Weekly

Nasdaq Calypso: A controlled agent environment for the trade lifecycle

Integration

MCP-based integration layer, including clients' own AI systems

Accessible records

Core trade, risk and collateral records

Initial capability

Natural-language assistant for platform data and documentation

Guardrails

Operational limits, real-time monitoring, strict sandboxing and no external data storage

Nasdaq has introduced an agent framework within Calypso for banks, brokers, asset managers and central banks that use the platform for front-, middle- and back-office trading, risk and collateral management. Institutions can use Nasdaq-provided agents or connect their own internal AI systems through the same MCP-based integration layer.

The sequence matters: the first capability is an assistant for querying platform data and documentation, rather than automated order placement. Nasdaq also says it will work with each client to review internal AI governance and control requirements before enabling these capabilities.

For institutions to adopt AI agents at scale, they need trusted, governed infrastructure within the capital-markets platforms they already operate.

Magnus Haglind, Head of Capital Markets Technology, Nasdaq (translated from the supplied Chinese edition)

Crunchbase AI: Private-market data enters agent workflows through MCP

Company coverage

More than 6 million companies, according to Crunchbase

Access channels

ChatGPT app directory, Claude, Perplexity and any MCP-compatible model

Traceability

Every result identifies its data source

Crunchbase has launched Crunchbase AI, a headless data layer that exposes funding information, company profiles, insights and predictions to agents and large language models through MCP. Its proposed uses are specific: weekly source-linked briefings on target companies, complete company lists screened against an investment theme expressed in a sentence, and tracking investor activity and round sizes within a sector.

For buy-side teams, the point is not simply an additional data source, but the ability to trace each answer back to the data. This is a company announcement; metrics such as its claimed 148% advantage in seed-round coverage have not been independently verified.

Blockworks MCP: Keeping data current for exchanges, brokers and banks

Data coverage

Crypto assets: prices, volume, unlocks, funding, governance votes and disclosures

Illustrative institutional uses

Listing-review packs, stablecoin briefings, tokenized-equity tracking and KYB counterparty profiles

Division of responsibilities

Clients retain the models, prompts and interfaces; the provider keeps the data current

Blockworks is making its datasets available to agents such as ChatGPT and Claude. Its article outlines six use cases for financial institutions. Those closest to traditional finance include daily stablecoin-supply briefings by issuer and chain, counterparty profiles for KYB teams before account opening, and tracking supply, holders and DEX volume for each tokenized stock alongside the underlying equity price.

This is a provider's product article, and the scenarios are proposed uses rather than demonstrated client outcomes. Its relevance lies in the division of responsibilities: the client owns the model and interface, while the provider maintains data freshness. That boundary is central to evaluating an external MCP data source.

Morningstar Direct AI: Three agents for three asset-management workflows

Availability

Launched for users globally

Initial agents

Product development, distribution and manager research

Data foundation

Morningstar data, analytics, research and ratings; also available to Claude, Microsoft Copilot and ChatGPT

Morningstar has rebuilt its flagship Morningstar Direct platform as Direct AI, with a browser-based interface and three agents for asset managers. The product-development agent combines flows and category data to explore where demand is strong and competition is limited, or how long comparable products took to reach $1 billion. The distribution agent helps sales teams identify gaps in model portfolios, illustrate the effect of adding a fund and prepare talking points for review meetings. The manager-research agent supports due diligence, examining changes in a fund's trading behavior and the market conditions that have helped or hurt performance.

These agents map to established product, distribution and due-diligence workflows rather than generic chat. Morningstar says they incorporate its analysts' evaluation frameworks. This is a vendor news release; statements such as its reference to more than 1,100 researchers are company-reported and have not been independently verified.

AlphaSense SuperAnalyst: From answering questions to monitoring assumptions

Deliverables

Investment memos, financial models, comparison tables, monitoring dashboards and board materials

Data foundation

Financial statements, filings, earnings calls, broker research, channel checks and 300,000 expert interviews

Key feature

Continues monitoring thesis assumptions and updates the analysis as markets change

AlphaSense describes SuperAnalyst as a persistent agent. Users supply an objective; the agent plans the work, finds and cross-checks evidence, calculates, writes and runs code, and places its output in documents that update over time. The announcement names portfolio monitoring, M&A due diligence and market assessment as use cases.

The release quotes two buy-side users. An investment director at Aberdeen Investments describes valuation models with consistent data sourcing and a clear structure. The research director at Talos Eurisko says it helps him decide within minutes where to spend his time, rather than replacing the analyst. Both are customer testimonials included in the vendor's announcement.

It does not replace the analyst's work; it lets me cover more ground in less time and with greater confidence.

Ian Lieberman, Partner and Director of Research, Talos Eurisko Asset Management (vendor-release testimonial; translated from the supplied Chinese edition)

Jump Trading: Long-running tasks for agents, final acceptance for people

Model

GPT-6 Astra, according to the OpenAI case study

Task duration

Can run continuously for several days

Acceptance principle

Treat outputs as signals: often informative, but potentially wrong

Lucas Baker, who leads LLM R&D at Jump, describes a workflow in which researchers define the problem, environment and evaluation criteria. One or more agents then run for extended periods, analyze their findings, assess them against the original proposal's criteria and adjust direction without waiting for a human review at every iteration.

The risk controls are described more concretely than the efficiency gains: agents work in a safe environment with freedom to produce results, followed by human review and acceptance. Their trading signals are scoped and reviewed like other signals before entering a tightly controlled execution environment. This is an OpenAI customer case study rather than an independent disclosure by Jump, and it provides no investment-performance or return figures.

If agents can freely produce results, but human review and acceptance remain at the end, we have confidence in the process.

Lucas Baker, Head of LLM R&D, Jump Trading (translated from the supplied Chinese edition)

Man Group: Cross-functional agent workflows with task-based permissions

Infrastructure

Proprietary multi-model harness launched in September, connecting models, tools, memory and execution environments

Team

About 450 engineers and more than 200 quantitative specialists

Permissions

Teams can build agents; access durations range from days to indefinite, depending on the task

Partnership

Anthropic partnership announced in February 2026

Man Group said in its half-year results that it would expand cross-functional agent workflows across the firm. At a Bloomberg summit on September 29, CFO and COO Antoine Forterre described investment research, quantitative research and operations teams redesigning their workflows. Operations teams are testing agent-to-agent processes with external administrators and custodians.

He gave two examples. A fund manager researching European mid-cap equities turned five years of decision notes into an agent workflow that retrieves structured and unstructured data and suggests questions for company management. Forterre said the manager's stock coverage had doubled, while noting that the sample was too small to attribute the result entirely to AI. A credit team used an internally built tool to identify changes in credit-facility terms hidden in annual-report footnotes. He emphasized governance, training and access controls under the firm's Gen AI policy as essential guardrails.

These are management statements made at a conference. The report provides no specific performance or cost figures, and the original company results announcement has not been directly checked for this edition.

BBVA: From individual use cases to a shared framework, The Frame

Scale

More than 80 million customers and 4,000 data specialists

Employee adoption

75% use AI regularly, saving an average of 2.4 hours per week

Development efficiency

Developer productivity up approximately 28%

Service and complaints

Assistants initially filter over half of in-app relationship-manager conversations; complaint-registration time down nearly 80%

Voice agents

Resolve about 90% of interactions in pilots in Peru and Mexico

At a presentation to analysts and investors, BBVA introduced The Frame: common standards for creating, deploying, managing and governing AI agents. Having focused on specific use cases over the past year, the bank now plans to replicate those capabilities across the group. BBVA also said Blue, a multimodal assistant for personal and business customers, is planned for a phased rollout in 2027.

All figures are BBVA-reported. The 90% figure for voice agents comes from production pilots and does not represent performance after a full rollout.

Barclays: Claude expands from knowledge assistance to engineering

Claude Code coverage

50% of developers by end-2026; a majority in 2027

Knowledge assistant

More than 16,000 employees and over 1 million searches processed

Global Markets

Classifies and routes approximately 120,000 client emails daily

Barclays is expanding its partnership with Anthropic to use Claude for software development, legacy-system modernization and operational efficiency. Two deployed examples are a retrieval-augmented employee knowledge assistant, launched in 2025 to support staff serving more than 20 million UK retail customers, and a Global Markets workflow that classifies incoming messages, enriches information and determines routing.

This is Anthropic's official partnership announcement. The reported figures describe usage rather than financial returns. Barclays emphasizes governance, security controls and human oversight.

ABN AMRO: Three evaluation questions before a multi-step agent pilot

Partner

Wonderful, Amsterdam: $550 million Series C completed September 2 at a $5 billion valuation

Pilot status

In preparation; specific workflows and start and end dates have not been disclosed

ABN AMRO's innovation lead Yorick Naeff outlined three questions on LinkedIn: can agents genuinely reduce repetitive work, can their actions be traced while people remain in control, and what would be needed to scale a successful result?

This case offers a contrast: the previous two banks discuss scale, while this one sets evaluation criteria and has not reported results. It is presented as an example of how a bank defines an agent pilot's acceptance criteria, not as evidence of proven impact.

DNB: Agent gains in KYC and coding accompany restructuring and around 400 FTE reductions

Division

Technology & Services

Workforce change

Approximately 400 full-time equivalents (FTEs)

Areas reporting gains

Customer data and KYC; technology development and coding

Timing

Reductions in Q4 2026; full cost effects from Q2 2027

Norway's largest bank, DNB, says it has adopted agentic AI in several parts of the group to handle tasks previously performed manually. It reports considerable efficiency improvements in customer-data and KYC work, technology development and coding. The bank will therefore adjust the organization and skills mix of its Technology & Services division, reducing approximately 400 full-time-equivalent positions.

The announcement says the bank will consult employee representatives and follow applicable procedures, with further financial effects to be disclosed during the year. It does not quantify AI-related savings or identify the specific roles being replaced.

AI is changing how we work and serve our customers. We are already seeing substantial benefits, so we need to adapt our organization to this new reality.

Kjerstin Braathen, Group CEO, DNB (announcement excerpt; translated from the supplied Chinese edition)

Goldman Sachs and MAS: New bankers manage agents from day one

According to The Straits Times, Kevin Sneader, who heads Goldman Sachs in Asia Pacific excluding Japan, told the Milken Institute Asia Summit that young employees now begin their careers managing agents and need to make effective use of this virtual workforce. He argued that management responsibilities are moving from middle managers to the front line. Reorganizing existing middle-management roles is a generational challenge across industries, and Goldman Sachs does not yet know how it will unfold.

MAS Managing Director Chia Der Jiun described a similar shift: operations employees need to become managers and supervisors of AI agents, while banks will need fewer graduates for analytical and preparatory tasks. He said Singapore is training bank employees in AI skills and working with universities.

These are public remarks, not disclosed operating data or formal plans. This item is based on The Straits Times report; the original conference record has not been checked for this edition.

MAS: AI risk guidance sets expectations for accountability, inventories and risk-based controls

Scope

All financial institutions and all forms of AI

Effective date

October 7, 2027; phased implementation of sections 5 and 6 permitted until October 7, 2028

Approach

Principles-based and proportionate to risk, rather than one-size-fits-all

MAS sets out four core expectations in its news release. First, boards and senior management should effectively oversee AI risk, with clear responsibilities, risk appetite and policies. Adequate existing governance structures can be used without creating a dedicated AI committee. Second, institutions should identify and manage risks throughout the AI lifecycle: inventory AI use, assess each use case's materiality and apply proportionate controls for data governance, testing, human oversight, cybersecurity, monitoring and change management. Third, institutions remain responsible for third-party AI and should be able to restrict, suspend or replace it where necessary. Fourth, controls should be proportionate: basic policies and procedures may suffice where poor performance or unavailability would not have a material impact.

The release also says institutions should regularly revisit controls as autonomous, tool-using agentic AI becomes more common. MAS plans to consult the industry on supplementary agentic-AI guidance in 2027. Deputy Managing Director Ho Hern Shin says clearer regulatory expectations can give institutions the confidence to innovate.

This item is based on the MAS news release. Refer to the MAS website for the full guidelines and the accompanying response to feedback.

Hong Kong: Reviewing whether existing laws cover AI-related harm

At a digital-infrastructure seminar during the AIIB Annual Meeting, Financial Secretary Paul Chan said the Hong Kong government would review whether existing laws adequately address liability for harm caused by AI. He emphasized room for innovation alongside clarity on responsibility and remedies for affected people. He also said the government would work with industry and professionals on sector-specific guidance and governance frameworks, with safeguards proportionate to application risk.

He also referred to a US$500 million AI subsidy scheme for eligible research institutions, start-ups and enterprises, and plans for a large data-facility cluster at Sandy Ridge in the Northern Metropolis. The full speech is available in the Hong Kong government news release.

Singapore: Studying stronger safeguards for high-risk AI uses

In a written parliamentary reply, Minister for Digital Development and Information Josephine Teo said Singapore is studying whether high-risk AI uses need additional safeguards, potentially including stricter testing, independently verifiable safety evidence, tighter deployment controls and stronger regulation. She emphasized a layered, risk-based approach: before deploying agents, assess the systems and data they can access, their permitted actions and the consequences of mistakes.

She also said frontier-model developers are not currently required to provide model access. These are areas of policy study, not new rules. This item is based on CNA's report; the parliamentary reply has not been directly checked for this edition.

ACRA: Practical guidance on responsible AI use in audit

ACRA's Audit Practice Guidance No. 1 of 2026 offers practical considerations for auditors and firms in five areas: integrity, professional judgment, accountability, transparency, and data confidentiality and security. It uses real incidents to illustrate how problems arise and includes reflective questions for self-assessment. Recognizing that firms adopt AI in different ways and to different degrees, the guidance focuses on practices they can apply in daily work.

This is not a rule for financial institutions, but banks and asset managers are among the organizations being audited. How auditors treat AI-generated working papers may indirectly affect institutions' expectations for documenting AI output.