Image generated by Deeptech Times using ChatGPT
The chatbot era was easy to demo. A model answered a question, sounded human and promised a cheaper future for customer service. Then companies tried to put those bots to work.
That’s where things got ugly. Customer records live in one system, payments in another, marketing journeys somewhere else, and compliance rules are buried in PDFs or in people’s heads. A bot that can’t see across that mess hits a wall fast.
Infobip’s AgentOS is a bet that the next phase of enterprise AI is orchestration, not conversation. The Croatian communications company pitches the platform as a control layer for AI agents across marketing, sales and support, connecting customer data, workflow automation and channels like WhatsApp, SMS, RCS, email and voice, then handing off to a human when the machine hits its limits.
Companies aren’t short of pilots anymore. They’re short of production systems. BCG’s 2025 research found that 45 per cent of APAC firms are experimenting with or deploying agentic AI. Separately, Adobe has reported that 88 per cent of Asia-based respondents say fragmented data limits their ability to deliver responsive, personalised customer experiences. Getting from demo to deployment is where the real work starts.
Who does what, and when
An effective AI agent needs more than a LLM. It needs permissions, memory, reliable data retrieval, audit trails and a clean handoff to a human. Skip any of those and automation just delays the moment a customer gets annoyed.
Ruslana Reznikova, Infobip’s vice president and general manager for APAC and Eurasia, says moving beyond pilots means wiring customer data, channels, workflows, security and human support into a single operating layer. It sounds like standard enterprise software talk, but it points to the real challenge in deploying agentic AI: making the bot accountable for action, not just conversation.
Infobip’s headline example is LAQO, a digital insurer that built LAQO GPT, a GenAI assistant on WhatsApp, in partnership with the company. The assistant grounds its answers in policy documents, draws on approved FAQ responses, and routes conversations by intent from the first message. An orchestrator was later added to coordinate a knowledge agent alongside two specialist agents for travellers and contractors.
The results, according to Infobip, are strong: the chatbot handled 30 per cent of inbound requests, supported a 330 per cent jump in sessions and resolved 90 per cent of queries within three to five messages.
In the first six months of 2026, LAQO GPT reportedly resolved about 40 per cent of all customer interactions on its own, alongside a 4.7 customer satisfaction score and a 26 per cent rise in completed travel insurance purchases. These are Infobip’s figures, drawn from its own client relationship, and worth treating as a case study rather than an independent benchmark.
A second customer, digital finance platform PesoRedee, has seen 80 per cent of inbound requests resolved entirely within the chatbot flow, Infobip says.
Going after the agent stack
Infobip isn’t new to this fight. The privately held company, led by co-founders Silvio Kutić and Izabel Jelenić, built its way to a billion euros in revenue before taking outside capital, becoming Croatia’s first unicorn in 2020 after a US$200 million equity round. In 2025, it secured a US$520 million senior secured facility led by funds and accounts managed by BlackRock and Blue Owl.
That history matters because enterprise AI is turning into a scale game. Infobip already runs a communications network spanning more than 850 carriers and 43 data centres, with over 3,000 employees across 20 offices in major APAC cities. Its edge is orchestrating the messy communications layer where customer interactions actually happen.
AgentOS supports the Model Context Protocol, which connects AI models, third-party tools and external systems, according to Krešo Žmak, Infobip’s chief innovation officer. Its LLM library supports models including OpenAI’s GPT, Llama and Gemini, letting enterprises pick models by task instead of locking their whole stack to one vendor.
The more consequential layer sits around the model. The platform’s Knowledge Agent uses retrieval-augmented generation to answer from company documents, and Infobip says its Agentic Chunker, paired with a cross-encoder, lifts vector search recall from 70 to 95 per cent while cutting document preparation time from months to hours. Active Recall carries conversation history through a session. A customer data platform unifies messages, transactions, and behavioural signals into a single profile. Guardrails check agent output against business rules before it reaches a customer.
Compliance features round out the pitch: role-based access control, data isolation, single sign-on and API-level authentication. Infobip says it holds SOC 2 and ISO 27001 certifications, is GDPR-compliant and encrypts data with AES-256. Web-interface actions are logged for up to five years, with tracing at the agent layer so a company can reconstruct why an AI responded the way it did.
The boring layer may be the battleground
AgentOS is priced like infrastructure, not software. Customers pay a platform fee for agents, orchestration, analytics, customer profiles and human-agent seats, a processing fee for messages moving through the orchestration layer, AI credits for model-driven work, and separate channel fees for WhatsApp, SMS, RCS and other delivery infrastructure.
None of it has the appeal of a friendly AI assistant with a name and an avatar. But that’s exactly what enterprise buyers are paying for now. The questions that matter are whether an agent can access the right data, operate inside governance rules, escalate cleanly when it should, and prove it’s worth the cost.
Infobip won’t have the field to itself. CRM platforms, marketing automation tools, contact centre software and cloud providers all want to own the orchestration layer for AI-driven customer engagement. Infobip’s advantage is its communications footprint. Its challenge is convincing buyers that the infrastructure behind an agent is worth as much as the agent itself.
Since launching AgentOS on April 1, Infobip says nearly 100 customers globally have adopted it, more than 20 of them in APAC. Whether that number keeps climbing will say a lot about where enterprises think the real value in AI agents lies: not in how well a bot talks but in what it’s allowed to do.
