B BROCENT Get a quote

Reference architecture · Hong Kong and Singapore

AI architecture for financial firms

How IT, data and AI fit together in a regulated financial firm of roughly 10 to 300 people, and the order in which to add AI so each step is safe before the next one starts. Choose your market and firm type, then move through the five stages to see what you already have and what comes next.

Read the picture from the bottom up. The IT foundation carries everything. Data starts with outside sources such as custodians, brokers and administrators, lands in business systems, is brought together in a data platform in your own cloud tenant, and is given one shared set of definitions before AI uses it.

AI sits at the top and is only as safe as each layer underneath it. Product names on this page are common examples, not recommendations.

The whole picture: seven layers from IT foundation to AI

Select a box to see what it is, common examples and the control that should be in place.

  • In place at the chosen stage
  • Added at this stage
  • Comes later
  • Selected: detail opens under its layer
People Who asks, decides and approves
  • Board and senior management Accountable for technology and AI risk
  • Front office Portfolio managers, analysts, traders
  • Operations and finance Processing, reconciliation, reporting
  • Risk and compliance Policies, monitoring, registers, returns
  • Clients and outside parties Receive reports, advice and filings
AI layer Added in stages, never ahead of the layers below
  • Stage 1 Assistants Claude, ChatGPT, Microsoft 365 Copilot or Gemini on business plans, behind company login
    Introduced
    Stage 1: Personal assistants
    Common examples
    Claude Team or Enterprise, ChatGPT Business or Enterprise, Microsoft 365 Copilot, Gemini for Workspace
    Control to have in place
    Business plans only; personal accounts blocked on managed devices; staff trained on what not to paste
  • Stage 1 AI administration Seats, cost, usage logs, prompt library, staff training
    Introduced
    Stage 1: Personal assistants
    Common examples
    Seat assignment from the directory, monthly usage and cost report
    Control to have in place
    One named owner for every AI platform; seats removed with leavers
  • Stage 2 Search and retrieval Finds passages in your documents, honours each user’s permissions, cites sources
    Introduced
    Stage 2: Connected knowledge
    Common examples
    Copilot over Microsoft 365, enterprise search connectors, a private document index
    Control to have in place
    Restricted folders excluded; every answer links to its source; queries logged
  • Stage 3 · later Document extraction Statements, notices, forms and agreements turned into fields, with human review
    Introduced
    Stage 3: Structured data
    Common examples
    Bank statements, fund notices, KYC packs, claim forms, contracts
    Control to have in place
    A person confirms extracted figures above a set threshold; original kept unchanged
  • Stage 3 · later Model gateway in your tenant Sensitive work calls models through your own cloud account; logged; contract confirms data is not used for training
    Introduced
    Stage 3: Structured data
    Common examples
    Azure OpenAI, Amazon Bedrock, Google Vertex AI, Alibaba Cloud Model Studio
    Control to have in place
    Region chosen deliberately; prompts and outputs logged; models listed in the AI register
  • Stage 4 · later Agents Own identity, narrow permissions, approval gates, full audit trail
    Introduced
    Stage 4: Workflow agents
    Common examples
    Reconciliation preparation, notice handling, review-pack drafting, alert triage
    Control to have in place
    A person approves every payment, filing, advice and external message

Microsoft 365 Copilot is licensed through Brocent. For the other assistants you hold the business subscription and we set it up and manage it. AI tools pricing →

Semantic layer Stage 3. One definition of each thing and how they relate
  • Client · investor
  • Legal entity
  • Account
  • Portfolio · mandate
  • Holding · position
  • Security · product
  • Fund · manager
  • Transaction
  • Counterparty
  • Document
  • Obligation · deadline
  • Limit · rule
Linked by: owns · holds · manages · advised by · evidenced by · due on · may be seen by
Data platform In your own cloud tenant
  • Stage 3 · later Ingestion Custodian and bank feeds, statement PDFs, SFTP, APIs, mailbox capture
    Introduced
    Stage 3: Structured data
    Common examples
    Daily position files, SWIFT or API feeds, shared mailbox for notices
    Control to have in place
    Every file logged with source, time and checksum
  • Stage 3 · later Raw zone Everything as received, unchanged, time-stamped
    Introduced
    Stage 3: Structured data
    Common examples
    Object storage with versioning and write-once retention
    Control to have in place
    No edits; retention matches record-keeping rules
  • Stage 3 · later Curated zone Reconciled positions, transactions, ledger and client data
    Introduced
    Stage 3: Structured data
    Common examples
    A data warehouse in your tenant
    Control to have in place
    Daily reconciliation; breaks owned by operations
  • Stage 3 · later Master data Clients, entities, accounts, securities, people: one identifier each
    Introduced
    Stage 3: Structured data
    Common examples
    Client and account master, security master, legal entity register
    Control to have in place
    One owner per master; changes approved and logged
  • Stage 2 Document index Searchable index of files, carrying each file’s access list
    Introduced
    Stage 2: Connected knowledge
    Common examples
    SharePoint or Drive index, private search index with access lists
    Control to have in place
    Index respects source permissions; rebuilt when permissions change
  • Stage 2 Catalogue Classification labels, where each figure came from, retention
    Introduced
    Stage 2: Connected knowledge
    Common examples
    Sensitivity labels, data catalogue
    Control to have in place
    Every dataset has an owner, a label and a retention period
Business systems Depends on the firm type
  • Portfolio and orders Positions, orders, pre- and post-trade checks, performance and attribution
  • Fund operations Fund accounting and NAV with the administrator, investor register, fees
  • Research Research notes, models, broker research, meeting notes
  • Investor relations Investor CRM, due-diligence questionnaires, reports and factsheets
  • Risk and compliance Mandate limits, trade surveillance, AML checks, personal dealing, returns
  • Collaboration Mail, chat and shared files, with recorded lines where required

Asset and fund managers · Where AI helps first: Draft due-diligence and RFP answers from approved text; Summarise broker research and earnings calls; Explain a portfolio against its mandate limits.

Outside sources Where most of the data starts
  • Prime brokers and custodians
  • Fund administrators
  • Execution brokers
  • Market and reference data
  • Investors and distributors
  • Regulators and exchanges
IT foundation Managed IT: the base for every layer
  • Identity Single sign-on, MFA, sign-in rules by device and location
    Introduced
    Already part of the firm today
    Common examples
    Microsoft Entra ID, Google or Okta
    Control to have in place
    MFA for every account; admin accounts separate
  • Devices Managed laptops and phones, encryption, patching
    Introduced
    Already part of the firm today
    Common examples
    Intune, Jamf
    Control to have in place
    Only managed devices reach firm data
  • Network Firewall, Wi-Fi, secure remote access
    Introduced
    Already part of the firm today
    Common examples
    Business-grade firewall, secure remote access
    Control to have in place
    Segmented guest and office networks
  • Mail and file security Anti-phishing, safe links, sharing controls
    Introduced
    Already part of the firm today
    Common examples
    Defender for Office 365 or equivalent
    Control to have in place
    External sharing restricted by label
  • Backup and recovery Mail, files and key systems, with tested restores
    Introduced
    Already part of the firm today
    Common examples
    Independent backup of Microsoft 365 or Workspace
    Control to have in place
    Restore tested at least every six months
  • Monitoring Sign-in and security logs, alerts
    Introduced
    Already part of the firm today
    Common examples
    Central log store, managed detection
    Control to have in place
    Alerts reviewed daily; logs kept for audit
  • Service desk Named engineers, on-site visits, joiners and leavers
    Introduced
    Already part of the firm today
    Common examples
    Managed IT service with agreed response times
    Control to have in place
    Access granted on day one, removed on the last day
Data moves upward: from outside sources into business systems, through the data platform, into one shared set of definitions, and only then to AI. Security controls and regulation apply to every layer.

Six kinds of financial firm, and where AI helps first

The systems differ by firm type; the layers above them do not. Typical licences are shown for orientation only.

Asset and fund managers

Hong Kong:
SFC Type 9 (asset management)
Singapore:
Capital markets services licence for fund management

Where AI helps first

  • Draft due-diligence and RFP answers from approved text
  • Summarise broker research and earnings calls
  • Explain a portfolio against its mandate limits
Business systems · Outside sources
  • Portfolio and orders: Positions, orders, pre- and post-trade checks, performance and attribution
  • Fund operations: Fund accounting and NAV with the administrator, investor register, fees
  • Research: Research notes, models, broker research, meeting notes
  • Investor relations: Investor CRM, due-diligence questionnaires, reports and factsheets
  • Risk and compliance: Mandate limits, trade surveillance, AML checks, personal dealing, returns
  • Collaboration: Mail, chat and shared files, with recorded lines where required

Outside sources: Prime brokers and custodians; Fund administrators; Execution brokers; Market and reference data; Investors and distributors; Regulators and exchanges

Family offices

Hong Kong:
Usually no SFC licence as a single family office; Type 9 if it manages assets for others
Singapore:
Qualifying single family offices are covered by a class exemption from 15 June 2026 (notify MAS, file an annual return); multi-family offices hold a capital markets services licence

Where AI helps first

  • Total exposure across every entity and bank
  • Read capital call notices and forecast cash needs
  • Find what the investment committee decided, and why
Business systems · Outside sources
  • Investments: Consolidated reporting across banks, deal pipeline, private-market commitments
  • Finance and accounting: Multi-entity ledger, consolidation, bill payment, treasury, tax packs
  • Legal and entities: Entity and trust register, contracts, committee minutes, signing
  • Risk and compliance: KYC and screening, personal dealing, risk and outsourcing registers
  • Family and philanthropy: Governance records, grants, impact measurement
  • Collaboration: Mail, chat and shared files, meeting notes

Outside sources: Private banks and custodians; Brokers; Fund managers and administrators; Market data; Legal, tax, audit, trustees; Regulators and registries

Private equity and venture capital

Hong Kong:
SFC Type 9, often with Type 1 or Type 4
Singapore:
Capital markets services licence for fund management, including the venture capital fund manager regime

Where AI helps first

  • Screen inbound decks against the investment thesis
  • Extract KPIs from portfolio company reports
  • Track side-letter obligations and deadlines
Business systems · Outside sources
  • Deal pipeline: Sourcing CRM, screening, investment committee papers
  • Due diligence: Data rooms, adviser reports, financial models
  • Portfolio monitoring: Company KPIs, valuations, board packs
  • Fund operations: Capital calls, distributions, investor portal, administrator
  • Legal and compliance: Fund agreements, side letters, investor KYC, returns
  • Collaboration: Mail, chat and shared files, meeting notes

Outside sources: Portfolio companies; Fund administrators; Investors in the fund; Advisers and data rooms; Private-market data; Regulators

Wealth managers and private banks

Hong Kong:
SFC Types 1, 4 and 9; banks are HKMA-authorised registered institutions
Singapore:
Capital markets services, financial adviser or bank licence

Where AI helps first

  • Prepare client review packs
  • Pre-check onboarding files for missing documents
  • Turn call notes into contact reports for sign-off
Business systems · Outside sources
  • Client management: Client profiles, KYC, suitability, mandates, contact reports
  • Advisory and portfolios: Model portfolios, proposals, order capture
  • Client reporting: Statements, performance reviews, fee calculation
  • Onboarding and AML: Document collection, screening, risk rating, periodic review
  • Compliance: Suitability records, communications monitoring, complaints, returns
  • Collaboration: Mail, chat and shared files, with recorded lines

Outside sources: Booking centres and custodians; Product providers; Clients and intermediaries; Market data and research; Screening data providers; Regulators

Brokers and securities firms

Hong Kong:
SFC Types 1 and 2; Types 4 and 6 for advice and corporate finance
Singapore:
Capital markets services licence for dealing in capital markets products

Where AI helps first

  • Triage trade surveillance alerts
  • Answer client service questions from approved content
  • Draft research summaries for analyst sign-off
Business systems · Outside sources
  • Trading and orders: Client orders, routing, algorithms, market access
  • Settlement and custody: Clearing, client money and assets, corporate actions, reconciliations
  • Client platforms: Web and mobile trading, statements, onboarding
  • Research and corporate finance: Published research, deal files, wall-crossing records
  • Risk and compliance: Margin, credit limits, trade surveillance, AML checks, best execution, returns
  • Collaboration: Mail, chat and shared files, with recorded lines

Outside sources: Exchanges and clearing houses; Custodians and banks; Clients; Market data vendors; Issuers and advisers; Regulators

Insurers and insurance brokers

Hong Kong:
Insurance Authority authorised insurer or licensed insurance broker company
Singapore:
MAS-licensed insurer or registered insurance broker

Where AI helps first

  • Read submissions and claim documents into fields, inside your tenant
  • Draft claims letters for handler approval
  • Check sales files for required disclosures
Business systems · Outside sources
  • Policy administration: Quotes, policies, endorsements, renewals, billing
  • Underwriting: Submissions, risk assessment, pricing, reinsurance
  • Claims: Notification, assessment, fraud checks, payment
  • Distribution: Agents, brokers, bank partners, commissions
  • Finance, actuarial, compliance: Reserving, capital, conduct monitoring, returns
  • Collaboration: Mail, chat and shared files

Outside sources: Policyholders and claimants; Agents and brokers; Reinsurers; Medical and repair networks; Market and risk data; Regulators

Follow one item through the layers

Three everyday cases. Each step shows what happens, where it sits in the picture and which control applies.

A custodian statement

  1. What happens
    A PDF or data feed arrives from the bank or custodian.
    Where in the picture
    Outside sources
    Control to have in place
    Sender and channel are on an approved list.
  2. What happens
    It is saved unchanged, with its source and time.
    Where in the picture
    Data platform: raw zone
    Control to have in place
    Nothing is edited; the original is the record.
  3. What happens
    Positions, cash and transactions are read into fields.
    Where in the picture
    AI layer: document extraction
    Control to have in place
    Runs inside your tenant; low-confidence fields are flagged.
  4. What happens
    Figures are reconciled against the portfolio or ledger system.
    Where in the picture
    Data platform: curated zone
    Control to have in place
    Breaks go to a named person in operations.
  5. What happens
    Each line is tied to its account, entity and security.
    Where in the picture
    Semantic layer
    Control to have in place
    One identifier per account and security; no duplicates.
  6. What happens
    Staff ask exposure and liquidity questions in plain language.
    Where in the picture
    AI layer: assistants and search
    Control to have in place
    The answer links back to the statement page it came from.

A client onboarding pack

  1. What happens
    Passport, proof of address and company documents come in by secure upload.
    Where in the picture
    Outside sources
    Control to have in place
    No documents by personal email or chat.
  2. What happens
    Files go to a restricted area for named staff only.
    Where in the picture
    Business systems: onboarding
    Control to have in place
    Restricted label applied on arrival.
  3. What happens
    Names, dates and ownership chain are read into the client record.
    Where in the picture
    AI layer: document extraction
    Control to have in place
    In-tenant model only; never a general assistant.
  4. What happens
    Screening runs. A compliance officer reviews and sets the risk rating.
    Where in the picture
    Business systems: compliance
    Control to have in place
    The decision is made and recorded by a person.
  5. What happens
    The client is tied to related parties, accounts and a review date.
    Where in the picture
    Semantic layer
    Control to have in place
    Access limited to the client team and compliance.
  6. What happens
    Missing documents and upcoming reviews are flagged automatically.
    Where in the picture
    AI layer: agents
    Control to have in place
    The agent prepares; compliance approves every action.

A question from a portfolio manager

  1. What happens
    “What is our total exposure to this counterparty?”
    Where in the picture
    People
    Control to have in place
    The user signs in with the company account and MFA.
  2. What happens
    The system checks who is asking and what they may see.
    Where in the picture
    IT foundation: identity
    Control to have in place
    Role decides which accounts are in scope.
  3. What happens
    “Exposure” and “counterparty” are resolved to agreed definitions.
    Where in the picture
    Semantic layer
    Control to have in place
    One definition, owned by risk, used everywhere.
  4. What happens
    The query runs on reconciled data, limited to permitted accounts.
    Where in the picture
    Data platform: curated zone
    Control to have in place
    Access rules are applied before the model sees anything.
  5. What happens
    The model drafts the answer with figures and sources.
    Where in the picture
    AI layer: model gateway
    Control to have in place
    Runs through your gateway; prompt and answer logged.
  6. What happens
    The manager checks the sources before acting on the number.
    Where in the picture
    People
    Control to have in place
    A person remains accountable for the decision.

AI adoption roadmap: five stages, each built on the one before

A stage is finished when its control is working, not when the tool is switched on.

  1. Stage 0

    Foundation in order

    What changesIdentity, devices and file permissions are cleaned up. A classification scheme is agreed.

    AI connects to · Control added
    AI connects to
    Nothing yet.
    Control added
    MFA everywhere, permission review, restricted folders identified.
    Timing
    Typically 4 to 12 weeks, depending on size and the state of permissions.
  2. Stage 1

    Personal assistants

    What changesEvery member of staff has an approved assistant on a business plan.

    AI connects to · Control added
    AI connects to
    Only what a user types or uploads.
    Integration
    Single sign-on; seats assigned from the directory.
    Control added
    AI policy, AI register, personal accounts blocked on managed devices, usage and cost report.
    Example
    Summarise a 60-page report; translate a term sheet.
  3. Stage 2

    Connected knowledge

    What changesAI can search the firm’s own documents and mail.

    AI connects to · Control added
    AI connects to
    Mail and files, with each user’s own permissions.
    Integration
    Copilot or connectors; an index that carries access lists.
    Control added
    Answers cite sources; restricted folders excluded; query log.
    Example
    “What did the committee decide on this in 2023, and why?”
  4. Stage 3

    Structured data

    What changesStatements and notices become data. Questions run over numbers as well as text.

    AI connects to · Control added
    AI connects to
    The data platform, through the semantic layer.
    Integration
    Feeds and extraction into your tenant; a model gateway you control.
    Control added
    Human check on extracted figures; a trail from every answer back to the source file.
    Example
    “Total exposure to one counterparty across all accounts.”
  5. Stage 4

    Workflow agents

    What changesAI carries out defined tasks end to end and hands over for approval.

    AI connects to · Control added
    AI connects to
    Named systems only, with its own account.
    Integration
    Agent identity, narrow API permissions, approval step, audit trail.
    Control added
    A person approves every payment, filing, advice and external message.
    Example
    Notice received, checked, payment prepared for approval.

Which data may meet AI, and when

Each firm sets its own rules; this table is where we begin the conversation.

Market and reference dataGeneral Stage 1: Personal assistants · Yes Stage 2: Connected knowledge · Yes Stage 3: Structured data · Yes Stage 4: Workflow agents · Yes
Research and internal knowledgeConfidential Stage 1: Personal assistants · With conditions Business plan only Stage 2: Connected knowledge · Yes Stage 3: Structured data · Yes Stage 4: Workflow agents · Yes
Positions and transactionsConfidential Stage 1: Personal assistants · With conditions No bulk uploads Stage 2: Connected knowledge · With conditions Summaries only Stage 3: Structured data · Yes Through your gateway Stage 4: Workflow agents · Yes With approval step
Client and investor recordsConfidential, personal Stage 1: Personal assistants · No Stage 2: Connected knowledge · With conditions Own clients only Stage 3: Structured data · With conditions Through your gateway Stage 4: Workflow agents · With conditions Approval on every output
KYC and identity documentsRestricted Stage 1: Personal assistants · No Stage 2: Connected knowledge · No Stage 3: Structured data · With conditions In your tenant, named staff Stage 4: Workflow agents · With conditions Compliance approves
Legal agreements and side lettersRestricted Stage 1: Personal assistants · With conditions Templates and public documents only Stage 2: Connected knowledge · With conditions Deal team only Stage 3: Structured data · Yes Through your gateway Stage 4: Workflow agents · With conditions Lawyer or owner approves
Inside or price-sensitive informationRestricted Stage 1: Personal assistants · No Stage 2: Connected knowledge · No Stage 3: Structured data · With conditions Case by case, compliance approves, in your tenant Stage 4: Workflow agents · No
Mail, chat and call recordsConfidential Stage 1: Personal assistants · With conditions Pasted extracts only, no client names Stage 2: Connected knowledge · Yes Own mailbox and shared sites Stage 3: Structured data · Yes Stage 4: Workflow agents · With conditions Drafts only, a person sends
Staff and family personal dataRestricted Stage 1: Personal assistants · No Stage 2: Connected knowledge · No Stage 3: Structured data · With conditions In your tenant, named staff Stage 4: Workflow agents · No

This table and the Stage 0 timing are Brocent’s default starting position for a regulated firm, not regulatory requirements.

AI readiness self-check: where are you today?

Tick what is true for your firm. Stages are sequential, so the result is the furthest stage you can reach without skipping earlier groundwork. Your ticks stay in this browser only.

Groundwork for Stage 1
Groundwork for Stage 2
Groundwork for Stage 3
Groundwork for Stage 4

Stages are sequential: a firm is ready for a stage when every item in the groups before it is true.

What the regulators in Hong Kong and Singapore expect

Instruments are listed by name so you can find them. This is an index, not legal advice: confirm the current text and whether it applies to your licence.

Technology risk

Hong Kong

  • SFC Code of Conduct and Management, Supervision and Internal Control Guidelines
  • SFC Fund Manager Code of Conduct for Type 9 firms
  • HKMA Supervisory Policy Manual TM-G-1, General Principles for Technology Risk Management (banks)

Singapore

  • MAS Notices on Technology Risk Management (binding; one for each type of institution, e.g. FSM-N05 for banks)
  • Technology Risk Management Guidelines (January 2021)
  • Proposed changes to the notices were consulted on in June and July 2026 and were not final as of November 2026

Cyber security

Hong Kong

  • SFC Guidelines for Reducing and Mitigating Hacking Risks Associated with Internet Trading (October 2017)
  • SFC report on the 2023/24 thematic cybersecurity review of licensed corporations (February 2025)
  • SFC circular of 2 June 2026 on enhanced cybersecurity measures against AI-enabled cyberattacks
  • SFC circular of 9 July 2026 asking internet brokers and virtual asset trading platforms to move to phishing-resistant sign-in
  • HKMA TM-C-1, Supervisory Approach on Cyber Risk Management (November 2024), and C-RAF 2.0 (banks)
  • Insurance Authority GL20 Guideline on Cybersecurity (authorised insurers; current version in effect from 1 January 2025)

Singapore

  • MAS Notice on Cyber Hygiene (binding; in force since August 2020)

Outsourcing and cloud

Hong Kong

  • SFC circular on use of external electronic data storage (31 October 2019): keeping regulatory records only with an outside provider needs SFC approval of the premises, a provider in Hong Kong or an undertaking from it, and two designated managers-in-charge
  • HKMA Supervisory Policy Manual SA-2, Outsourcing (banks)

Singapore

  • Guidelines on Outsourcing (Financial Institutions other than Banks), in effect since 11 December 2024
  • Banks follow MAS Notice 658 and the Guidelines on Outsourcing (Banks)

Continuity and resilience

Hong Kong

  • SFC report and circular on operational resilience and remote working (October 2021)
  • HKMA Supervisory Policy Manual OR-2, Operational Resilience, and TM-G-2, Business Continuity Planning (May 2022, banks)

Singapore

  • MAS Guidelines on Business Continuity Management (June 2022)

AI specifically

Hong Kong

  • SFC circular to licensed corporations on use of generative AI language models (12 November 2024)
  • HKMA circular on consumer protection in respect of use of generative AI (19 August 2024) and its 2019 high-level principles on AI (banks)
  • No AI-specific Insurance Authority guideline published as of November 2026

Singapore

  • MAS Guidelines on Artificial Intelligence (AI) Risk Management, issued 7 October 2026: oversight, identifying AI use and the AI inventory from 7 October 2027; lifecycle controls and infrastructure by 7 October 2028
  • MAS FEAT principles on fairness, ethics, accountability and transparency (November 2018)

Personal data

Hong Kong

  • Personal Data (Privacy) Ordinance
  • PCPD Artificial Intelligence: Model Personal Data Protection Framework (June 2024)
  • PCPD checklist on guidelines for the use of generative AI by employees (March 2025)

Singapore

  • Personal Data Protection Act
  • PDPC Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems (March 2024; they do not cover generative AI)

Financial crime

Hong Kong

  • Anti-Money Laundering and Counter-Terrorist Financing Ordinance (Cap. 615)
  • SFC and HKMA AML / CFT guidelines

Singapore

  • MAS AML / CFT notices (revised set in effect from 1 July 2025)

Books and records

Hong Kong

  • Securities and Futures (Keeping of Records) Rules (Cap. 571O)

Singapore

  • Record-keeping duties under the Securities and Futures Act and its regulations

Family offices

Hong Kong

  • No licensing regime specific to family offices; a single family office that manages only its own family’s assets generally does not need a Type 9 licence, though other activities can require one

Singapore

  • Revised framework for single family offices, in effect from 15 June 2026: qualifying single family offices are covered by a class exemption, notify MAS and file an annual return

Five ways AI is joined to your systems

Each stage uses the patterns introduced before it.

  1. Used from: Stage 1

    Identity federation

    Staff sign in to every AI tool with the company account; seats follow joiners and leavers.

    Why it matters: Fewer orphan accounts; access is removed with the leaver process.

  2. Used from: Stage 2

    Delegated connectors

    The assistant reads Microsoft 365 or another system on behalf of the signed-in user.

    Why it matters: AI sees only what the person asking may already see, so permissions must be right first.

  3. Used from: Stage 2

    Permission-aware retrieval

    Documents are indexed with their access lists; only permitted passages reach the model.

    Why it matters: Answers come with sources that can be checked.

  4. Used from: Stage 3

    Model gateway in your tenant

    Sensitive work calls models through your own Azure, AWS or Alibaba Cloud account, with logging.

    Why it matters: You choose where data is processed and keep the record.

  5. Used from: Stage 4

    Agent with its own identity

    A named account, the minimum permissions for one task, and an approval step.

    Why it matters: Every action is attributable; a person stays accountable.

How Brocent works on this

We put the foundation first because AI is only as safe as the layers under it. For firms of this size we take over the whole IT function: service desk, Microsoft 365 administration, laptops and phones, joiners and leavers, network and hardware, with named engineers and on-site visits in Hong Kong and Singapore.

We work with Claude, ChatGPT, Microsoft 365 Copilot and Gemini, and with DeepSeek, Qwen and Kimi. Microsoft 365 Copilot is licensed through Brocent. For the other platforms you hold the business subscription and we set it up and manage it for you.

We use AI in our own operations: on the service desk, in sales support, in sourcing and in accounts payable. Each AI worker has its own identity and only the permissions its role needs, everything it does is logged, and anything sent to a client or supplier is approved by a person first. The controls on this page are the ones we run ourselves.

In financial services we run managed IT for a global financial services firm across Beijing and Hong Kong, a global insurance group across three cities in China and a global investment bank in Beijing. The case studies are anonymised.

Financial services case studies →

Questions financial firms ask about AI

Which AI stage can a regulated financial firm start with?

Most firms start at Stage 0: MFA on every account, managed devices and a review of file permissions. Approved assistants on business plans come next. A firm moves on when the controls of the stage before are working, not when a tool is switched on.

Can staff use ChatGPT or Claude with client data?

Not at Stage 1 in our default position: client and KYC records stay out of general assistants. From Stage 2 a person may search the documents of their own clients, and from Stage 3 client data is processed only through a model gateway in the firm’s own cloud tenant. Each firm sets its own rule.

What does the SFC circular on generative AI expect?

The SFC circular of 12 November 2024 sets four principles for licensed corporations: senior management responsibility, AI model risk management, cybersecurity and data risk, and management of third-party providers. It treats AI used for investment recommendations, advice or research as high-risk.

When do the MAS AI risk management guidelines apply?

MAS issued its Guidelines on Artificial Intelligence (AI) Risk Management on 7 October 2026. Oversight, identifying AI use and keeping an AI inventory are expected from 7 October 2027, and lifecycle controls and infrastructure by 7 October 2028.

How long does Stage 0 take?

Typically 4 to 12 weeks, depending on the size of the firm and the state of its permissions. This is Brocent’s working estimate, not a regulatory requirement.

Does Brocent sell Claude or ChatGPT?

Microsoft 365 Copilot is licensed through Brocent. For Claude, ChatGPT, Gemini and the other platforms you hold the business subscription and we set it up and manage it. Prices are on our AI tools pricing page.

What we confirm in an assessment

  • Which core systems are in use, and which offer feeds or APIs
  • Number of entities, booking centres, custodians and outsourced providers
  • Microsoft 365 or Google Workspace licence tier and the current permission structure
  • Which AI subscriptions already exist, who pays, and what staff use informally
  • Which data must never reach an external model, and where data must reside
  • The current outsourcing register, technology risk assessment and AI inventory

Basis for this picture

Regulatory instruments are named from regulator publications checked in November 2026. The Stage 0 timing and the data table are Brocent working defaults, not regulatory requirements. Nothing on this page is legal advice.

Sources: MAS: AI risk management guidelines (media release) · MAS: technology risk and cyber hygiene · MAS: outsourcing guidelines (non-banks) · MAS: single family office framework · SFC: thematic cybersecurity review · SFC generative AI circular (Clifford Chance summary) · HKMA: TM-C-1 cyber risk module · Insurance Authority: GL20 · PCPD: AI resources

Find out which stage your firm can start at

Tell us your firm type, market and the systems you run. We start with a short assessment of your Microsoft 365 and current AI use, then propose a fixed monthly scope.

Book an AI readiness assessment