Is APAC behind the US on AI adoption?

In the US, AI is increasingly embedded directly into research, diligence, RFP, and document-review workflows. Across many European conversations, usage still appears more assistant-led, something you query rather than infrastructure built into the process. APAC sits between the two.

That distinction matters more than any adoption percentage.

What is actually stopping firms from expanding their use of AI?

In a recent DiligenceVault poll of asset manager RFP and DDQ professionals and allocators, incorrect or hallucinated answers ranked as the biggest barrier to broader AI adoption. Difficulty validating sources, inconsistent outputs, and data privacy followed closely. Cost ranked materially lower.

Source: DiligenceVault industry poll.

That ordering is important. The obstacle to enterprise AI is not whether the technology can generate an answer. It is whether people can trust, validate, and govern that answer once it is generated.

We saw the same pattern in two workflows where AI is already creating measurable value.

In RFP and DDQ response, generating a first draft from approved content is the easy part. The harder problem is knowing which information is current, where it came from, who owns it, and how it gets reviewed.

In fund research, the value is not another auto-generated memo. It is arriving at a meeting with the right context already assembled, so the analyst spends more time on judgment and less time on retrieval.

Putting an AI interface on top of existing data was never going to be enough. What makes the difference is the data underneath, the workflow it is embedded in, and the review built around it.

Anthropic's own work on self-service data analytics reinforces that point. Feeding Claude more historical context, including prior queries, notebooks, and dashboards, moved accuracy by less than a point. Adding human-written skills, a semantic layer, and human sign-off moved accuracy from 21% to 95%. The same system lost 30 points of accuracy after one month without upkeep.

Source: Anthropic, “How Anthropic Enables Self-Service Data Analytics with Claude” (June 2026).

Someone still has to define what good looks like, structure the information, and maintain the system over time.

AI capability has a maintenance burden, just like any other production system.

Why does distribution matter as much as model quality?

Singapore and Japan made this particularly clear.

Microsoft Copilot came up repeatedly, not because teams necessarily considered it the strongest model, but because it was already embedded in the environment they used every day, connected to Microsoft 365, Power Automate, and the broader stack.

As model quality continues to converge, ease of access and fit into existing systems become more important.

Japan: R&D, model capability, and error rates

Japan stood out for the amount of R&D attention going into AI itself.

We heard firms discussing investments in their own model capability alongside commercial LLMs, reflecting a broader desire for greater control over language, data, performance, and deployment.

The conversation was less about choosing one external model and more about building the right model stack for the organization.

Just as interesting was the focus on measurement.

One firm compared AI errors to a manufacturing defect rate, something to measure continuously and systematically drive down.

As AI moves into higher-consequence workflows, that error rate number may matter more than the adoption rates.

Model selection is becoming an architecture decision

The US conversations brought another issue into focus: economics.

Teams running AI across high-volume workflows are beginning to hit consumption limits and actively manage model usage and cost. That came up much less frequently in APAC.

Independent benchmarks show why cost is becoming a design decision. On some comparable tasks, the highest-cost models can run roughly 60x the price of the lowest-cost options.

Source: Artificial Analysis Intelligence Index.

That does not mean the cheapest model is the right model.

It means firms should stop treating model selection as a single enterprise-wide choice.

Research, extraction, classification, and routine drafting may require one level of capability. Complex reasoning and higher-risk investment analysis may require another.

Routing the right work to the right model, at the right cost and with the right oversight, is becoming part of the architecture.

What is the most valuable dataset an investment firm can build?

Access to capable models is becoming broad and increasingly commoditized. Proprietary organizational data is not.

For investment firms, that includes prior diligence, manager interactions, risk assessments, meeting notes, and the reasoning behind past decisions.

Private markets make the challenge especially visible. Unlike public-market data, private-market information remains fragmented across PDFs, decks, financial statements, side letters, and manager-specific reporting formats.

AI can only reason reliably across information that has been identified, structured, reconciled, and maintained.

The teams furthest ahead were not simply asking AI to summarize a portfolio. They wanted systems that could retain the context behind a decision: why a position was sized the way it was, what the team's conviction was at the time, which risks were knowingly accepted, and what eventually changed the view.

That creates a governance question worth asking directly:

Most firms are not there yet, but know they need to be.

Over time, the record of what a firm believed, what it worried about, and why it made a decision may become one of the most valuable datasets an investment organization owns.

The bottom line

Four takeaways from our time across Asia Pacific.

The technology is becoming similar everywhere.

Where it sits in the workday, what information it can reach, and how much a firm trusts it to operate with limited supervision are not.

We will continue this research across markets through the rest of the year. If your team wants to compare notes on where AI sits in your own workflow, or see the full DiligenceVault poll results, we would be happy to share what we are seeing.