The Signal

Hello Betamax,

I didn’t incorporate AI into my workflow until late last year. I was, to a great extent, adamant that my writing process – from research to finalizing my drafts – be free from the influence of LLMs. (Sure, Grammarly was installed on my computer as a browser extension, but that was pretty much it.)

These days, I’ve onboarded Claude as a thinking buddy: It gives me updates on discourses I’m following, discusses story angles with me, and writes research briefs on people and companies I’m profiling.

Instead of pondering how AI might influence my work, I find myself wondering how my work is shaping these models. Does it? What exactly am I handing over when I interact with the chatbot? How valuable is this information? I can continue expanding my list of questions, but you get the point.

Companies could benefit if their employees asked more of these questions. Today, more workers are embedding AI into their workflows; many use tools their organizations haven’t approved.

It has created a new risk for enterprises: shadow AI, where individuals disclose their work to public AI platforms such as ChatGPT and, in the process, compromise company safety.

This year, shadow AI-related incidents accounted for 43% of security incidents at breached organizations, up from 20% the previous year. It drove average breach costs up to US$5.39 million this year from US$4.63 million last year, per IBM’s report.

Image credit: Ulla

No wonder many are looking to install private AI, which runs within a company’s own infrastructure. Companies know their workers benefit from AI, so it’s best to embed the tools within the company. This way, usage does not come with the risks.

Against that backdrop, the market for private AI infrastructure is vast and growing. It was reportedly valued at US$33.72 billion in 2025 and is projected to grow at an annual rate of 18.7% through 2030.

That market growth says a lot about the priority level of private AI.

Ivan Lee, founder and CEO of San Francisco-based AI company Datasaur, argues that stakes could be high.

Datasaur builds private AI infrastructure for enterprises. It also runs a data labeling and annotation business.

“What if your company’s largest business contract was copied and pasted into ChatGPT?” Lee asks. “How would your security team feel about that?”

In this month’s edition of The Signal, Tech in Asia speaks with Lee about what it takes to build connections that let corporations benefit from AI without exposing themselves to potential security breaches.

If you work in highly sensitive sectors such as healthcare, legal, or finance – as many of Datasaur’s clients do – a reliable and secure connection is as important as, if not more than, an efficient one. How does one strike a balance between the two?

It’s an art of choosing the right trade-offs, apparently.

Yunindita Prasidya, editor

Private AI hits an inflection point this year, as risks loom large. Shadow AI is one; a changing regulatory landscape is another.

In June, Anthropic reversed a zero data retention guarantee for its Fable model, which many enterprises see as a prerequisite. The move sparked backlash among enterprise customers who feared that retained data could train future models and compromise their intellectual property.

In the same month, Anthropic blocked two of its models for all foreign nationals after an order from Donald Trump’s administration. However, it was lifted by the end of the month.

These developments have eroded users’ trust in public AI systems.

“Imagine if you’re a hospital and you’re dependent on Anthropic models and suddenly, for reasons outside of Anthropic’s control, the US government just turns it off for two weeks,” Lee says. “That can be disastrous.”

Enterprise interest in private AI has risen significantly because of these developments, Lee adds without disclosing any numbers.

Here’s the challenge for him: Private AI infrastructure can look very different from one client to the next, and a reliable and secure network has to be custom-fit. Scaling AI adoption, it turns out, isn’t just about better models. It takes the right compute, cloud architecture, security controls, and connectivity – all working together.

No one-size-fits-all in privacy

“We always talk about privacy as being a spectrum,” Lee says.

“There are certain clients where, as long as it’s on their private cloud – their AWS, their Google Cloud Platform – it’s good enough,” he adds. “We also have clients who say, ‘Nope, I’m not even allowed to put this on Azure. This is too sensitive an information. But we do have these Dell hardware racks. Can you install something there?’ So we’ll do that. We do all the above.”

With the latter approach, though, there’s the additional challenge of establishing a dependable connection. The software may not work, or the connection between the graphics processing unit (GPU) and the model is unstable, requiring more debugging.

Virtual private cloud, on the other hand, has streamlined GPU setup, making it much easier for Datasaur to deploy private AI solutions on these platforms.

Datasaur must build different types of infrastructure and tools to suit customers’ preferences, whether they are in the cloud or on-premises, and whether they want chatbots or agents.

Last year, the company was still deploying many private AI chatbots. This year, Datasaur began nudging clients toward AI agents as models become more powerful.

How do customers pick between one or the other?

Chatbots were a default by necessity. When enterprise customers think of AI, it is what comes to mind.

Reliable agentic workflows only became viable industry-wide around November last year, when Anthropic, Google, and OpenAI released competing frontier models – Opus 4.5, Gemini 3 Pro, and GPT-5.1 – within about a week of each other. Datasaur expanded its offering in response.

Lee explains that agents are more effective in accomplishing tasks. When you’re building a private AI infrastructure, it helps to have always-on agents doing background work, making reliable and trustworthy connections more seamless. But chatbots are still available if that’s what the customer prefers.

Complex problems expand what’s possible

For one client, Datasaur had to access information available only on a web portal that didn’t have an API or an MCP server – it had no working port Datasaur could connect to. No one was maintaining the 20-year-old system.

The team ended up using a framework called Browser Use, which lets models control web browsers like a human and extract the data needed.

“Browser Use has been demonstrated and proven in much larger frontier models,” Lee says. “The way we did it was using Google’s Gemma model. We took the same concept of browser use, but we were able to engineer a way to push this tiny little model to perform browser use successfully.”

Frontier models like Claude’s Opus, Fable, and Mythos have multiple trillions of parameters. In comparison, Google’s lightweight open model Gemma 4 has up to 30 billion parameters – a much smaller model. Datasaur chose it because it’s one of the models Datasaur was allowed to use, given its client’s security requirements.

For context, a model’s parameter count influences how much complexity it can handle. A higher count means a model has greater capacity for capturing nuances, complexity, and patterns.

But Datasaur had to work with what it had.

“We were able to do this locally deployed, without any data going to OpenAI or Anthropic, or even to the cloud at all,” Lee adds.

This was one of Datasaur’s toughest challenges. It highlights two key lessons: Connectivity is not available by default, and a working portal is often a product of ingenuity.

Lee reckons the toughest cases are often the most satisfying ones.

Building with Lego blocks

Datasaur works with clients across various industries: healthcare, finance, legal, and the public sector. The challenge of delivering for clients has more to do with individual workflows than the industry itself, says Lee.

How complex can it be?

“We are currently in discussion with a client that has to follow an internal playbook for this workflow, and it’s over 600 steps,” Lee shares. “Within those steps, they also have to talk to many different systems, go through different departments.”

The project will likely take Datasaur three months, which is an anomaly for the firm today. Lee aims for projects with a quicker timeline – anything between four and six weeks.

Part of what makes faster delivery possible is what Lee and his team learn from every job. Projects may look completely different at a glance, but they often have commonalities.

For example, reading invoices and handling a customer support ticket may look like very different projects at first glance. But they both may need integration with Google Drive, WhatsApp, or a Salesforce CRM. Once integrated, they can be reused across projects – building connectivity, evidently, is a long-term investment.

“We talk about them internally as Lego blocks,” Lee explains. “There will be times when a Lego block is unique to this client. We’re never going to use it again. That’s fine.”

“But if you think that this Lego block could be reusable, make sure you document it, share it with the team, and the next time another project requires that same Lego block, we reuse it,” he adds.

That same logic applies to Datasaur’s internal operations. Lee and his team look for manual work that recurs across different projects and build reusable tools for future ones.

The speedy delivery is also a nod to Datasaur’s AI-native operation, where it leverages agents to work faster and better.

“What took my AI engineers six weeks to deliver in 2025, an agent can get 90% of the way there in just six hours,” Lee notes. “A year ago, we wouldn’t have been equipped to handle a 600-step process. That would have been insane for us to tackle.

“Now, it’s still a lot of work, but we know how to do it,” he adds.

Lee estimates his team’s productivity has tripled in the last 12 months, while his headcount increased only by 10%.

All things considered, private AI is anything but simple, cheap, or quick to get right. It requires orchestrating heavy compute, flexible cloud environments, strict security, and seamless connectivity.

While it can look different for different companies – requiring dedicated time to debug flaky GPU connections, a new engineering workaround to connect to legacy systems, or a 600-step process – it is clear that more companies are seeing private AI as a non-negotiable.

And it’s better to pay for a reliable, trustworthy connection than to shoulder breach-related costs.

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

1. Don’t stack up too many agents

OpenAI’s Tara Seshan and Nan Yu frame AI adoption as having a ceiling tied to how many separate agents a person can supervise at once. Their fix is to bundle related tasks into a single conversational interface rather than scattering them across standalone agents, with coordination features that handle the supervision automatically.

Try this: Fold an AI agent into a tool your team already uses before rolling out another standalone one.

2. Give every agent an employee ID

Certis treats its AI agents the same way it treats staff. At the same time, the company implements role-based training and clear communication about why each tool is deployed, on the premise that adoption stalls when employees don’t trust how the rollout is handled.

Try this: Before scaling agents across a team, give each one its own identity, restricted permissions, and a named owner – the same way you’d onboard a new hire.

3. Treat long-running agents as potential insider risks

Microsoft’s Satya Nadella warns that long-running AI programs create internal security threats, one that needs the same containment process as a software release: sandboxing new agents before they go live, limiting their access to only what a task requires, and tracking their activity closely.

Try this: Before letting an agent run unsupervised, give it the same offboarding discipline you’d give a departing employee: minimum access, logged activity, and someone watching for anything unusual.

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.

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