
The governance structure is deliberate: a human reviewer sits at each stage of the agent pipeline, and quality-assessment agents audit the output of other agents. The goal is not to remove humans from the loop but to restructure where they sit in it.
Ippen Digital: the architecture question
Markus Franz, CTO of Ippen Digital and founder of its incubator lab, made the case for a different mental model. Stop thinking about individual AI tools. Agents are the operating system; skills (reusable, composable capability modules) are the applications built on top.
Ippen has built a suite of agents covering content research, SEO, audience analytics, and editorial quality assurance, orchestrated by a conductor agent that manages task allocation among specialist sub-agents. A content research workflow that previously took a journalist two to three hours now takes five minutes, scanning 300 articles and surfacing relevant material with full sourcing transparency.
Key practical observations from Franz:
Agents should be treated like junior employees: they require onboarding, context, defined personas and explicit guidance on tone, priorities and editorial standards.
Multi-agent systems require each agent to carry an “agent card,” a defined identity that enables agent-to-agent communication across departments.
Skills (reusable capability modules) are more durable investments than one-off agent builds.
Interface design is the current bottleneck: most agent outputs are still delivered as walls of text that editors cannot work with efficiently.
The adoption curve follows a predictable pattern: exploration, then specialised use cases, then genuinely novel workflows that emerge from the people using the system day to day.
Where the gap sits
The distance between the most and least advanced publishers in the room is not a technology gap. The tools are accessible, getting cheaper, and vendors are moving fast. What separates Hindustan Times’ 97.4% AI-generated code output from a publisher still running manual marketing workflows is not access to better software. It is the organisational decision to treat adoption as non-negotiable, and to build the governance, the incentives, and the measurement frameworks that make it stick. None is waiting to see how it plays out.