We hosted the AI for Finance Summit Tokyo on October 22 at LY Corporation in Kioi Tower. This was the first Japan edition of the series. Over 100 investors and practitioners attended. More than half of our guests were from the buy-side, sell-side, or asset owners, and two in three held director-level roles or higher. Sixteen speakers led a fireside chat, a talk on agentic AI, three panels, and two demos, all under the Chatham House Rule. Most sessions were in English with live Japanese subtitles. The buy-side panel and the talk on agentic AI were held in Japanese. Where Los Angeles put credit on stage next to equity ten days earlier, the question in Tokyo was how Japan's institutions can build an edge with AI. Most information that moves Japanese stocks is still disclosed in Japanese first. A model can now read all of it in seconds.

The discussion focused on where the investment edge goes now that a model can read everything. It also explored what work still needs a person. Six findings stood out.

The first is that reading Japanese early is no longer the edge. Companies are disclosing more information and improving their accounting. Japan is also viewed as a key part of global supply chains, so much of the demand for insight now comes from overseas. Fast traders abroad feed Japanese disclosures directly into models, accepting they might be wrong almost half the time. Meanwhile, asset managers still worry about model hallucination. The information gap is narrowing, but it is not yet closed.

Second, the edge moves to what has not been encoded. Japan was described as a high-context market, where management signals its intent through actions and omissions. A single board appointment can mark a transformation the consensus has missed. The biggest bias named was the belief that a company that has not changed in twenty years never will. Anything that can be encoded eventually will be. Taste has to develop faster.

Third, the next reader of research is an agent, and the room was split on how to respond. One side argued for structuring research so machines can read it. This means tagging the claim, the assumption, and the confidence level, and linking each to supporting evidence. The other side warned that simply feeding the machine drives down the price of research. They argued that a memorable idea is what gets paid. The machine demands structure. The human still pays for a story.

Fourth, agentic workflows are being built by hand, inside conservative budgets. With IT budgets going to the core business first, one team built its own. Each morning, agents brief the team on every name held or under review. These briefings explain how companies relate to one another and what each has done in the past. The system does work an analyst should do but rarely can, and it supports human judgment, not automated trading. Accuracy comes from running it like a human team: split the work, push calculation to tools, and send back anything that falls short.

Fifth, the first bill is a surprise, and routing is the answer. Costs ran far ahead of plan when every task was sent to the largest, most expensive model. The fix is to route each task based on performance and cost. Simple jobs go to smaller or open-weight models. Reliability is built in a sequence: market data first, where liability is low, then internal data, with client-facing use last. Getting the cost down is what lets the system scale.

Sixth, a named person signs off, but how the next generation learns judgment is an unsolved problem. A model's signal is a useful opinion, to be checked like a junior analyst's report. If an agent trades on documents obtained through a data breach, the person is held responsible, not the AI. The instinct for knowing when a number is wrong came from the experience of building models by hand. Every panel raised the junior-to-senior career path, and nobody claimed to have an answer.

Our live demo with AWS searched Japanese company disclosures in the original language. It cited every answer back to its source document and built a peer comparison on licensed data. The demo also showed the memory and managed runtime needed to run agents inside an institution.

Two tensions defined the close. The first was convergence. If every investor reads the same disclosures with similar models, their conclusions will converge. One view was that this leaves room for active managers who add first-hand research and engagement. Another was that a fully encoded market becomes very hard to beat. The second tension involved a firm's own data, most of it unstructured and untouched. One view was that the next two years will be spent tagging this data, because one wrong tag in an automated decision is costly. The other view was that intelligence will become cheap and abundant. In that world, the bottleneck is the handoff: AI output is packed into slides, only for someone else to use AI to extract the information again.

Hong Kong follows on November 5, with Milan and Singapore closing out the year.

Learn more at linqalpha.com or contact support@linqalpha.com.