At Infinum Slovenia’s latest Business Breakfast, executives and tech leaders gathered in Ljubljana for a candid, no-frills conversation about AI – where it’s genuinely delivering value, and where the gap between promise and production reality remains wide.
On May 6th, Infinum welcomed an invite-only group of business and tech leaders for a panel discussion on the practical realities of AI adoption. The format was deliberately unpolished: no vendor pitches, no curated case studies – just an honest read of where most organizations actually stand today.
The event was organized as part of Infinum Slovenia’s Business Breakfast initiative, a series created to give executives and decision-makers a more grounded, realistic perspective on digital transformation – one that goes beyond glossy presentations and sales promises.
“I strongly believe that today it is more important than ever to have honest and open conversations within the business community about AI and digital transformation,” said Tamara Lah Momčilović, Managing Director of Infinum Slovenia and the driving force behind the initiative.
The idea behind Business Breakfasts is to create a space where executives and decision-makers can gain a broader, more realistic perspective on what is actually happening in practice. Because the events are closed to the media, the atmosphere naturally becomes more open, authentic, and constructive, giving this format its greatest value.
TAMARA LAH MOMČILOVIĆ, MD, INFINUM SLOVENIA
The morning’s discussion brought together practitioners with hands-on experience shipping AI in production environments.
Željka Stiblik, Conversational AI Lead at Infobip, drew on over 18 years of experience in digital products and conversational AI to ground the conversation in what she sees working – and failing – with clients across industries, while Branimir Akmadža, AI & Data Engineering Team Director at Infinum, brought an engineering-first perspective that comes from actually building these systems.
Tamara Lah Momčilović, Managing Director of Infinum Slovenia for the past decade, framed the discussion around the decisions executives face before a single line of AI code is written. Guiding the conversation throughout was Željko Plesac, Partnerships Director at Infinum, whose questions kept the panel anchored to what practitioners in the room actually needed to hear.
The model is 20% of the solution
One of the panel’s most direct – and most quoted – moments came early, when Branimir reframed the entire AI conversation.
The model is maybe 20% of the solution. The other 80% is everything around it – your data pipeline, your integration layer, your evaluation framework, your feedback loops.
BRANIMIR AKMADŽA, AI & DATA ENGINEERING TEAM DIRECTOR, INFINUM
It set the tone for what followed: a systematic dismantling of the idea that AI is something you plug in.
The companies getting results, Branimir argued, are the ones treating AI as an engineering discipline – with architecture, testing, iteration, and clear KPIs. The ones getting burned are the ones who expected to connect an API and transform their business by Friday.
Why data is the real obstacle
One topic the panel kept returning to was data – or rather, the lack of it in a usable form.
The pattern Branimir described was immediately familiar to most in the room: a client arrives excited about AI, and within the first week of an infrastructure audit, the team discovers data siloed across five systems, inconsistently labeled, partially duplicated, and governed by no one.
“This isn’t a technology problem,” he said plainly. “It’s an organizational one.”
Željka reinforced the point from the conversational AI side: fragmentation is the norm, not the exception.
Data lives in spreadsheets, PDFs, legacy systems, and – critically – in the institutional knowledge of experienced employees. An AI agent without access to complete context cannot fully solve a problem, no matter how capable the underlying model.
ŽELJKA STIBLIK, CONVERSATIONAL AI LEAD, INFOBIP
She summed it up plainly: “Garbage in, garbage out.”
It’s a problem Infinum has been working to solve firsthand. Knowledge Hub, Infinum’s enterprise AI product, was built around one core premise: AI is only as good as the knowledge you give it.
From chatbot to agent: the personalization gap
Željka Stiblik brought a perspective shaped by 18 years at the intersection of digital products and customer communication – and she didn’t soften the picture.
She traced the arc of conversational AI from systems that could answer exactly 20 predefined questions, to today’s agents that understand context, intent, and tone – and can proactively initiate and execute actions. The progress is real. But so is the gap.
Technology has advanced faster than most organizations have been able to follow. We still see a significant distance between what’s technically possible and what’s actually implemented.
ŽELJKA STIBLIK, CONVERSATIONAL AI LEAD, INFOBIP
She was precise about where generic AI hits its ceiling. Off-the-shelf solutions can handle a meaningful share of automation – around 30% deflection rate on customer queries is achievable, and not insignificant. But reaching 70–80% deflection, and the real business impact that comes with it, requires a different category of solution entirely.
The difference is between a chatbot that knows how to respond and an agent that knows how to resolve.
ŽELJKA STIBLIK, CONVERSATIONAL AI LEAD, INFOBIP
True AI agents don’t just answer – they act. They check statuses, update records, book appointments, escalate issues, send messages. Getting there, Željka argued, demands more than a capable model. It demands clean, accessible data; well-defined guardrails; rigorous evaluation; and a platform built for production-level reliability, security, and regulatory compliance.
Which brought her to the point she returned to most: “A good demo is easy to make. Production is a different story.”
Almost any organization can build an impressive proof of concept in a week. What separates that from a system that delivers real business value is everything that comes after – data preparation, quality evaluation, continuous monitoring, and a partner or in-house team that genuinely understands your processes.
The technical bar for a demo has never been lower. The bar for production has never been higher.
Building AI you can actually trust
Hallucinations – instances where AI systems produce confident but incorrect outputs – came up repeatedly as one of the most damaging trust barriers in production environments.
Branimir outlined the multi-layer approach Infinum uses to address them.
The framework starts with RAG (Retrieval-Augmented Generation), anchoring model responses to an organization’s actual documents and data rather than general training. Every claim the system makes should point to a verifiable source. A separate model instance then runs continuous quality evaluation – tracking relevance, faithfulness, and response quality through tools like Langfuse. And human-in-the-loop feedback closes the loop, with real users flagging errors that feed back into the system.
Two layers of quality assurance: automated evaluation running continuously, and human oversight catching what automation misses.
The unglamorous work that actually matters
The thread connecting every discussion across the morning was consistent: most organizations aren’t failing at AI because of the technology.
They’re failing because of fragmented data, low adoption, and the distance between what a demo looks like and what production actually demands.
Branimir closed with a frame that stuck: “AI is not a product you buy – it’s a capability you build.” The companies that will win aren’t chasing the latest model release.
They’re investing in data foundations, defining clear success metrics before writing a single prompt, and treating AI as a team sport between engineers, domain experts, and end users.
The technology, he said, is ready. The question is whether organizations are ready to do the unglamorous work that makes it useful.
Planning your own AI roadmap and not sure where to start? Explore how we can work together.
