He’s Led AI Transformations Across 10 Countries. His Award-Winning Projects Weren’t the Most Advanced.

The AI industry has become very good at selling more: more models, more agents, more automations, more tools. But every addition creates something else a business has to understand, govern, maintain, and eventually fix.

Jonathan Chew takes the opposite approach.

A self-described “AI Minimalist,” he has led AI transformations across businesses in 10 countries, is a two-time AI and innovation award winner, and was behind Tune.Ai’s AI Project of the Year recognition. With an MSc in AI and Machine Learning alongside an EMBA, his work spans context engineering, AI governance, agentic systems, and taking AI applications from prototype into production.

His experience has led him to a deliberately unfashionable conclusion: the most impressive AI system is not necessarily the one doing the most. Often, it is the one solving a clearly defined problem within clear boundaries, in a way the business can still understand and maintain after the person who built it leaves.

In this conversation with NewInAsia, Jonathan explains why expanding AI stacks can become liabilities, why proprietary context could matter more than the model underneath it, and why “boring” AI may ultimately be the kind businesses should want.

More AI Tools Mean More Problems You Have to Own

Q: You describe yourself as an “AI Minimalist” at a time when companies are being encouraged to adopt more AI everywhere. What are businesses adding today that you think they’ll regret having to maintain tomorrow?

A: Right now, companies are adding three things at speed.

First, overlapping subscriptions: three tools that each do 80% of the same job, none mastered, all renewing.

Second, automations built on processes nobody ever documented. When we automate a mess, we get a faster mess — and now that automation is load-bearing.

This is further augmented in the age of AI.

Every tool you adopt has a carrying cost that never shows up in the demo: licences, security reviews, training, the context-switching tax on your team, and potentially one person, or a select few, who quietly becomes the only one who knows how it works.

Regret will come from focusing on acquiring more tools without first spending the hard, boring effort of designing the shape of work with intent and judgement.

His Award-Winning AI Wasn’t the Most Advanced AI

Q: Your work has received multiple AI and innovation awards, including AI Project of the Year for Tune.Ai. What did those projects teach you about where meaningful AI innovation actually comes from — better technology, better problem selection, or something else entirely?

A: None of the projects that won awards were the most technically advanced things we could have built.

What they had in common was an unglamorous problem chosen well: frequent, painful, measurable, and owned by someone who wanted it solved. Then we scoped ruthlessly until the system was small enough to explain in one sentence — and small enough to maintain.

My somewhat controversial take is this: the model is the least differentiated part of the build. Everyone has access to the same frontier intelligence through a subscription and the same APIs they can pay for.

But innovation is not access. It is judgement — the layer of decisions about where AI is allowed to act, what it is grounded in, and where the human stays in the loop.

I have watched this play out across three technology waves now — dot-com, mobile, and AI — and the pattern repeats.

The winners were rarely the teams with the most advanced technology. They were the teams who picked problems people actually had and shipped things people actually used.

The Model Is Rented. Context Is Owned.

Q: You specialize in context engineering — what AI knows, remembers, and is trusted to do. As foundation models become increasingly capable, could a company’s proprietary context become a more meaningful competitive advantage than the model it uses?

A: It already is.

The model is rented. Context is owned.

Foundation models are converging into a utility. Every quarter, they get more capable and less differentiated. Whatever advantage a widely available model gives you is an advantage your competitor can access too.

What nobody can rent is what your company knows.

That proprietary context includes your decision history, pricing logic, customer tone, standard operating procedures, and the underlying reasons behind them. The judgement call that went wrong once and became a rule.

That is context — and context engineering is the discipline of deciding what your AI knows, what it remembers, and what it is trusted to do.

Inside companies, the bottleneck is almost never model capability. It is that the company’s knowledge lives in people’s heads, scattered inboxes, and undocumented habits. AI cannot leverage what was never written down.

So the most valuable AI work I do often looks suspiciously like documentation: grounding systems in real documents so every answer can be traced to a source instead of a guess.

The compounding effect is the point.

A better prompt helps you once. Context compounds: every SOP captured, every guardrail defined, every decision recorded makes the next AI system cheaper to build and safer to trust.

And it travels — when a better model ships, you swap the engine and keep everything that made it yours.

The model is the engine. Context is your knowledge of the roads.

Engines are for sale everywhere.

Agentic AI Doesn’t Remove Management. It Exposes It.

Q: You work closely with leaders navigating AI transformation and the shift toward agentic systems. What misconception about agentic AI do you find yourself challenging most often?

A: The expectation that agents are autonomous employees — and that an agentic workforce can be achieved in a single quarter.

There are inflated expectations of what AI can do without a full understanding of the progressive steps required to get there.

Technology is rarely the bottleneck. Agentic AI industrializes delegation.

It does not remove management; it exposes the quality of it.

If we cannot brief a capable human on a task — the scope, the context, the boundaries, what done looks like — we cannot brief an agent either.

Vague delegation to a person produces one bad outcome. Vague delegation to an agent produces bad outcomes at scale, confidently.

The version that works is bounded delegation.

Narrow scope. Explicit guardrails — what the agent may never do matters more than what it can do. Defined handoffs, so everyone knows where the agent stops and the human begins. And a named owner for the outcome, because accountability does not transfer to software.

A second misconception hides inside the first: that more agents mean more value.

Multi-agent swarms demo beautifully and fail quietly in production, because every extra agent multiplies the ways the system can go wrong.

Add context and tool bloat to the mix, and the points of failure scale further.

If It Only Works on Its Best Day, It’s Still a Demo

Q: After implementing AI across businesses in 10 countries, what do you now consider the clearest dividing line between an impressive AI demo and an AI system that can be trusted to do real work?

A: A demo is judged on its best day, on an ideal path. A system is judged on its worst day — edge cases and scalability.

Every AI system will be wrong eventually. The technology is probabilistic, unlike rule-based applications, which were deterministic.

Trust, therefore, is not about accuracy alone. It is about whether failure is visible, bounded, and recoverable.

Can the user tell when an output is shaky? Are claims traceable to a source they can check? Is there an escalation path to a human? And does anyone notice within a day when it breaks — or do you find out from a customer three weeks later?

After implementations across ten countries, the pattern is the same everywhere.

Demos are built for the happy path: the builder’s laptop, the builder’s prompts, the clean example.

Real work is the messy input, the second user, the edge case, the new hire who was never trained on it.

Demos optimize for wow. Systems optimize for boring — predictable, repeatable, checkable.

In this business, boring is a compliment.

I generally start my evaluation with three questions:

1. Can the least technical person on the team use it without the builder in the room?
2. When it fails, who finds out, and how fast?
3. If the person who built it resigns tomorrow, does the system survive the handover?

If the honest answer to that last one is no, you do not have an AI system.

You have a demo with a dependency.

That is why my definition of success in AI systems is unglamorous: time handed back, quietly, every week, to people who stopped noticing the system because it simply works.

Final Thoughts: In AI, “Boring” Might Be the Highest Compliment

Jonathan’s argument ultimately challenges a familiar assumption about AI adoption: that greater capability should automatically lead to greater complexity.

Instead, the advantage may lie in restraint — choosing a problem worth solving, capturing the context that makes the solution useful, defining where AI can and cannot act, and ensuring the system can survive failure and handover.

As models become easier for everyone to access, those decisions become harder to copy. The technology can be rented. The understanding behind how a company applies it cannot.

And perhaps the strongest sign that an AI system has succeeded is not that people keep talking about it.

It is that they stop noticing it altogether — because it quietly gives them time back every week.

Read the Chinese article here.