Weekly

Millennium | Embedding personal AI assistants in everyday workflows

Digital Twins in use

Over 1,600

Initial pilot

Over 150 twins

Daily interaction in pilot

97%

Millennium has made Digital Twins available to all interested employees. Working through email, they retain personal context and support research, monitoring and preparation. More than 1,600 are in use; 97% of the initial user group interacted daily, demonstrating sustained engagement within that cohort.

Each twin has a separate identity and audit trail. Access is granted system by system within the employee’s entitlements, enabling personalised assistance with clearly defined permissions.

Ardian | Bringing institutional knowledge into diligence and drafting

Drafting tasks covered

Over 70% less time

Priority use cases

12

Knowledge base

~30 years of deal experience

According to implementation partner JAKALA, Ardian Buyout organised historical deal materials into a knowledge layer with citations and version records to support investment drafting and due diligence. The team tested architectures using parallel specialist agents, opposing arguments and an independent evaluator to challenge investment assumptions.

The system drafts from existing documents and deal materials, while investment professionals review, challenge and finalise the work. Implementation partner JAKALA reports over 70% less time on the drafting tasks covered, highlighting efficiency gains in that part of the workflow. Professionals retain responsibility for judgement, argument and final delivery.

Scale VP | Connecting external data with internal deal history

Example query time

~30 min → under 1 min

Connected sources

Vendor data + CRM + notes

Controls

Authentication + call logs

Scale Venture Partners built an MCP server linking PitchBook, Salesforce deal records and partner notes to a research assistant. One specific question: which co-investors in a company had previously led rounds in deals the firm passed on? The assistant cross-references investor lists, deal history and the reasons for passing.

The firm reports that this cross-system query fell from roughly 30 minutes to under one minute, illustrating retrieval efficiency after external data and internal records were connected. The system checks identity per request, applies per-user rate limits and records tool calls. It returns clearly defined structured fields, with tools designed around actual research questions.

Balyasny | Evaluate first, then let investment teams tailor agents

Evaluation dimensions

More than 12

Platform adoption

~95% of investment teams

Deployment model

Central platform + team customisation

An OpenAI customer case study describes how Balyasny evaluates models against internal data and benchmarks across more than 12 dimensions, including predictive accuracy and numerical reasoning, before production deployment. A central team builds shared components and compliance guardrails; investment teams customise agents by asset class.

The case study reports that roughly 95% of investment teams use the AI platform, indicating adoption coverage across those teams. A central team provides shared capabilities, while individual teams tailor applications to their research tasks.

Bloomberg | Delivering financial data with interpretable context

Coverage

100m+ securities

Data fields

50,000+

Access foundation

Data License Plus

On 29 September, Bloomberg launched Enterprise MCP, adding a standardised AI access layer to Data License Plus. Alongside licensed data, the interface supplies field definitions, calculation methods, currencies, price types and data timestamps, helping agents assess whether a number fits the task.

The same “price” may refer to different exchanges, currencies or points in time. Delivering those definitions with the data supports reconciliation and explanation across sources. MCP is a standard protocol for connecting AI to external data and tools; institutions can assess completeness, interpretability and ease of use against specific research tasks.

S&P Global Energy | Experts define the data; agents combine the queries

LNG data domains

7 agents

Domain experts own

Definitions + query examples

Access governance

Unity Catalog

In a case study published by Databricks, S&P Global Energy divides commodity datasets into narrowly scoped Genie agents by business domain. Domain experts maintain field descriptions, business definitions and trusted query examples. MCP then brings multiple agents together through a unified query interface.

For LNG, seven agents cover assets and contracts, cargo, tenders, outages, supply and demand, netbacks, and prices. Cross-domain questions can call the relevant data groups together. Underlying access retains Unity Catalog permissions, while benchmark questions support ongoing quality checks.

FactSet | Bringing licensed financial data into research workflows

Initial coverage

9 datasets

Beta adoption

45 firms

Beta users

800+ institutional users

FactSet’s original launch announcement describes a native MCP server offering unified access to nine datasets, including fundamentals, estimates, ownership, M&A, prices and supply chains. Institutions can call these datasets within AI workflows, reducing the need to rebuild connectors for each application.

The preceding Explorer beta involved 45 firms and more than 800 institutional users, indicating participation during testing. FactSet positions the service as governed, production-grade access, with institutions connecting content under their data entitlements. This earlier practice is shown with its original launch date.

DBS Hong Kong | Adding agent capabilities to corporate banking assistance

DBS Joy

Phased rollout from 26 Sep

digibot

Planned for 2027

Activation

Authenticated + customer-initiated

DBS Hong Kong announced a phased introduction of agent capabilities for its corporate banking assistant, DBS Joy, from 26 September. Customers can ask about payment status, transactions and fees, with the assistant retrieving and analysing relevant account information to reduce navigation across pages.

Agent capabilities for the retail assistant digibot are planned for 2027. Both rely on authenticated, customer-initiated interactions and retain access to human service. The announcement reports a 16% reduction in hotline calls from Joy’s existing service, providing context for its established service base as the new agent capabilities roll out in phases.

Shanghai | Putting AI approvals, reporting and outsourcing controls into practice

The Shanghai office of the National Financial Regulatory Administration issued 16 measures on AI applications. Generative AI used in public-facing services or high-risk scenarios must be reported to the office before launch. High-risk applications require an admission control process and approval by the institution’s risk management committee.

External compute, model and data-processing services should fall within IT outsourcing risk management. The document addresses relevant financial institutions in Shanghai. Institutions can use the original text and related provisions to incorporate scenario approvals, launch reporting and supplier management into their processes.

HKMA | Delegating to agents does not transfer management accountability

In his presentation at the HKIB Annual Banking Conference, Arthur Yuen distinguishes four AI roles: assist, recommend, decide and act. The material emphasises board and senior-management accountability for the full action pathway, including authorisation boundaries, approvals and escalation, traceability and challenge, and human intervention.

The first 2026 cohort of GenA.I. Sandbox++ lists 36 use cases, 30 financial institutions and 27 technology partners, extending exploration to end-to-end agentic workflows. This supervisory discussion material offers a framework for management accountability and agent governance that can inform authorisation and operating processes.

Fed’s Waller | Agent payments need new mechanisms for trust

In his Sibos speech, Waller distinguishes agent-assisted shopping from shopping and payment autonomously undertaken by an authorised agent. For the latter, he raises three central questions: how to prove payment authority, who bears responsibility for mistaken purchases, and how to adapt fraud detection built around human behaviour.

Speaking in a personal capacity, Waller raises questions for industry standards and infrastructure in a policy discussion on agent payments. Institutions can use these questions to identify design priorities for payment authority, responsibility and fraud detection as they explore the field.