Most conversations I have with other founders about AI land quickly on the same question: what can we automate? It is a reasonable instinct. But for founders and leadership teams building AI-native businesses, the more consequential question is a different one: if AI handles more of the execution, what exactly are humans for?

The answer is more specific than “creativity” or “vision,” two words invoked so often they have lost their precision. The human skills that matter most in an AI-first business are strategic thinking and systems thinking. Together, they form the architecture layer that determines whether an AI-native business compounds in value, or simply runs faster toward the wrong outcomes.

Why these two skills, specifically

There is a structural reason why these skills remain irreplaceable, beyond familiar arguments about nuance and context.

AI is backwards-looking by design. Every model, every output, every recommendation is built on patterns extracted from past data. You can feed it new context and real-time signals, and it will process them intelligently. But its underlying reasoning is always a function of what has already happened.

Strategic thinking is fundamentally forward-looking. It requires forming a perspective on where a market is going before the evidence is conclusive, on what customers will want before they can articulate it, on which bets to make when the data is incomplete by definition. The best strategic decisions are made precisely in the space where historical patterns are the least reliable guide. That is the space AI cannot occupy.

Systems thinking adds a different dimension: the ability to see how parts of a business interact, how a change in one area creates second and third-order effects elsewhere, and how the overall system produces outcomes individual components cannot explain on their own. Without someone thinking at the system level, AI initiatives tend to solve isolated problems while creating new friction at the handoffs between them.

Both skills are intuition-heavy, built on real-world experience, judgment grounded in a specific business context, and a tolerance for ambiguity that cannot be trained into a model on historical data.

Also Read: What to actually prioritise when your board wants AI and everything feels urgent

How to choose

The practical question for any leadership team is what goes to AI, what stays with humans, and what requires both. A simple Signal vs. Execution layer model structures every capability across three levels.

• Execution is where defined tasks get done: content generation, data retrieval, report formatting, customer query handling, workflow automation. AI operates excellently here. The inputs are bounded, the success criteria are measurable, and the patterns are well-established. It is now reasonable to default to AI at this layer and free your team from the cognitive overhead it was consuming.
• Orchestration is where workflows are coordinated across functions, systems, and agents. AI can support this layer, but someone has to have defined the logic, the rules, and the exception-handling criteria that make coordination work. Systems thinking is what makes orchestration coherent. A human needs to own this layer even as AI tools increasingly execute within it.
• Direction is where the business decides what it is building, why, and in what sequence. AI can inform direction with data and scenario analysis. It cannot own it. Direction requires the forward-looking intuition that AI structurally lacks, and the systems-level awareness to understand how today’s choices shape tomorrow’s constraints.

A practical example: we recently rebuilt the GTM stack for one of our SaaS businesses to be fully AI-native at the Execution and (partially) Orchestration layers. Prospecting, lead qualification, outreach sequencing, and follow-up logic all run through AI-driven workflows. The team now spends 30 to 45 minutes per week on that function instead of roughly 20 hours, with a 1.8x higher response rate and 1.6x higher close rate. The Direction layer did not change. Who to target, what positioning to lead with, and which segments to prioritise remained entirely human. What changed is that AI executes that judgment at a scale and consistency no manual process could match.

The errors I see most businesses make are automating Direction (outsourcing strategic and systems choices to tools that optimise for past patterns) or leaving Execution to humans (wasting human capacity on tasks AI handles better). The framework is a forcing function to avoid both.

Also Read: Think with AI: The new skill for social entrepreneurs

What to pay attention to in the next 12 months

Competitive pressure will push many teams to deploy AI broadly and quickly. Some of that pressure is real. A lot of it will produce systems that move fast in the wrong directions.

Three things worth keeping close.

• Audit your direction layer: How many strategic choices are currently being made by default, through inertia, or by deferring to tools and benchmarks? If your team spends most of its cognitive energy at the Execution layer, Direction is likely underinvested.
• Build systems thinking as an organisational capability: The businesses that compound from AI are those where someone is consistently asking: how does this initiative interact with everything else we are building? This thinking needs to be embedded in how you design and iterate on AI deployments, not just exercised at the top.
• Resist the pull toward AI as a strategy: Using AI is not a strategy. It is a capability. Durable competitive advantage comes from a clear view of where you are going and a coherent system for getting there, with AI as a powerful layer within that system.

The point

AI will keep getting better at execution, and it will increasingly support orchestration. The layer it cannot occupy is Direction, and the reason is structural: it is built on the past, and Direction requires a view on the future.

Strategic thinking and systems thinking are what make the difference between an AI-native business that compounds and one that scales its existing assumptions faster.

These are the skills worth protecting, developing, and keeping close to the center of how you make decisions.

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post The two human skills that make AI-native businesses actually work appeared first on e27.

In modern meeting rooms today, there is a scene that feels both strange and normal at the same time. Someone opens a laptop, types a few sentences, and within seconds, a business proposal appears, along with market analysis, source code, presentation designs, and even marketing strategies that once required a small team working for days. On the other side of the room, someone sits quietly for a moment before speaking. They are not slower. They are thinking: “Is this the right decision for the people who will be affected by it?”

Moments like that are beginning to change how we understand human value.

For decades, humans were valued for their ability to process information. The faster someone could calculate, memorise, write reports, or organise data, the more valuable they became inside organisations. But AI is beginning to shift that foundation. Machines can now write faster, analyse more broadly, read thousands of documents without fatigue, and even generate ideas that appear creative. Many professions are starting to experience a quiet discomfort: if AI can perform most intellectual tasks, then where exactly does human value still exist?

That question is no longer philosophical. It is becoming personal.

A programmer watches AI generate hundreds of lines of code in seconds. A designer sees AI create illustrations from a single prompt. Analysts watch dashboards and insights appear automatically. Even writers quietly wonder whether readers can still distinguish between human writing and machine-generated text.

Yet the longer we live alongside AI, the more something interesting becomes visible: speed was never the true core of human value.

AI is incredibly powerful at answering. But humans still live in a much greyer territory: deciding which questions are worth asking.

And that difference matters.

AI can help companies optimise profits. But humans decide whether those profits are achieved in ways that damage or strengthen society. AI can help underwriting systems assess risk within seconds. But humans understand what it feels like to be afraid because of illness, unemployment, or the desire to protect one’s family. AI can create highly efficient business strategies. But humans decide whether organisations still retain a sense of humanity.

That is where I have started to believe human value is not simply “the things AI cannot yet do.” That definition is too fragile. AI capabilities will continue to evolve. If human value is defined only by the remaining gaps in machine capability, then our identity will continue shrinking every year.

What feels more fundamental is this: humans give meaning to decisions.

Also Read: What to actually prioritise when your board wants AI and everything feels urgent

AI generates possibilities. Humans choose the consequences.

And choosing consequences means bringing morality, empathy, fear, hope, life experience, personal wounds, culture, and even love into the process of work. Those things are difficult to measure in spreadsheets. Yet they are often what separates systems that are merely efficient from systems worth preserving.

The problem is that the narrative around “uniquely human skills” sometimes sounds too comforting. Many conferences claim the future will belong to empathy, creativity, leadership, and communication. It sounds beautiful. But honestly, AI is already entering those spaces too. AI can now speak warmly, write poetry, act as a “companion,” and generate fairly creative ideas.

So are we truly redefining human value? Or are we simply rebranding the tasks that automation has not reached yet?

I do not think the answer is clear.

And perhaps that uncertainty is important to acknowledge.

Because there is a strong possibility that a large portion of human work will change dramatically. Not only repetitive jobs. Creative and strategic work is being reshaped too. Many people are quietly experiencing a professional identity crisis. They once felt valuable because of specific expertise. Now that expertise can be replicated by AI at low cost and extraordinary speed.

At that point, humans are forced to confront a question deeper than “How do I remain competitive?”

The question becomes: “Who am I when my abilities are no longer rare?”

And that is not a question AI can answer for us.

I see this shift directly in technology itself. In the past, engineers were valued primarily because they could write complex code. Today, AI can generate much of the boilerplate work. But the engineers who remain truly valuable are becoming those who understand business context, operational risk, stakeholder conflict, long-term implications, and decision-making under uncertainty.

In other words, human value is shifting from production toward judgment.

Also Read: Think with AI: The new skill for social entrepreneurs

It is no longer about who can build something the fastest. It is about who can wisely decide what should be built in the first place.

Strangely enough, those abilities are often born not only from formal education, but from life itself. From failure. From loss. From making bad decisions. From leading people during difficult situations. From understanding that behind every dashboard metric are real human lives.

Perhaps that is why future organisations will change how they hire people.

Not only by measuring technical skills, but by evaluating the ability to think across contexts. The ability to navigate ambiguity. The ability to maintain moral direction while systems become increasingly automated. The ability to build trust in a world flooded with synthetic content and algorithmic decisions.

Ironically, the more advanced AI becomes, the more expensive human trust becomes.

Because in an era where almost everything can be generated by machines, people begin searching for something that feels real.

Not merely correct according to data, but emotionally sincere.

Not merely efficient, but humane.

Not merely intelligent, but morally accountable.

At the same time, I do not want to romanticise humanity too much. Humans are also filled with bias, ego, manipulation, greed, and error. Many of the world’s worst decisions were made by humans, not AI. So not everything “human” is automatically good.

Which is why the future may not become a battle between humans and AI.

It may instead become an ongoing negotiation between machine capability and human wisdom.

And we still do not fully know what that final shape will look like.

Perhaps many jobs will disappear, while entirely new roles emerge that we cannot yet imagine. Perhaps organisations will become far smaller yet more productive because they are supported by AI agents. Perhaps one person will be able to build a global company with only a handful of people and a network of AI systems. Perhaps degrees, titles, and traditional corporate structures will slowly lose meaning.

Or perhaps humans will simply grow exhausted from living inside worlds that feel overly automated, and begin valuing slower, more authentic, imperfect human interaction again.

I do not know.

But one thing feels increasingly clear.

Also Read: How AI is changing what an SME team actually looks like

Human value may not come from being faster than AI.

Nor from being smarter than machines.

Perhaps human value emerges from our ability to retain consciousness, responsibility, and meaning in a world where almost everything can be automated.

And perhaps the most important question is not:

“Will AI replace humans?”

But rather:

“When almost everything can be done by machines, what are the things we still want to preserve as deeply human?”

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post In a world optimised by AI, what is left for humans? appeared first on e27.

For a while, differential privacy was spoken about as though it might do for data privacy what encryption did for data in transit. A hard technical answer to a messy institutional problem. The theory was elegant, the guarantees were rigorous, and the early signal from major adopters was powerful. 

That gap matters because it tells us something larger about technology strategy. Differential privacy did not stall because the mathematics was weak. It stalled because organisations kept treating it as a universal privacy answer when it is really a precision instrument for a narrower class of problems. It is very good at answering one hard question. How do you release useful aggregate information while limiting what can be learned about any one person? That is not the same as solving privacy in the round. 

It solved a narrower problem than the market wanted

Most organisations do not actually need a mathematically formal guarantee for every privacy question they face. They need a working combination of access controls, minimisation, retention discipline, contractual restrictions, governance, and operational trust. Differential privacy sits inside that world. It does not replace it. The mistake was to imagine that a strong formal guarantee at the output layer could make the wider privacy problem feel settled.

In practice, most institutions still need to govern collection, purpose, access, sharing, deletion, model use, and accountability separately. That is why differential privacy often ends up as a specialised control rather than the centre of the privacy operating model.

This is also why the technology feels simultaneously important and oddly non dominant. It addresses a real problem, but not the whole one. Strategists often back technologies that appear to simplify governance. Differential privacy usually does the opposite. It sharpens one guarantee while leaving the surrounding organisational obligations very much alive. That makes it more honest than many privacy narratives, but also harder to sell as a universal answer.

Also Read: How to build customer trust with improved data privacy

The trade off is too visible to ignore

Differential privacy makes privacy expensive in a way organisations can actually see.

The noise added in the wrong way can either weaken privacy or make the data less useful. Leader need help understanding the trade offs inherent in differential privacy. Its earlier discussion of open challenges goes further and says broader use will require better processes both for measuring utility and for helping users work with differentially private outputs. That is the part many executives dislike. Differential privacy does not let them pretend privacy is free. It forces an argument about how much accuracy, granularity, or downstream usefulness they are willing to give up.

In other words, differential privacy did not fail commercially because it was too academic. It failed to become ubiquitous because it was too honest. It puts the privacy utility bargain in plain sight. In large institutions, that usually means politics. Product teams want fidelity. Analysts want details. Revenue teams want more segmentation. Policy teams want stronger protection. Noisy outputs are not only a technical choice. They become a budget, power, and accountability conversation. 

Epsilon is mathematically neat and managerially awkward

Differential privacy relies on parameters that matter deeply and are still difficult to explain outside specialist circles. There is still no consensus answer on what epsilon should mean in practice or how it should be set. 

That is not a minor educational problem. It is a strategy problem. A control rarely becomes ubiquitous when the key parameter cannot be translated cleanly into board-level language, regulator language, procurement language, and customer language all at once. Differential privacy is strong where the institution can tolerate technical nuance and invest in interpretation. It struggles when leaders want a simpler sentence than the truth allows.

The engineering burden is still heavier than the story suggests

Differential privacy is easy to describe and hard to implement well. That should not surprise anyone who has actually watched privacy technologies move from paper to production. The hard part is rarely just the mechanism. It is the surrounding system. Contribution limits, privacy accounting over time, query controls, data schemas, public versus private feature choices, and monitoring all need to line up. This is not plug-and-play privacy. It is a design-heavy privacy.

Also Read: Data privacy for startups: Simple steps to protect sensitive documents

It asks for governance maturity that many firms do not yet have

Differential privacy is not just a maths layer. It is a governance discipline masquerading as a technical feature.

Explaining the protections to end users and other stakeholders is difficult because the guarantees are not absolute and need contextualisation. This is one of the most under-appreciated barriers to ubiquity. Differential privacy demands that an organisation know what data is being used, what counts as a contribution, what is being released, who decides the privacy budget, how utility is evaluated, and who signs off when those choices carry consequences. Many firms still are not good at that level of definitional discipline.

That is why differential privacy often lands best in institutions that already think like stewards rather than extractors. Official statistics agencies, mature research environments, and large platforms with dedicated privacy infrastructure can absorb the overhead. A typical enterprise trying to move fast with fragmented data ownership usually cannot. The challenge is not only whether it can add noise correctly. It is whether it can define responsibility clearly enough to use the technology honestly.

It becomes politically hardest where it matters most

The utility loss from differential privacy can fall harder on underrepresented groups, both in private data summaries and in differentially private machine learning. In plain terms, the smaller or less represented the subgroup, the more likely the noise is to hurt the usefulness. That makes deployment especially delicate in precisely the settings where fairness, public accountability, or high-stakes decisions matter most.

This is one reason differential privacy remains easier to justify in some telemetry and aggregate analytics settings than in high-consequence operational systems. A technology does not become ubiquitous just because it is principled. It becomes ubiquitous when the trade-offs are politically boring. Differential privacy is not there yet. In many important contexts, it still makes the distribution of cost too visible.

Also Read: How to unlock possibilities through data privacy enhancing technologies

So what is really going on

The more strategic reading is this. Differential privacy was sold as a privacy solution, but it behaves more like a discipline of institutional restraint.

It forces organisations to answer questions they would often prefer to blur. What are we actually trying to learn? How much precision do we truly need? Who gets to decide the privacy loss? Which users bear more of the utility cost? What other controls still matter because differential privacy does not solve them? Those are healthy questions. They are also exactly the sort of questions that stop a technology from becoming frictionless and universal.

So the right conclusion is not that differential privacy is disappointing. It is that the market misunderstood what kind of success it was likely to have. Differential privacy was never going to become ubiquitous in the simplistic sense of appearing everywhere sensitive data appears. It is becoming something else. A serious control for specific settings where aggregate insight matters, formal guarantees matter, and the institution is mature enough to live with visible trade-offs. 

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post Differential privacy was supposed to solve it: Why it is not ubiquitous yet appeared first on e27.

Jiang Tao did not set out to build a music AI company. He set out to give his wife a gift. That detour, a decade in the making, has produced one of China’s most talked-about AI music platforms and a quietly ambitious global expansion play.

There is a moment in most founder origin stories where the mission and the person become indistinguishable. For Tao, founder of Shanghai and Beijing-based AI startup Initiai.on, that moment happened not in a boardroom or an accelerator cohort, but in a recording session with his daughter.

Also Read: How AI and Web3 are rewiring music’s infrastructure for a new creative economy

“I decided to train a model to generate a song for my wife as a gift,” Tao told e27 on the sidelines of BEYOND Expo 2026. “I used three years to train models that could generate melody and lyrics. And then my daughter and I sang the song together — the melody, the lyrics, all generated by the model.”

He pauses before adding, almost as a footnote: “I had never studied music before.”

What he did have was twenty years of machine learning experience, rooted in a PhD focused on speaker recognition, the science of identifying who someone is from their voice. That technical foundation, combined with a personal obsession he freely admits started as a romantic gesture, became the thesis behind Initiai.on: that AI could unlock musical self-expression for people who had never had access to it before.

From Tencent Music to founding his own model

Tao spent part of his pre-startup career at Tencent Music, one of the largest music streaming and entertainment platforms in the world, where he was part of an internal team exploring music generation. It was also around that time that he watched a pivotal case study unfold in the global AI music space.

“Have you heard of Suno, the most popular music generation company?” he asks. “At first, they did not want to generate music; they wanted to train a speech generation model. They uploaded a model named Bark to GitHub. People found that Bark could generate songs with vocals and background music. From that point, Suno turned to generate songs, not speech.”

The accidental discovery that audiences wanted AI-generated music, not just AI-generated speech, was a signal Tao took seriously. Combined with his own years of research and his Tencent experience, it gave him the conviction to go foundational: to build Initiai.on’s models from scratch rather than layer applications on top of existing APIs.

Also Read: How Wubble AI is transforming commercial music creation

“At that point, I knew I must do this full time,” he says. “It was so exciting.”

Yinchao: millions of users, and a grandfather singing to his grandchildren

Initiai.on’s flagship consumer product, Yinchao — a one-stop AI music creation and consumption platform built on the company’s self-developed large model — has crossed a user base in the millions. The platform also generated the official theme song for the 2025 World Artificial Intelligence Conference (WAIC), a signal of both the model’s technical maturity and its growing institutional credibility.

But the users Tao describes with the most evident pride are not the engineers or the institutional clients; they are the everyday people who had never thought of themselves as musicians.

“Even people without music knowledge can use this model to generate a song, to record their story,” he explains. “You record 30 seconds in our app, and we can embed your voice, and use that to generate the sounds. And many people on our app generate songs for their fathers, mothers, grandparents, grandchildren.” He smiles. “I like every day to read their stories, listen to their stories through their songs.”

The platform also serves professional musicians at the other end of the spectrum — artists with a hundred musical ideas and the budget to properly produce only a handful. Yinchao lets them generate full demos rapidly, stress-testing concepts before committing to full production.

On the devaluation question

The obvious provocation in any interview about AI music is the question of whether the technology devalues human creativity, commoditising the most intimate form of human expression. Tao does not dodge it, but he reframes it in a way that reflects both his engineering background and what sounds like genuine conviction.

“In China, people always pay for their emotions,” he says. “Some people don’t care whether the music was generated by a human or by technology. All they care is whether this music makes them happy? Does it make me want to cry? Does it satisfy my emotions?”

His second argument is economic access. Before tools like Yinchao, commissioning a custom song as a personal gift could cost upwards of US$2,000. “Only a few people could do this. Now everyone can.” He is not dismissive of human artistry, though. “I think genuine creative people can think of patterns that a model cannot generate. That is the most valuable thing in humanity.”

Going global, starting this week

For most of its existence, Initiai.on has operated primarily in the Chinese market. That is changing. At BEYOND Expo, Tao confirmed the company is launching a new product called Hitok, aimed at international markets including Australia, India, and Europe.

The monetisation model will differ by region. In Western markets, the platform operates on a credit-based system for generating music and music videos. In China, users pay for other formats of engagement and content. The model already supports multiple languages — Chinese, English, Japanese, and Korean — with more in the pipeline.

Also Read:

Southeast Asia is also on the radar. “Some partners have invited us to join Singapore,” Tao says. It is not a formal announcement, but it is not a non-answer either.

Also Read: Faster tech, slower brains: The biological blind spot of the AI race

Behind the product is a team that reflects the company’s unusual intersection of disciplines: Tsinghua-trained PhDs who specialise in GPU chip generation and also happen to be singers; professional musicians from Shanghai Conservatory of Music and Guangzhou Xinghai Conservatory, brought in specifically to evaluate whether the model’s output clears the bar a human ear would set.

It is, in its own way, a mirror of Tao himself, a machine learning veteran who spent three years training a model not because a market research report told him to, but because he wanted to give his wife a song she would remember.

The post Yinchao’s millions: AI music that lets anyone be a composer appeared first on e27.

The post Funded: SEA founders need a capital sequence, not another funding scramble appeared first on e27.