Beyond the Tornado: As Agentic AI enters its Cambrian age, focus shifts to creativity; pricing, forecasting and control; governance wins, bill shock bites; autonomy meets ambiguity
Agentic explorers: Omnicom Media Group's Alex Pacey and T2 Tea's Peter Randeria
Australia's marketing leaders are moving agentic AI from experiments into business operations. But the executives deploying the technology say the biggest challenges have moved well beyond getting autonomous systems to work. Instead, organisations are grappling with governance, cost control, agentic drift, workflow redesign and the practical realities of managing agents at scale. Still, they stress it's worth the conquest. Speaking on Mi3's latest podcast, T2 marketing director Peter Randeria and Omnicom Media Group chief product officer, Alex Pacey described a market that has matured rapidly over the past six months. Agents are increasingly being used to support forecasting, pricing, workflow management, customer engagement and decision making, with some organisations now treating AI as a core operating capability rather than a technology experiment. The shift is producing an unexpected lesson. While governance was widely viewed as a potential brake on innovation a year ago, both executives argue the organisations moving fastest today are often the ones that invested earliest in governance frameworks, employee education and data foundations. As deployment accelerates, the conversation is shifting toward the harder questions of trust, accountability, economics and organisational change.
What you need to know:
- Agentic AI has shifted from quiet experimentation to mainstream enterprise deployment, with marketers now focused less on proving value and more on governance, oversight and cost control.
- On Mi3's latest podcast, T2 Tea's Peter Randeria reveals AI use is now embedded in daily operations, but scaling it safely requires strict controls, approved connectors and active human supervision after an agent accidentally deleted finance data during testing.
- Omnicom Media Group's Alex Pacey argues governance is accelerating, not hindering, adoption, with early discipline around compliance and security enabling faster deployment at scale.
- The industry's vision of vast autonomous multi-agent ecosystems remains largely aspirational, with leading organisations instead centralising orchestration to avoid complexity, drift and operational chaos.
- Agent drift remains a major concern as systems can confidently pursue incorrect outcomes, misinterpret instructions or take unintended actions without seeking approval.
- Businesses are increasingly applying similar disciplines to AI agents as they do to employees, requiring onboarding, training, monitoring and performance management rather than set-and-forget automation.
- Token-maxxing was dumb. Cost management is emerging as the next battleground, with firms grappling with token consumption, AI budgeting and growing concerns over escalating usage costs as subsidies fade.
- T2 has moved AI spending into a dedicated P&L line and imposed token limits on users to improve visibility and prevent runaway costs.
- Despite the challenges, agentic AI is delivering measurable gains, with T2 citing stronger pricing and promotional performance while Omnicom credits AI with sparking a new wave of creativity and problem-solving.
- The next phase of competitive advantage will belong to organisations that can govern, supervise, integrate and economically manage AI at scale rather than simply deploy it.
We've gone from it being the dirty little secret that everyone was using to now becoming part of the rhythm of the organisation.
A few brave pioneers aside, a year ago, the main agentic focus of many marketing organisations, assuming they were doing anything at all, was simply about building the confidence to even experiment with new workflows built on famously hallucinatory technologies.
Now, with agentic solutions moving firmly into the mainstream, the challenge is governing them, supervising them, managing drift and chaos, and paying for them.
Mi3 dived deeply, and early, into agentic AI, kicking off with a research report based on interviews with more than 60 industry leaders around the world and published as Inside the Tornado. Now, a year later, it has wrapped up that coverage with a podcast featuring Peter Randeria, marketing director at T2 Tea, and Alex Pacey, chief product officer at Omnicom Media Group.
Their message for marketers was clear: The hard part is no longer proving agentic AI works. The hard part is making it work safely, consistently and economically at scale.
Twelve months ago, according to Randeria, AI was often used quietly and unofficially inside organisations. "The use of AI, certainly the use of agentic, was sort of a dirty little secret," he said. "You sort of kept it hidden, you kept it behind closed doors and you didn't sort of talk about it."
Today, he says, "In our organisation, it’s been incredibly normalised now. It’s become part of the rhythm of the way that we do business, certainly by the people who are using it. … But the fact is that we’ve embraced it and are using it not just a sort of day-to-day basis, but minute-by-minute basis oftentimes."
That shift from experimentation to operational deployment is forcing executives to confront a new reality. The challenge is no longer access to technology. It is governance. For years, governance has often been viewed as the bureaucratic counterweight to innovation. Yet both executives argued the opposite is proving true.
Pacey said the tension between moving quickly and maintaining compliance, security and governance had been a defining feature of Omnicom's AI journey. While some employees found the early controls frustrating, those constraints ultimately accelerated deployment later.
There’s very much tension within any business, but certainly one of our size and type, particularly as it pertains to AI, he suggested. "Wanting to be as cutting edge as possible on the one hand, but on the other hand, the tension of the requirement for governance, compliance, data security, and a sustainable model built over time."
For Pacey, effectively, the training wheels have been let off. "I think ultimately what I’ve seen in our business is that, whereas for some initially, perhaps those who are more impatient, that could be incredibly frustrating, it is in fact that early discipline that has freed a speed and a nimbleness later on that has seen us move at the speed you’re talking about.
"If I had to talk about one thing that’s enabled that, for us the number one enablement has been legacy... What that means is we’ve gone through expert systems, machine learning, narrowband AI, generative AI, and now agentic AI, all building and iterating within the same platform. That also means that the infrastructure and the enterprise memory exist within our business. Above everything else, that has been a key enabler."
Don't mention the scaffolding
When Mi3 began its research with a month-long road trip interviewing founders, heads of AI, product VPs and tech CMOs in the States, the vision the industry wanted to paint was on of millions and even billions of agents working autonomously – and yet in unison – for the better of brands and customers alike.
It's a common refrain still. And it's nonsense, still. The scaffolding didn't exist a year ago, and while there have been improvements over the year, particularly around orchestration, the peak remains elusive. So does the base camp, you could argue.
There were some truth tellers of course. Boomi's Michael Bachman warned attendees to Boomi World in May last year, "If you don't have a hub, and governance and control in the middle, that is not going to be spaghetti on your plate, that is going to be a spaghetti factory that you've never witnessed the likes of before."
Cognizant's AI Chief Babak Hodjat told us in August he had already dined at the bistro. His firm, a huge global consultancy with over 350,000 provided a cautionary tale. Departments eager to deploy generative‑AI tools each chose their own: sales brought in AgentForce, IT relied on ServiceNow, HR tinkered with custom bots. The upshot: digital bedlam.
“People can’t be walking around with a Rolodex in their hands going, 'okay, now I’m gonna go talk to that agent',” Hodjat told Mi3. The firm’s AI lab built a remedy, an 'intranet of agents' called Neuro San, designed to corral disparate bots, govern their behaviour and make them discoverable.
In the world of real things, where executives like Pacey have to deliver actual products, it helps to stay anchored to operational reality.
According to Pacey, "Without sounding flippant, the way we avoid having unmanaged multi-agent systems is by not building them ... I think about this in terms of two speeds. If I take orchestration within our platform, as a first speed or the first parallel, the way we overcome that is by having a single agent that orchestrates and coordinates all others. So at the moment we work in an infrastructure whereby the agents that people will use on a daily basis are built centrally. Okay, so they all have clear roles, they all have clear boundaries and they all impact a certain console of our platform, which speaks to a certain area of our business."
Next, he said, "Over the top of that we have an AI agent coordinator, which enables a user to effectively use natural language to ask and instruct that single agent that will then interact with the other agents that will interact with the capabilities, be they machine learning or AI capabilities. So that means there’s no person within our business directly instructing or constructing an agent that could go for one of the better term, haywire or point in a different direction to another agent.
"That is great from a governance perspective. It’s fantastic. It’s what clients want to hear from a security and compliance perspective, but it does – or can be seen – to shackle people from an experimentation perspective. So what we also have within our platform is an area where people can go and build their own agents or even let’s be honest, things they call agents that are actually chatbots, but do that in a safe environment before they’re let off the training field into the wild.”
We are now in the business of managing intelligence, full stop, whether it’s human or whether it’s machine. And that these agents and these large language models are enablement capabilities. They’re not replacement capabilities.
Get my drift
One of the recurring themes throughout Mi3's year-long research project has been agent drift. Systems that perform well initially can evolve in unexpected directions over time. Others complete tasks in ways users never anticipated. When a group of US Northeastern Universities teams with their colleagues on West Coast to create a power-packed Red Team to unleash on agentic systems, the result was a study they called Agent of Chaos.
Think of it as the the Dunning-Kruger effect, but for bots. AI agents are not just error-prone, they are also prone to being decisively wrong and overconfident in situations they do not understand. Worse, if you unleash agents into an ambiguous environment, agentic systems can misinterpret authority, leak sensitive data and trigger system-level failures, often declaring success even as underlying outcomes unravel.
While the severity varied, all of the agentic early adopters we spoke with for that story had experienced some form of agentic wobble.
Lendi's chief product officer Travis Tyler told us at the time, agentic systems become, "materially more useful once they have tools and workflow access,” but he also warned, “that is also the point where the risk profile changes – the failure mode shifts from ‘bad answer’ to ‘confident action in the wrong direction’.”
Per Tyler, one of the biggest operational constraints is context quality. “We’ve seen agents struggle when business rules, current-state behaviour or ‘what must not change’ are implicit rather than explicit,”
He added that hallucination and incomplete synthesis remain persistent problems, particularly in “document-heavy or judgement-adjacent use cases,” and said Lendi had treated those risks explicitly in products such as Contract Analyser and customer-facing systems.
Likewise, Randeria has seen the limitations up close. He described one incident involving connectors into T2's finance systems. "We started deploying and exploring connectors into some of our key enterprise systems, particularly our NetSuite finance system, probably a couple of months ago. One of our senior team members prompted it and got it started doing a search for an answer on a certain thing. What it ended up doing was going deep into our finance system and then basically removing data from our finance system pretty quickly. There was a whole layer of trading data we lost in short order.
"Now, it was not irrevocable. We managed to understand what had happened. Luckily, it wasn't days of work or anything, and we could retrieve it."
His point though, is that T2 learned quickly the need to take much more careful control. "One thing we've said, even to that pilot group, is that there is absolutely no use of connectors unless they are pre-approved. So all of us are absolutely paranoid. In that particular instance, we managed to catch it just in the nick of time. But yes, it gave us the heebie-jeebies, I can tell you."
More broadly, Randeria said organisations are discovering that agents pursue objectives with remarkable determination.
"What surprises me a lot is just how far agents go to complete a task," he said. "They don't ask for permission. They're just hell for leather doing it."
To complicate matters, sometimes firms encounter the opposite of drift - systems that do exactly what they are told to do, but without the benefit of human insight.
Take Lendi, which has already loosed one of the most advanced agentic enterprise systems up Australia's home loan market as an example. Unlike many Australian firms that are terrified of ever acknowledging any failure, Lendi's exec team is commendably transparent.
As we reported in January this year, Lendi’s first attempt at Guardian, its flagship agentic home-loan experience, began life as a kind of super-agent. Tyler told attendees at a Lendi partner and customer event: “We first started off with one agent, and that agent had access to 270 tools, had to traverse all these different disciplines, had to sell, had to be compliant, and had to do all these different things. And it was too much.”
In this instance, the issue was not so much that the AI refused to do what it was told, but rather that it did exactly what it was asked to do. (Once again, imagine you are instructing an intern on a mission-critical system.)
Both of these behaviours are forcing companies to rethink management itself. At the risk of anthropomorphising, companies are starting to understand that agents are less like the set-and-forget software packages of the past and more like employees who occasionally go rogue.
Canibuild is an example here. The Australian construction tech firm has overhauled its sales and marketing funnel with AI agents that instantly respond to inbound leads, replacing hours- or days-long delays, driving strong performance gains: Demonstration conversion rates are up 68 per cent, while revenue per employee has increased more than 400 per cent.
But Mark Deacon, the firm's chief commercial officer, said Canibuild treats AI systems much like new hires. Agents are trained on the company’s products, customers and processes, then monitored and refined over time.
“In the same way, if you hire a new person for your company, you don’t just sit them at their desk and go good luck,” Deacon said. “You train them on your business, your customer, your product. It’s the same with an AI agent.”
That training process can take one to three months, followed by ongoing optimisation. The company has assigned a dedicated employee to maintain its internal knowledge base, which acts as the primary source of truth for its AI systems. When performance dips, it is often a signal that the knowledge base is out of date.
“Whenever we see that percentage start to dip, we know it means that the knowledge hub hasn’t been kept up to date,” Deacon told Mi3 at the time.
Both Randeria and Pacey suggested organisations are entering an era in which leaders will increasingly manage mixed teams of humans and machine intelligence.
Pacey has begun framing the challenge internally as managing intelligence regardless of whether that intelligence is human or machine.
"We are now in the business of managing intelligence, full stop, whether it’s human or whether it’s machine. And that these agents and these large language models are enablement capabilities. They’re not replacement capabilities," he said.
"Look, commercial pressures notwithstanding, and perhaps I’m too pure from that perspective ... but the way we position these are, these are to augment your intelligence. They are enablement for you to be better. What I’m talking to clients about time and time again is what these agents do is get you to a much more elevated start point so that when the human comes in with the time that they have left, they are able to take this to a whole new level."
As to what that means in terms of supervision and evaluation, Pacey told Mi3, "I want people to consider that effectively this is just a new intelligence in their group they have to consider the same way they would a human. The implication is significant. Rather than replacing employees, agentic systems are increasingly being viewed as tools that elevate starting points while leaving humans responsible for judgement, supervision and critical thinking."
Randeria shared the view about the human oversight viewing agents as powerful but potentially risky tools that require active management, monitoring, and critical human judgement.
"You have to work out how much human effort is required to continue to monitor and sense-check everything that's coming through, in order to make sure it's delivering what you need it to do. So I think that's the challenge. It isn't a panacea where you can just let it go and have a coffee for the next three hours. It's work for humans, but work in a different way."
He said, "It's really interesting in terms of what it requires now of leaders and managers to think critically about what's coming back and ask: Is this right? Is this accurate? Is this on brand? Where has this come from? How did it achieve these results?
"And I think that's where the humanity will remain an important aspect of what we're doing here. I think those of us who get terrified of the dystopian future forget that actually, at the heart of this, it requires critical thinking and a level of assessment that the AI on its own cannot do."
It's such a live conversation we're grappling with in the moment. What I'd say is I don't think certainly our marketers are clear on the economics of where we're heading from an AI perspective, the cost per piece of content, cost per image, the actual physical cost that things [and they will] start to stack up.
Cost as gravity
If governance is emerging as one challenge, economics is rapidly becoming another. The industry's early experimentation phase was subsidised by investors in the LLM firms. Those subsidies then extended out to the bond markets as the hyperscalers and the model giants started borrowing money to fund the architecture underpinnings AI.
As this was happening, firms who should have known better were indulging, and in some cases encouraging, the insanity of token maxxing where developers and project managers bragged about the money they were burning on LLMs.
It lasted a few months, tops. Then Uber announced it has burned through a year's worth of tokens in four months, and Microsoft started culling access to tools like Anthropic's Claude as costs careened out of control, and firms started paying attention to the expense line in the spreadsheet, instead of the blue sky above the top of the hickey stuck in the PowerPoint pitch deck.
According to Digital Economy, a Stanford University publication researchers Erik Brynjolfsson, Sandy Pentland, and Jiaxin Pei in a paper called How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks found agentic tasks are “uniquely expensive, consuming 1000x more tokens than code reasoning and code chat.”
Cue economic gravity. In recent months, businesses have started paying closer attention to token consumption and operating costs. And the original assumption that the tech sector would deploy its decades-long superpower of accelerating efficiency to lower cost is starting to fray.
One uncomfortable result, Pacey noted, is bill shock. "Jevons paradox, which is something that's getting thrown around quite a lot in our business at the moment ... is the idea that improvements in resource efficiency actually lead to an increase rather than a decrease in total resource consumption. And that's what we're beginning to see now.
"We're seeing the end of cheap AI because, at the end of the day, the big four have got to at least start to be seen to cover costs, never mind make any actual money. And it is quite interesting to think that the AI that we have now, later on, could be dictated by who makes money and who doesn't. We could end up with one or two rather than four. When you're putting all your bets in one area, that becomes quite challenging."
According to Pacey, "I think we are certainly seeing that across our business, and there is a challenge here for me from a visibility perspective because so much of this is still centralised. We're still very much in a place where I don't think, as an industry ... we're clear on what this means for charging and commercial models."
For marketers generally, the implications may be familiar. Just as digital media spending eventually required formal governance, budgeting and accountability, AI usage is heading in the same direction.
It's an issue brands are already dealing with. Vera Skocic, Techtronic Industries (TTI) head of customer and strategy told Mi3 earlier this month: "I was looking at my SaaS budget, and it's going to look quite different over the next year. We were asking: What do we forecast for next year? We probably won't be working with so-and-so because we cut back here or increase there, right? Then it was like, 'Hold on, where are the tokens?' How does that have to be accounted for? The business hasn't landed on that yet."
AT T2, Randeria said, "It's such a live conversation we're grappling with in the moment. What I'd say is I don't think our marketers are clear on the economics of where we're heading from an AI perspective, the cost per piece of content, cost per image, the actual physical cost that things [and they will] start to stack up.
"We're just trying to navigate that now. We've made some critical decisions we think will give more transparency to this, but I don't think we have all of the answers. We've moved the AI cost centre out of the IT line in the P&L into its own standalone P&L line to give it transparency. That was an important decision we made recently. And it's helped to draw some real focus on how we're managing that."
The issue of cost goes back to the point about control and the very tight group of pilot users with access. "They have very clear token limits on how many queries they can make. And that's all the way up to the end of the business," said Randeria. "You get to a certain number, you just can't do any more. And we're clear on that because we've seen these horror stories of organisations deploying Claude out to the entirety of the employee base and having to get a bill of half a billion dollars with all of the usage that was incurred.” *
It's the rebirth of creativity and creative problem solving. I think for a long time now, probably since the birth of programmatic digital, we've spent an awful lot of time talking to clients about the pipes and how they go together and how we do things, as opposed to what we're doing and why.
Actually, it's good news
For all the challenges that need to be overcome, agentic AI, is delivering real business benefit. It's clear from speaking to both Pacey and Randeria, that agentic (and the generative capabilities aligned to it) is not a transitory technology wave, its a change in the way business operates.
For T2, some of the strongest returns have emerged from pricing and promotional optimisation, helping improve commercial outcomes and revenue performance.
Per Randeria: "We’ve been deploying Agentic in pricing and promotional optimisation, and I can tell you it’s a cornerstone of data analytics and finance. It has instantly delivered really tangible commercial results. We’ve got smarter, we’ve got faster. It’s helped us define the right promotional and pricing levers that help us deliver the commercial outcome, by helping us to reshape our brand to be worth paying more for.
"It’s helping us drive unit revenue really significantly."
For Omnicom, Pacey pointed to something less tangible but potentially more transformative: A renewed focus on creativity.
"For me, and I've referenced it before, what I have seen across our business is potentially not what Peter talked about in terms of immediate, obvious commercial impact, but it's the rebirth of creativity and creative problem solving. I think for a long time now, probably since the birth of programmatic digital, we've spent an awful lot of time talking to clients about the pipes and how they go together and how we do things, as opposed to what we're doing and why."
He said the ability to start with a blank canvas and to build agents in a safe environment means, "people are coming to me and talking to each other about something they built in an afternoon or in an hour that's completely changed the way they think about something."
Furthermore, he said, "The ability to understand an entire category and what has worked and not worked historically, then think about a problem, or look at a piece of data rapidly and in more depth than they've ever been able to before, and therefore come up with an idea, is so exciting. I really hope the year ahead sees a real creative renaissance in terms of how we use this technology to solve problems, because the potential is absolutely there."
A year after agentic AI first emerged as a serious enterprise discussion, that may be the most important lesson. The technology itself is no longer the biggest obstacle. The winners are increasingly likely to be the organisations that can govern it, supervise it, integrate it into workflows and understand its economics before the costs, risks and complexity scale alongside its potential.
*The half a billion Peter references was the total cost of AI in the month, not just tokens.
