Three senior media executives working across Asia-Pacific put the share of their work now handled by artificial intelligence agents at between 10 and 20 percent, and none of them scored their own organisation above six out of ten, in a panel discussion published on August 6, 2026.

The session, titled AI in Practice: What's Actually Working in APAC Media, ran 33 minutes and 44 seconds and was recorded at ATS Singapore 2026 before being posted to the ExchangeWire television channel. Independent technology journalist Eileen Yu moderated. The panellists were Eileen Ooi, APAC president of PHD, part of Omnicom; Ganga Chiravurri, president of product and solution development for APAC at dentsu; and Vincent Pang, managing director for APAC at Locala. The published transcript is machine generated, and speaker names appear inconsistently within it.

A scorecard nobody pushed above six

Yu opened by asking each panellist to grade how well or how badly AI was performing for their team, on a scale of one to ten. Ooi answered six, saying her group sits slightly ahead of the market while conceding that the market as a whole is still working things out. Chiravurri gave the same range, five or six, and questioned whether ten was the correct ceiling at all. Pang also said six, then added that the score which matters is the one agencies award his company rather than the one his company awards itself.

That is a low set of numbers from three people whose organisations have spent two years telling clients that AI changes everything. It is also consistent with what followed.

Yu framed the discussion with figures she attributed to Salesforce, drawn from a Singapore study of 100 respondents: half of Google searches now carry AI summaries that bypass brand websites, 84 percent of Singapore marketers say AI is reshaping search optimisation strategy, and 86 percent have begun optimising for AI generated responses on surfaces including ChatGPT and Google's AI Overview. The sample is small and the claims self-reported, which matters when the headline number is set against independent measurement.

That measurement points in the same direction at different magnitudes. A randomised experiment involving 1,065 desktop Chrome users found that AI Overviews cut outbound organic clicks by 39.8 percent, while the Datos and SparkToro desktop panel recorded United States zero-click searches at 22.4 percent in March 2026. The definitional gap between such datasets is itself a live dispute.

What the agents are actually doing

Ooi described PHD's position through Omni, the marketing orchestration platform Omnicom has built over an extended period. The holding company has held enterprise-level partnerships with the large language model providers for the last two years and is now moving into multimodel work, she said, with all of its agentic solutions running inside Omni. She listed a research console intended to surface insights beyond what she characterised as human bias, and an audience console covering synthetic audiences.

Omni's positioning has been reshaped by consolidation: Acxiom's RealID identity graph, inherited through Omnicom's $13.25 billion acquisition of IPG completed on November 26, 2025, now sits at the centre of the combined platform.

Asked what share of tasks agents have taken over, Ooi gave a figure of roughly 10 to 20 percent in overall productivity terms, concentrated in insights reporting. She attributed the ceiling to two factors: the strength of the data powering the agents, and the capability of the user operating them. Reporting automation predates the current cycle, she added, and has been accelerated rather than newly invented.

The ratio holds elsewhere. LG Ad Solutions disclosed in November 2025 that its media effectiveness agent reduced campaign report compilation from two days to five hours, while none of the agents it designed handled money.

On model selection, Ooi said the group runs the three largest language models, with a localised Chinese model used for the China market.

The definition problem

Chiravurri took a different line, arguing that the term itself has lost precision. The definition of AI, he said, is "fairly bastardized right now". He traced a continuum running from high-level statistics through data science to predictive modelling and, over roughly the past 18 months, generative work, noting that predictive systems are already embedded across the digital platforms media teams operate every day.

He was equally direct about the gap between marketing language and system architecture. Situating a process on top of a language model does not produce an agent, he said: "that's just an AI assisted tool." A genuine agent, in his framing, requires independence in the orchestration layer, the ability to hold embeddings, fine-tuning against proprietary documents and information, and the capacity to advance a workflow without step-by-step handholding. Summarising open web information, he said, does not qualify.

Between true agents and AI-assisted tools, Chiravurri put dentsu's coverage at a good 15 to 20 percent of the work performed.

That figure sits close to what independent reporting has found. Digiday reporting from November 12, 2025 placed agents largely in advisory and information-retrieval roles, advising rather than placing spend. Ooi made a version of the same point, describing as a myth the client assumption that agents will simply sort everything out. Agents orchestrate and recommend, she said, but "it changes the way we work but it doesn't solve everything". Chiravurri drew the parallel line between pilots and production: reaching something operable hands-free takes further work across everyone in the chain.

A grocery account, 1,800 stores, and a weekly cycle

Pang supplied the panel's most specific operational account. APAC contains roughly 15 to 20 key markets, he said, and consumer journeys, business potential, store networks and media consumption vary widely between Singapore, Malaysia, Australia, Thailand and Indonesia, and within each market as well.

His example concerned a grocery client of two years standing, operating 1,800 outlets across convenience stores, supermarkets and hypermarkets. The engagement began at 300 stores, with plans adapted to each store's local reach, business potential and competitor intensity. Built manually, that structure came to 300 lines of tactics across three creatives and three messaging variants, and orchestrating it by hand took three days and three nights. The same workflow now scales across 1,800 stores in a couple of hours, with each weekly campaign wave generating thousands of reports automatically.

The results he cited: cost per visit halved, seven dollars of revenue returned for every dollar of media investment, and more than 2.5 million incremental visits across the partnership. These are supplier-reported figures, presented without independent verification.

Pang's framing of where the value sits was deliberately unglamorous. The most valuable applications occur during or before the buying itself, he said, in understanding market differences and recommending local strategies. "It's not spectacular but it works," he said. With budgets under pressure, doing more does not mean "spreading the same plan everywhere".

He identified the failure point as execution rather than strategy. Planners on both brand and agency sides often have strong strategic planning, he said, but omnichannel setup across mobile, social, CTV and out of home flattens the nuance and creates operational friction. The objective, in his words, is preventing a plan from becoming "a beautiful strategy that disappears before it reaches execution".

The cheesecake that kept turning into a cat

Asked where AI had disappointed, Chiravurri, who said he has spent 25 to 27 years building technology products, offered a diagnostic anecdote rather than a complaint.

Roughly 18 months before the panel, his team was building a product with generative AI as a core element, two weeks from production. Asked to generate variations of a slice of cheesecake, the system returned three cheesecakes in the requested colours and one image of a cat, every time. Testers and developers worked through a Friday night and Saturday morning without resolving it.

The explanation arrived from an unrelated subreddit thread asking, in his recounting, "Why is it that people in the US always name their cats Cheesecake?" Most model training at that point had been conducted largely on United States data, text and images.

The point he drew from it was architectural. A base model taken at face value is not the unit of analysis; what matters is the proprietary data, training technique and process design layered on top of it. Failure modes have shifted rather than disappeared, he added, citing hands with too many fingers as a current example.

Ooi identified the corresponding organisational risk as "outsourcing thinking, taking the outputs for granted". A cat appearing in place of a cheesecake is easy to spot. Media data accepted at surface level is not, and she said the concern sharpens in optimisation agents handling millions of dollars.

Human review as a hard gate

On that basis, Ooi described human involvement as non-negotiable at PHD. Agents that help set up campaigns are not permitted to go live without a human vetting the work, and gaps are still caught, she said, because the agents themselves were built by humans.

The sell side has been building the same constraint into infrastructure. PubMatic, which launched AgenticOS on January 5, 2026, shipped guardrails on August 5, 2026 that block AI agents from transacting on unapproved inventory. The IAB Tech Lab formed a Programmatic Governance Council on April 21, 2026 with Omnicom, WPP, dentsu, Amazon Ads, The Trade Desk and Magnite among its members. Whether protocol fields or platform-level checks should carry enforcement has divided the industry since October 2025.

The cost question the panel could not close

The most consequential exchange concerned money. Yu raised a figure attributed to Uber, saying the company had exhausted its annual AI budget within four months. No source or reporting period was given on the panel for that claim.

Ooi's response was the sharpest assessment of the session. The industry is not talking about the cost problem enough, she said, and is overly focused on efficiency and headcount savings. Technology costs have historically been absorbed by the agency. "The cloud costs money," she said, as does storing the data that keeps models performing. Pricing therefore has to move away from paying for service and human cost: "There is a need to have a conversation around blended cost now." She placed the industry "at a tipping point", to be worked through jointly with clients.

That conversation is already under way in public, with three incompatible answers emerging. Dept declines to pass token costs to clients at all. S4 Capital's Monks builds tokens into tech-and-subscription pricing. Large holding companies fold the cost into broader commercial structures, and PMG deployed a tool across its company under a $50-a-day token cap per user. Omnicom chief executive John Wren told analysts on July 29, 2026 that the market has not yet seen what AI actually costs; the group reported third-party service costs of almost $2.9 billion in the first half of 2026, against $1.7 billion a year earlier. Digiday reported on August 3, 2026 that principal media has returned to the centre of holding company strategy, with AI services offered as the inducement.

Pang, speaking from the vendor side, noted that cost tends to be pushed down the chain regardless, and reframed the question as one of value exchange and alignment. Agreement is needed, he said, on who uses AI, what data powers it and where human validation occurs. Without that, "AI is going to just create another layer of complexity instead of removing one".

Chiravurri added a technical dimension. Calibration is poor on both sides, he said: "we're not very well calibrated in terms of our usage", nor are the model providers on inputs and outputs. Token dashboards exist and consumption is measured, but whether tokens should be priced as they currently are remains unresolved. His example was a one-sentence question about 500 years of history returning 60 pages of billed output. His forecast, offered personally rather than for dentsu, was that intermediary companies will emerge to compress queries before they reach the model: "I do expect a lot of efficiency companies to come in the middle."

Engineering work in that direction is already visible. A Dutch agency published an architecture diagram on August 2, 2026 arguing that collapsing 12 protocol calls into a single buyer agent holds token consumption down. Magnite chief executive Michael Barrett, interviewed on July 15, 2026, placed 2027 agentic advertising spend near $700 million. Gartner has forecast that more than 40 percent of agentic AI projects could be abandoned by 2027, with escalating costs among the reasons.

Process change and the governance gap

Asked what would have raised his score, Chiravurri pointed at process rather than technology. Major technological shifts affect processes hardest, he said, and organisations remain wedded to theirs. Working with one client last year, dentsu built a software system after standardising the client's process for both automation and AI-assisted work. According to Chiravurri, that reduced an almost 35-day process to 30 minutes, standardised across five different markets.

Ooi declined the hindsight framing entirely. Nobody predicted the speed of change, she said. What she identified as the missed opportunity was governance: clarity on legislation at organisational, industry and government level, and a method for qualifying which agents meet enterprise-grade standards rather than treating any agent built by any employee as equivalent.

That question of agent quality has commercial weight for dentsu specifically. The group reported a ¥327.6 billion net loss for fiscal 2025 driven by a ¥310.1 billion goodwill impairment on international operations, abandoned efforts to sell that international business, and installed Takeshi Sano as president and global CEO with effect from March 27, 2026. The restructuring programme targets ¥42 billion in annual cost savings, following 3,400 job cuts announced in August 2025representing 8 percent of the overseas workforce.

Pang closed on sequencing. Last year was AI assistance, he said; this year brought live testing of AI agents; next year will be past that. The strongest teams, he argued, will apply AI to work that was previously too complex, too time consuming or too manual, rather than to work that was already fast: "how can we do more with AI and not just do things faster".

Why this matters for media buyers

For programmatic buyers and marketing leads, the panel is useful precisely because it is unflattering. Three executives with commercial incentive to describe agentic AI as transformative instead reported coverage of 10 to 20 percent, concentrated in reporting and pre-buy analysis, with human sign-off retained before any campaign goes live.

That figure is a usable benchmark. Mediaocean's H1 2026 outlook, which surveyed 320 marketing professionals in November 2025, found 54 percent planning to increase investment in AI media while the same respondents struggled to implement generative tools internally. Research covered in July 2026 recorded 87 percent of marketers believing their organisation uses generative AI effectively, against a MiQ survey from November 2025 in which only 45 percent felt confident. A panel of practitioners marking itself at six out of ten is a rarer data point than a commissioned survey at 87 percent.

The unresolved item is commercial. Nobody on the panel could say who pays for inference, on what basis, or how it is priced into a media fee. That is the same gap visible in holding company disclosures, in agency token caps and in the return of principal media as a margin mechanism. Until it closes, the productivity figures quoted at conferences describe only one side of the ledger.

Timeline

Summary

Who: Eileen Ooi, APAC president of PHD, part of Omnicom; Ganga Chiravurri, president of product and solution development for APAC at dentsu; and Vincent Pang, managing director for APAC at Locala, in a panel moderated by independent technology journalist Eileen Yu.

What: A 33-minute panel discussion in which the three executives rated their own organisations' AI performance at five or six out of ten, placed agent and AI-assisted tool coverage at 10 to 20 percent of workflows, described human sign-off before campaign launch as non-negotiable, and stated that the industry has not resolved who pays for model inference or how it should be priced into client fees.

When: Recorded at ATS Singapore 2026 and published on August 6, 2026.

Where: Singapore, covering agency and vendor operations across approximately 15 to 20 Asia-Pacific markets including Singapore, Malaysia, Australia, Thailand and Indonesia, with a separate localised model deployment referenced for China.

Why: Agency and vendor self-assessments of AI maturity have generally been published through vendor-commissioned surveys reporting high effectiveness. This panel produced practitioner figures substantially below those benchmarks, alongside an explicit statement from a holding company executive that technology costs previously absorbed by agencies now require a different commercial arrangement with clients.