1 min left

Field notes · Agentic AI in Practice

Beyond the Hype: Where Agentic AI Meets the Enterprise

Agentic AI pays off only when it touches real workflows, and the hardest part is not the technology but earning trust.

Fabian Teja Boegershausen, Country Director Malaysia & Brunei at AgodaMilan Dhingra, Chief Product and Technology Officer at TeleportFreddy Loo, Head of AI at Pos Malaysia BerhadTristan Eljed, Head of Data & AI Enablement at SEEKAna Anastasia Litovchenko, Head of AI Strategy & Enablement at Xsolla

With Fabian Teja Boegershausen, Milan Dhingra, Freddy Loo, Tristan Eljed, Ana Anastasia Litovchenko

From CDO Vision Kuala Lumpur 2026
From CDO Vision Kuala Lumpur 2026 · 17 Jul 2026

The short version

01

Value comes from real workflows, not shiny pilots

Teleport cut dangerous-goods false positives by up to 75%, and Pos Malaysia caught medical-claim anomalies within months. Both wins came from narrow problems with money attached

02

Trust is the real bottleneck

Xsolla found that users want to see and steer what an agent is doing, not just get a one-step answer. Pos Malaysia’s answer to who owns an agent’s mistake is explainability, traceability and shared oversight.

03

The skeleton is universal, the edge is vertical

Every deployment needs the same unglamorous basics: guardrails, governance, data foundations, human feedback and observability. Competitive advantage comes from going deep into your own industry’s data and keeping humans in charge of judgment.

Contents
Agentic AI as an operating model
And governance
  • That agentic AI earns its keep only when it touches real workflows — and that the hardest problems are trust, accountability and governance
Models
  • The models
That agentic AI earns its keep only when it touches real workflows — and that the hardest problems are trust, accountability and governance, not the models.

There is hardly a boardroom anywhere that does not discuss AI agents every week. Yet, as moderator Fabian Teja Boegershausen, Country Director for Malaysia and Brunei at Agoda, put it at the outset, companies differ sharply in what they think "agentic" means, how they build it, and where it has actually worked.

Fabian Teja Boegershausen, Country Director Malaysia & Brunei at Agoda
Moderator

Fabian Teja Boegershausen

Country Director Malaysia & Brunei, Agoda
Milan Dhingra, Chief Product and Technology Officer at Teleport
Panellist

Milan Dhingra

Chief Product and Technology Officer, Teleport
Freddy Loo, Head of AI at Pos Malaysia Berhad
Panellist

Freddy Loo

Head of AI, Pos Malaysia Berhad
Tristan Eljed, Head of Data & AI Enablement at SEEK
Panellist

Tristan Eljed

Head of Data & AI Enablement, SEEK

His panel brought together four practitioners from very different industries: Milan Dhingra, Chief Product and Technology Officer at Teleport; Freddy Loo, Head of AI at Pos Malaysia Berhad; Tristan Eljed, Head of Data & AI Enablement at SEEK; and Ana Anastasia Litovchenko, Head of AI Strategy & Enablement at Xsolla. What emerged was less a celebration of autonomy than a sober account of what it takes to move agents from slideware into production.

The tone was set early. Dhingra said that if agentic AI does not improve orchestration, visibility or intelligence in Teleport's operations, it is "just a toy." Loo relayed his boss's brief even more bluntly: find new money or cut costs, because shiny AI problems have no place on the books.

From single tasks to systems that decide

Eljed offered a working distinction. An AI agent, he said, handles a single task with a single outcome, such as helping a recruiter screen CVs. An agentic system strings decisions and workflows together: it plans, reasons and anticipates next steps based on feedback. That is where complexity, and the need for governance, comes in — though a human still makes the final call.

Partner with The CIO Room

Reach the leaders this piece is about

Four or five partners per room, alongside 50–100 CIOs and enterprise AI leaders. Tell us who you want to meet.

Teleport's first production case shows what that looks like on the ground. The company, Southeast Asia's largest e-commerce logistics specialist, ships parcels in the spare belly capacity of passenger aircraft across AirAsia and around 50 other airline partners. Every parcel must be screened against IATA rules for dangerous goods, because the cargo flies with passengers.

Conventional airport scanners are deliberately conservative, and Dhingra said 70 to 75% of their flags turn out to be false positives. Each one means an operator opens a bag, inspects the parcel by hand and repacks it — costing time, money and delivery deadlines.

Teleport built a retrieval-augmented system with a semantic layer grounded in IATA's dangerous-goods guidelines and its own internal feedback loops. It reads often misclassified marketplace product data, assigns the correct commodity code and flags genuine risks. The result, Dhingra said, is up to a 75% reduction in false positives, faster airport clearance and a better chance of meeting the delivery promises platforms make to shoppers.

Feedback loops, and the customer who wants to watch

Decisions are only as good as the feedback behind them, and Eljed argued every AI system needs two kinds. User feedback asks whether a recommendation or search result was actually useful. Operational feedback asks whether the system is high quality, fair and serving business outcomes.

Governance metric
  • That second loop, he said, becomes the governance metric
Product metric
  • Just a product metric
That second loop, he said, becomes the governance metric, not just a product metric.

Litovchenko brought the customer's view. Xsolla, which builds payment and commerce tools for game developers worldwide, has spent about two years replacing chatbots with customer-facing agents. Its starting hypothesis was that fewer steps — ideally one — would delight users.

The lesson was the opposite. People want to trust the agent, and to trust it they want to see and steer the flow: what the agent is doing, how long they will wait, whether their refund is actually being processed. The job, she said, is not only to predict the user's next step but to answer what is on their mind while the work happens, and then deliver the outcome.

When the agent gets it wrong

Boegershausen put the hard question to Loo: when an agent makes a mistake at an organisation as public as Pos Malaysia, who owns it? Loo, eight weeks into the job, admitted Pos is early in its agentic journey, and that his reflexive answer — the business — did not survive a conversation with his boss.

He reasoned by analogy to an employee. A bad decision usually means someone was given a task they were not trained for, authority they should not have had, or the wrong information. Trace those back and responsibility may sit with whoever granted the agent its rights, not with the business user. Pos's working answer is explainability and traceability of every decision, with oversight shared between the business, the AI team and IT.

An audience member from the telecoms industry raised a problem from the other direction. After posting a data analyst role on SEEK and receiving more than a thousand applications, his first three interviews were with candidates whose AI-tailored CVs matched the job perfectly but who were not suited to it.

Eljed said he has faced the same problem in his own hiring for months. Some candidates now send AI avatars to interviews that answer on their behalf. SEEK's stance, he said, is that AI should surface the best matches and cut repetitive work, but the final decision stays human — and face-to-face judgment still catches red flags machines miss.

A boring skeleton, a vertical edge

Is there a universal playbook for deploying agents? Dhingra said no — but there is a universal skeleton, and it is deliberately boring: guardrails, governance, data foundations, human feedback and observability across every deployed agent. The same discipline applied to RPA and blockchain before it.

The real value, he argued, comes from going deep into a vertical. For Teleport that means turning unstructured operational data and context into structured data and context. Once that is done, agent outputs become more real-world, more transformative and more trustworthy.

Loo described an agent to his 80-year-old mother as a highly intelligent four-year-old with a sponge-like brain and the autonomy to act — which is why governance matters. He also cautioned that not every problem is an agent problem; most have been solved somewhere already. Pos used an agent to catch medical-claim anomalies, such as one person visiting several clinics in a week with family members. Rules and RPA could have done it, but the agent cut build cost and time to market, delivering savings within a couple of months.

Asked what will be unrecognisable by 2030, Litovchenko pointed to a metric Xsolla's board is now discussing: decision-making throughput. Xsolla runs a company-wide AI foundation with vertical plug-ins for each business branch. As agents take on execution, she expects growth to hinge on how quickly and how well leaders decide on priorities, risk, people and reputation — the part that will remain human. Xsolla plans to start experimenting with ways to measure it next year.

Buying time, honestly

The sharpest moment came from the floor. One audience member asked Loo what Pos Malaysia's right to win is against J&T, Ninja Van and the in-house logistics arms of Shopee and Lazada — rivals built without legacy networks or regulatory baggage. Will AI close the gap, or just buy time?

Loo joked that the question might cost him his job, then answered candidly. Domestic parcels are cutthroat, Pos carries a legacy network and regulatory limits, and AI alone will not change that. His plan is to bank unglamorous wins — medical claims, fuel efficiency across a large fleet — that prove AI can return real money. That proof matters: no initiative gets past the board without it, at a time when most AI pilots never reach production.

Those wins, he said, are table chips. They fund the capabilities needed for the real prize, a new operating model: customer service rebuilt around AI, an AI-enabled sales organisation, and a leaner, more nimble Pos. "Whatever that we're doing right now really is to buy time," he conceded.

That candour captured the panel's broader message. Across logistics, recruitment, gaming and travel, the leaders on stage were not selling autonomy. They were describing disciplined work: picking problems with money attached, building feedback and governance in from day one, keeping humans in charge of judgment, and earning trust one workflow at a time.

Compare this with our AI conference for CIOs.

Agentic AI as an operating model
Share

Leaders in this piece

Meet the people behind the quotes

Fabian Teja Boegershausen, Country Director Malaysia & Brunei at Agoda

Fabian Teja Boegershausen

Country Director Malaysia & Brunei, Agoda

Joined Kuala Lumpur · 17 Jul 2026View profile
Milan Dhingra, Chief Product and Technology Officer at Teleport

Milan Dhingra

Chief Product and Technology Officer, Teleport

Joined Kuala Lumpur · 17 Jul 2026View profile
Freddy Loo, Head of AI at Pos Malaysia Berhad

Freddy Loo

Head of AI, Pos Malaysia Berhad

Joined Kuala Lumpur · 17 Jul 2026View profile
Tristan Eljed, Head of Data & AI Enablement at SEEK

Tristan Eljed

Head of Data & AI Enablement, SEEK

Joined Kuala Lumpur · 17 Jul 2026View profile

Keep reading

More from the rooms

All field notes
What It Will Take for Agentic AI to Earn Enterprise Trust
Agentic AI

What It Will Take for Agentic AI to Earn Enterprise Trust

6 min read

Beyond Generative AI to Agents: CDO Vision San Francisco
Panel recap · Agentic AI

Beyond Generative AI to Agents: CDO Vision San Francisco

7 min read

Agentic AI Is Easy to Deploy, Hard to Govern: Dallas 2026
Panel recap · Agentic AI

Agentic AI Is Easy to Deploy, Hard to Govern: Dallas 2026

5 min read

Past The CIO Room room in Singapore
Next rooms · 13 Nov 2026

Have this conversation in the room

The next rooms open in Singapore, New York, Sydney.

Partner with The CIO Room