The Signal

Hello Demo,

The first AI agent demo I’ve ever witnessed was pretty amusing. It was for an agent created to order a margherita pizza.

It happened while I was attending a developer gathering hosted by Cloudflare at its London office in April 2025. An ElevenLabs engineer was showcasing how you could build an AI agent that integrates the software company’s voice tools into chatbots like Claude via its recently released model context protocol server. Basically, you can type in a prompt and have the agent do the call for you.

Back then, companies had just started exploring different applications for an AI agent. What exactly could it do?

The bar has since risen. It’s no longer enough for an agent to play personal assistant. We want it to audit workflows and catch capital leakage, with companies scrambling to onboard agents into their processes.

But the results have been less than stellar: 95% of AI pilots failed to deliver measurable returns, according to an MIT Media Lab report. Over 40% of agentic AI projects will be canceled by the end of 2027, Gartner says, citing escalating costs, unclear business value, and inadequate risk controls.

That is likely because many companies focus on building the AI application layer without digging into the invisible connectivity that sits underneath.

In contrast, Axium Industries, an industrial AI startup based in Singapore, treats connecting fragmented data as the hardest and most valuable part of the work.

Image credit: Ulla

“We spend 80% of our time working on data, only 20% on AI,” says Jeremy Ang, co-founder and CEO of Axium. “If any company comes in to say that they have it the other way around, I think customers should start to walk away knowing they’re in for a bad time with that approach.”

For this month’s edition of The Signal, Tech in Asia speaks with Ang, alongside Tobias Leong, Axium’s co-founder and CTO. The company took inspiration from the forward deployed engineering model (FDE) invented and popularized by Palantir, which sends engineers to clients’ factories and lets them figure out what needs solving. The co-founders say most of the heavy lifting is in data engineering.

It’s part of a bigger conversation about what it means to deliver AI-driven productivity. Is it a productive day when you burn tokens to outsource your pizza order to an AI agent? Or is it so when your agent finds a way to increase your company’s revenue by, say, 10%?

These hardly feel like a fair comparison. Yet they’re both part of the same conversation about what AI agents are actually for.

We’re no longer interested in asking, “What could it do?” The more intriguing question is, “What should it do?”

Yunindita Prasidya, editor

Despite talks of an AI agent bust, Axium Industries, which builds AI agents for industrial clients, is proving that providing AI agents as a service still works.

Over the past year, the company has grown from working on two to 19 projects, and its yearly contract sizes have increased from six to seven figures in USD. It found product-market fit early – the firm broke even in the first six months of its operation. Its clientele has also broadened from oil and gas companies to those in manufacturing, maritime, retail, and aerospace.

Axium helps clients identify capital leakage in their procurement, inventory, and resource allocation for production. Once the team finds a way to improve the production capacity of an existing factory by, say, 10%, that trickles down into millions of increased profit and revenue for the client, Ang and Leong tell Tech in Asia.

That’s why the business takes off: because results show up on the balance sheet.

Start with the unscalable

“We start by building things that might seem unscalable at the start,” says Leong.

By sending engineers into clients’ premises to build, design, and ship custom agentic solutions, it is taking a page out of Palantir’s FDE playbook. Axium’s team spends considerable time on the factory floor, observing workflows, understanding how things operate, and talking to people.

In the process, they check for alignment and misalignments between what the data is telling and what is happening on the ground.

This process is especially important when you are working with high-volume industrial data. These factory floors are equipped with sensors that produce billions of data points every single day.

For AI agents to be useful, they need to receive data via reliable connections. Then they must understand the data’s context.

Without these, even the most capable AI agent works with an incomplete picture. Some companies might be tempted to skip these steps.

“A lot of times it is the structuring and contextualizing of data – providing it at the right time – that becomes one of the most difficult parts to solve,” Leong explains.

It’s laborious work because data comes in different formats, and you can’t always predict what they will be.

For one client, Axium had to review more than 12,000 documents written in several languages. The data also had different attributes – not everything was recorded digitally. Some of the scanned handwriting was difficult to read, even for humans.

Another case, one of the hardest ones Axium had to solve, involved a customer that didn’t have the right data pipelines. The client’s ERP, which tracks business data, had no API that Axium could connect to.

In other words, there was no built-in way for outside software to connect to the system and pull information out. To do that, Axium had to figure out an engineering workaround.

“Assume you need to read a PDF on your computer,” Ang says. “You need to send it to me, but there are no ports on this laptop, and you cannot send it through the internet. What’s the best way to give that PDF to me? I have to take pictures of your screenshots again and again until I get the full picture of that PDF.”

“That’s what we needed to do,” he adds. “Multiply that by almost a million times.”

Axium’s experiences show how companies must lay the right foundation before benefiting from AI. This includes digitizing processes and ensuring that data can reliably travel between devices. As operations become more distributed and data-intensive, moving that information reliably between machines, systems, and locations has become more crucial than ever.

Earning trust

It is no surprise that skepticism abounds. With AI agents’ ROI shortfalls and the rise of “agent washing” – a practice in which vendors rebrand basic chatbots into agentic AI – customers are becoming wary of false promises.

Axium recognized this sentiment among its clients. Some of them had shopped around, talking to over 20 so-called AI companies.

They weren’t satisfied with the options on the table. The clients didn’t want another AI wrapper solution.

“You cannot be hand-wavy about it,” Leong explains. “You either know exactly what the decision threshold is for, say, concentration level for some chemical, or you don’t. You have to get it absolutely right.”

That means an AI company catering to industrial clients like Axium must understand the physical reality of the factory floor and the process parameters that come with it.

“You have to know all this otherwise you can’t get the deal,” notes Ang.

Since they started the business, Ang and Leong have surrounded themselves with a ring of experts from big industrial companies with 20 to 30 years of experience in the space. These experts guided the two co-founders on what to build, what to solve, and what metrics to optimize for.

Still, clients have to be won over.

“We had a client challenge us,” Leong recalls. “They withheld a lot of the important information. They said, ‘Let your AI agent figure it out, and then we’ll compare notes.’ It’s honestly quite scary. Our hearts were beating very fast.”

“But when we told them the number, they were totally shocked,” he adds. “It was spot on. Exactly the right number they were looking at. It proves to them, as well as us, that this actually really works.”

Place your bets strategically

Axium is selective in investing its efforts. It places bets only in areas that could generate substantial economic value or where the margin is.

“AI is good for three types of workloads: what humans cannot do yet, what humans don’t want to do, and what humans shouldn’t do because they’ve got better things to do,” Ang says. “We always aim for the first one. That’s the money maker.”

“We knew the power of AI, the power of data engineering,” Leong adds. “We wanted to find a problem that was big enough, and we knew that’s going to be something industrial-related because that underpins a large chunk of the GDP here in ASEAN.”

Manufacturing is the largest contributor to the regional economy. Part of the broader industrial base Axium targets, it accounts for over 20% of several Southeast Asian countries’ GDP.

There is a huge unlock in value creation when AI agents optimize this area, but only once the factories are connected enough to make that possible. The sensors, machines, and legacy systems on the floor have to talk to each other before an AI agent can do anything useful. It’s not just connecting infrastructure; it’s part of what determines whether AI can operate reliably at scale.

Currently, Axium is building its internal capacity to do more of that. The firm is hiring in several cities in Asia. It estimates that it will have 40 to 45 team members within the next few quarters.

Axium aims to be as efficient as possible. Each of its existing employees is running a few projects. The company’s bet on expanding its team has more to do with serving customers in different countries, each of whom requires hyper-localized services.

Much work remains to be done when it comes to implementing AI: There’s a 95% failure rate in delivering returns, 90% unsatisfactory deployment, and 40% of projects risk being canceled.

These figures seem to be the result of companies chasing the AI layer first, when the real foundation starts earlier: connecting the data, systems, and infrastructure that allow AI to work reliably in the first place.

Ideas from conversations we’ve tuned into this month, and what your team could try with them.

1. Give your AI agents a spending limit

ClickHouse CEO Aaron Katz argues that as agents start buying their own software, companies need to treat them like financial actors. Without that, purchasing turns into costs nobody approved.

Try this: Before an agent provisions its own tools, give it what you’d give a new hire with a company card: an identity, a budget, and a manager who checks the receipts.

2. Turn your team’s knowledge into a product

Cloudflare’s Justin Joyce says AI’s real value doesn’t lie in automation. It’s in converting knowledge stuck in one expert’s head into something searchable and consistent.

Try this: Pick one process only your best person does well. Write down the decision rules before handing it to an agent.

3. Plan your AI roadmap three months out, not one year

OpenAI product lead Tara Seshan warns that building for where models are today and where they’ll be in a year both fail. The only horizon that survives is two to three months, tested against real usage. A longer bet often means shipping something the next model release makes obsolete.

Try this: Cut your next AI feature’s roadmap to what’s realistic this quarter. Ship the simplest version and let usage decide what’s next.

This newsletter is made possible thanks to XLSMART, the Best Mobile Network 2026 – awarded by Ookla.

Through continuous investment in digital infrastructure and the expansion of its 5G network, XLSMART is building faster, more reliable connectivity across Indonesia – helping people and organizations stay connected, collaborate, and work faster.

Learn more about XLSMART’s 5G network & coverage.