The AI Test for GPs: Why Judgment, Access and Hard Moats Still Decide Returns

At almost every session of this year’s DealStreetAsia Asia PE-VC Summit in Singapore, the conversation eventually turned to AI. Much of it was about investing in AI companies. The more consequential question for our industry is a different one: how AI changes the way general partners (GPs) themselves source, underwrite and hold investments — and what limited partners (LPs) should expect to see as proof.

That was the brief for our panel on 24 September, “The AI test for GPs: Underwriting disruption, transforming portfolios and proving value to LPs”, where I joined Yi Pan of Neuberger, Charles Allen of TPG and Xuan Ye of Collyer Capital, moderated by DealStreetAsia founder and editor-in-chief Joji Thomas Philip.

My view, sharpened by that discussion, is simple. Basic AI will become a commodity tool — table stakes for every manager. Durable advantage will depend on what competitors cannot obtain simply by using the same model: proprietary knowledge, trusted relationships, protected rights, physical capabilities and strong teams. And for material investment decisions, a human must still approve the commitment and answer for the outcome.

1. Give a rival the same AI

As growth-stage investors, we enter every deal with the exit in mind. My starting assumption now is that competitors will have access to broadly similar AI. So the underwriting question becomes: what advantages would a rival still need to build, acquire or earn over our holding period — and how long would that take, what would it cost, and can the company keep improving while others catch up?

For years the mantra was that software eats the world. I think the paradigm has shifted: it is software and hardware that eat the world, and the most defensible businesses bind the two together. In healthcare, that means testing advantages an algorithm cannot shortcut:

• Full-stack control. Does the company control the chain from design and manufacturing through assembly, packaging, branding and distribution? When those stages are connected, performance, cost and customer feedback can reinforce one another.
• Resilient supply. Are there multiple qualified suppliers that can switch in practice? Two suppliers that depend on the same upstream source may still share a single point of failure.
• Clinical switching costs. Will a surgeon abandon a trusted device — often a small fraction of the cost of the overall procedure — for a new entrant with no track record?
• Rights and regulatory moats. Even a comparable product must still generate clinical evidence and earn regulatory approval in each market it enters. Licences, defensible IP and exclusive rights can take years to replicate — they are not software sprints.

Each advantage needs testing, not assuming. I would give a rival the same AI, assume cheaper substitutes appear, and examine what happens to pricing and retention. If AI lowers the cost of entry, new competitors can arrive faster and valuation multiples can compress even while a company keeps growing revenue. A valuable business can still be a poor investment at the wrong price.

The same discipline applies to AI-driven productivity claims inside portfolio companies. Consider a purely hypothetical example:

Hypothetical figures for illustration only; not AGP or portfolio-company results.

Measurable gains are possible — a large study of 5,172 customer-support agents published in The Quarterly Journal of Economics found AI assistance raised issues resolved per hour by 15% on average, with wide variation across workers. But the investment question is how much of the improvement the business actually retains once competitors and customers respond.

2. Keep the thesis alive during the hold

Because AI can shift a market within a single holding period, we make the key assumptions explicit at entry and monitor the evidence that could invalidate them. If customers start replacing a paid feature with a bundled alternative, the growth and margin case needs another look. We agree material triggers in advance and revisit product investment, follow-on capital or exit plans when they are breached.

We also encourage portfolio companies to extend their first-mover advantage into adjacent products and services — a flywheel that compounds the moat while competitors are still clearing the hurdles above. But the decision cadence should match the business. Not every model release justifies a new valuation.

3. Screen for the inflection point, overnight

Growth investing is ultimately about identifying the moment a company hits its inflection point. This is where AI has changed our daily workflow most visibly.

Every night, an on-premise AI system works through a large universe of potential leads and filters for companies showing the revenue signals that fit our mandate. By morning, the team has a short, manageable list to investigate and narrow further. Work that once required a sizeable bench of analysts screening manually can now be run by a small number of AI-native team members who build and maintain the pipeline.

That shapes how we hire. Rather than scaling headcount in step with deal flow, we look for driven, AI-native juniors with STEM backgrounds — engineering, mathematics, physics, biology, chemistry — who are comfortable running open-weight models on local, secure hardware. Our ambition is for a lean team to operate with the reach of a far larger one.

4. Designing an AI-native GP: human in the loop

We founded August Global Partners in late 2023, just as capable models were emerging, which gave us the opportunity to think about the firm with AI in mind from the start. My own background made this natural: my PhD was in semiconductors, and in the mid-1990s I was designing early neuromorphic chips — brain-inspired hardware that was decades ahead of the computing power and data needed to make it useful. Watching that idea finally become practical struck a chord.

If I were building an investment firm from scratch today, I would design it around four principles:

• A controlled knowledge system. Connect original evidence, investment assumptions, dissent, decisions and outcomes, so the firm learns from every deal — including the ones it passed on or got wrong. Today’s open-weight models are increasingly capable of much of the routine analytical work our industry does, and deploying them on-premise, with segmented access for sensitive material, keeps proprietary data inside the firm. But private deployment protects an asset; it is not automatically the asset. The moat is valuable knowledge and its disciplined use.
• Juniors who argue with the machine. Analysts frame hypotheses, put them to the model, and challenge its recommendations before anything reaches the investment committee. When the committee pushes back — and it should — the team returns with evidence. The best outcome is a junior who can use AI to show a senior partner, convincingly, where they have got it wrong.
• Talent grooming and succession by design. Research and first-draft work will change substantially, so apprenticeship must be rebuilt around judgment, founder exposure and responsibility. Senior investors need incentives to teach their practical know-how, so that the strengths of key people become a capability the firm retains as individuals move on.
• Human in the loop. The deal lead owns the recommendation; the investment committee owns the approval of capital and position size. AI informs and challenges; it does not own the decision. If we invest well, LPs reward us with their capital. If we invest badly, the accountability is ours alone.

That view was broadly shared across the panel. AI today functions as decision support that sharpens investment-committee debate, not as a vote in the room. Governance matters as much as capability: firms should be clear about which models, agents and datasets informed any given analysis, so that no one mistakes where human judgment begins.

5. What LPs should test — and what AI cannot collect

LPs are becoming AI-enabled too, and I welcome scrutiny that helps them understand a manager’s real sources of advantage. For portfolio companies that have listed on public markets, AI tools can surface developments almost instantly — and LPs do call us when they see them. That is healthy. Consistent, analysable data helps: the industry’s ILPA Performance Template, for instance, standardises performance metrics and the underlying cash flows, and its guidance calls for Excel or another analysable digital format.

The risk is treating the most complete dataset or the most polished explanation as evidence of the best judgment. A polished AI-generated narrative does not establish repeatable skill. Small samples, unrealised valuations and changing market conditions make skill hard to isolate. LPs should test access, judgment, post-investment contribution, team depth and succession against real decisions and outcomes — including the failures.

Above all, the value of a specialist GP lies in what AI cannot collect: the conversations around a board meeting, the founder relationships built over years, the sector networks that surface a deal before it is marketed. As the panel repeatedly noted, AI is only as good as the data it is fed — and in Asia, the most valuable data is rarely public. That proprietary knowledge drives a chain no model can replicate on its own:

• Judgment — conviction that a particular company is worth backing;
• Access — the relationships needed to secure an allocation in a competitive round;
• Execution — the sector knowledge to help the company create value once invested.

When all three come together, the returns show up in the data. That is the track record LPs ultimately underwrite.

The real test

Five years from now, basic AI will be table stakes. The moat will be proprietary knowledge, relationships, rights, physical capabilities and teams that competitors cannot obtain with the same model. The AI test for GPs is whether a manager uses these tools to see inflection points sooner, underwrite durable moats more rigorously and free its people for judgment and relationships — while remaining unambiguous about who is accountable for every commitment of capital.

AI informs the decision. People approve it, and answer for it.