Infobip’s CTO on Moving From Demo to Production
Most enterprises have already deployed AI tools. Far fewer have built the systems, workflows, and governance needed to run them as part of daily operations. That gap — between adoption and real operational scale — is the focus of a recent Q&A with Izabel Jelenić, Chief Technology Officer at Infobip, on what it actually takes to move agentic AI from pilot to production.
The Problem Isn’t Adoption — It’s Scale
Enterprise investment in agentic AI keeps growing because the underlying business pressure hasn’t gone away: organisations still need to serve customers faster, personalise engagement, manage rising interaction volumes, and cut operational complexity across marketing, sales, support, and contact centre functions.
But the results so far have been uneven, and Jelenić points to a specific reason: most early AI deployments were launched as isolated tools rather than as part of a connected operating model. A chatbot in customer service. A recommendation engine in marketing. Automation in a back-office process. Each can create value on its own — but without a connection to enterprise data, customer context, and downstream workflows, that value stays capped.
McKinsey’s latest global AI research found that 88% of organisations now use AI in at least one business function — yet most haven’t scaled it across the enterprise. The challenge today isn’t getting started. It’s getting AI woven into how the business actually runs.
Isolated Tools vs. a Connected Operating Model
- Isolated AI tools solve a narrow task well, but stop there — no shared context, no downstream action
- A connected operating model links agents to live customer data, business systems, and the workflows that follow a decision
- Without that connection, scaling a successful pilot to enterprise-wide use multiplies the same limitations rather than the value
Why Governance Has to Come Before Scale
Operationalising agentic AI isn’t only an engineering problem. Jelenić frames it as a question of governance and oversight as much as technology: who is accountable for an agent’s actions, how decisions are reviewed, and where a human needs to step back in.
This becomes especially important in customer-facing functions like support and engagement, where an agent’s mistake reaches a real customer in real time. Enterprises that skip governance design in the rush to deploy tend to discover the gap only after something has gone wrong in production — a far more expensive way to learn the lesson than building the safeguards up front.
Human Handovers Still Matter
A recurring theme in enterprise agentic AI rollouts is the handover point — the moment an AI agent recognises it has reached the edge of its competence and needs to bring a human into the loop. Designing this well is not a minor UX detail. It determines whether customers experience AI as a helpful first line of support or a frustrating barrier between them and a real solution.
Getting handovers right requires the same connected context described above: the human picking up the conversation needs the full history the AI agent has already gathered, not a customer forced to repeat themselves from scratch.
What Enterprise Leaders Should Take Away
- Audit existing AI deployments for isolation: Are your tools connected to shared customer data and workflows, or operating as standalone point solutions?
- Build governance before scaling: Define accountability and review processes for agent actions before expanding beyond a single pilot
- Design handovers deliberately: Make sure agents recognise their limits and pass full context to a human — not just the ticket
- Treat connection, not adoption, as the real benchmark: 88% of organisations already use AI somewhere. The differentiator now is whether it’s stitched into how the enterprise actually operates
| “If these tools were not connected to enterprise data, customer context, and downstream workflows, their impact was often limited.”— Izabel Jelenić, CTO, Infobip |
The Bottom Line
Agentic AI pilots are easy to launch and hard to scale — not because the models aren’t capable, but because most organisations haven’t built the connective tissue between AI, data, and workflow that production deployment demands. As Jelenić’s perspective makes clear, the next phase of enterprise AI isn’t about proving agents can work. It’s about engineering the governance, context, and handover design that let them work reliably, every day, at scale.
