I wrote the AIRG while leading AI risk supervision at the Monetary Authority of Singapore, after almost two decades in risk management and writing regulations. More recently, I co-authored a paper on runtime governance for agentic AI and helped a little with MAS's Safeguards for Agentic Finance at Runtime (SAFR) paper, the white paper on safeguarding agentic finance at runtime.

This executive guide is for the people who own strategy and the adoption decision - board directors and senior management. Who also have to set the appetite, approve the spend, and answer for the outcome. It works through the decisions you may have to make, and what the AIRG and the Project MindForge toolkit cover.

Governance and adoption are not in tension. Rapid, useful adoption of AI is what good governance supports: you move faster, not slower, when there's greater trust. Make adoption responsible but in a way that can scale and deliver value.

Governance lifts the return on AI as use cases grow; without the capacity to govern, the return plateaus early.

How it fits together in Singapore. The FEAT principles, fairness, ethics, accountability and transparency, are the underlying principles. The AIRG is the expectations built on them. MindForge is the practice, based on and aligned to the AIRG. Principles, then expectations, then practice.

FEAT sets the principles, the AIRG the expectations, MindForge the practice.

The diagrams throughout are mine, drawn for this guide. Each one carries its copyright embedded in the file; please reuse them only with clear attribution.

I keep this openly available rather than behind a sign-up. The people who have to make decisions about AI should understand it well, and the system works better when they do. Use it, and pass it on. And if you would like to follow me for more of such stuff, come find me on my Substack.

These are a starting set of the decisions and questions a board and senior management have to think about, not a fixed or complete list, and I expect it to grow and change over time. Each topic opens with a question you may recognise from your own meetings, and each one links to the fuller write-ups on this site. Use the table of contents below to jump to a decision, or step through them in order. The companion guide for the people who oversee and challenge AI, the second and third line, is here.

I am keeping this as a live document, so drop me a note (at gary@quaintitative.com) anytime if you have a question, want to add or correct something, or think there is a decision or question that should be here. Your feedback helps improve the guide for everyone.

What this covers, and what else is here

These decisions are a start. Behind them sits the AIRG in full and all MindForge considerations, written up one by one, plus the wider Singapore picture, comparisons with the global frameworks, and the tools you need to actually use any of it. A few of the highlights:

The full MindForge toolkit, consideration by consideration. The decisions here draw on key considerations. The rest are written up in the MindForge toolkit.

The companion guide for your control functions. The same material for the people who oversee, challenge and assure AI, the second and third line: AI risk management and governance for risk and control functions.

Crosswalks that map one framework to another. The AIRG to MindForge crosswalk, clause by clause; the AIRG, MindForge and the CRI framework together; and the AIRG to FSB crosswalk.

How the AIRG compares with the global frameworks. The AIRG compared with the EU AI Act, NIST and ISO 42001, so a firm across jurisdictions maps once.

The resource library and the ebooks. Everything in one place: AI risk management in finance, the resources, and the plain-English books.

Contents

The six decisions

Decision 1 of 6

What is changing, and how to oversee it

The landscape, and where the board and senior management sit in it.

How can the board and senior management oversee something that moves so fast?

What is changing. Generative AI produces open-ended content and leans heavily on third-party models. Agentic AI goes further: it plans, calls tools and takes actions across steps, with less human involvement at each one. Both open real value in finance, in service, research, operations and compliance, and both raise the stakes, because capability and risk rise together. Whether it is Generative AI today, Agentic AI tomorrow, or something else in the future (for example, recursive self-learning or self-improving AI), the decision at your level is not about the technical details of the tools or technology, but simply about whether it is good enough for the tasks in your firm. And what is good enough for a given task or use case does not change with each generation of technology.

Generative AI suggests and produces content; agentic AI plans, calls tools and acts across steps, with less human in each.

MindForge names five roles. The toolkit splits the firm into five: Executives, Builders, Custodians, Use Case Owners, and Business Users. Builders make the AI, Use Case Owners are accountable for a given use case, Custodians are the oversight and control functions, Business Users apply it, and Executives, you, set the direction and carry the accountability. Knowing who holds what keeps the board and senior management from either rubber-stamping or trying to do another role's job.

MindForge's five roles; the board and senior management are the Executives.

What you own. The AIRG set out a short list of responsibilities for the board and senior management. The board approves the overall approach, sets how AI fits the firm's risk appetite, makes sure roles and accountability are clear across the three lines, and keeps enough understanding to challenge what it is told. Senior management runs the framework the board approves: it puts controls across the AI lifecycle, escalates material incidents to the board in good time, and makes sure the firm has the skills and resources for the job. The board owns the appetite and the approach; senior management owns whether the firm actually operates within them. Neither of you builds or runs the models; you own whether the firm is fit to adopt them.

The board owns the approach and the appetite; senior management runs the framework; both sit over the three lines of defence.

What to do. Ask for a short list of where AI is used in the firm and how much each use case matters. Ask for what you can actually interrogate: the inventory, the risk materiality ratings, and evidence that someone competent and independent reviewed the work, not a polished demo. Then think about the firm's approach to AI, how much AI risk is acceptable, and who is accountable for AI across the three lines.

In an ebook: Agentic and Generative AI for Finance.

Decision 2 of 6

Should we build this, and what will governing it cost

Proportionate adoption: risk materiality first, then the full cost of running it safely.

A business case lands on your desk for a new AI use case. What should you weigh before approving it?

Start with the risk materiality. Before weighing the value or ROI, ask for the risk materiality based on at least the dimensions in the AIRG - impact (on the firm and its customers), complexity (how complex or novel it is) and reliance (how much autonomy and how heavily the firm will rely on it). MindForge also assesses the risk materiality of use cases with these dimensions. Risk assessment is not just paperwork; it's a tool that helps the firm decide how much governance the use case needs, and what it will cost to run safely. It also helps the firm to scale responsibly. Your use cases will grow over time, and the only way to scale is to be clear about the risk materiality of each use case and scale the intensity of controls based on the risks of each use case.

Impact, complexity and reliance set the risk materiality, and that sets how much control the use case needs.

Count the governance cost, not just the build cost. A high-materiality use case needs a higher intensity of controls across its lifetime. This includes independent review before go-live, monitoring after, human oversight, and the people to do all of these. A business case that only takes into account the build and licence cost understates what any use case will actually cost the firm. The right question is whether the value still makes sense once the cost of governing the use case, based on its risk materiality, is accounted for. Sometimes a lighter, lower-materiality design may be able to capture most of the value with a far lower intensity of controls.

Does it make sense to assign the same risk materiality to both systems even if the use case is the same?

Check the firm can actually host it. Whether the infrastructure is adequate is part of the decision: can your environment run, secure, log and monitor this use case to the standard that the risk materiality requires, and keep sensitive data where it belongs? If it cannot, either the additional infrastructure cost needs to be accounted for, or the use case may need a third party to help. But remember that outsourcing it to a third party gives rise to another set of controls.

For third party AI, decide what to demand. The more material your use case, the more disclosure you should require from a vendor, such as an AI Card, or a model or system card. Buying it does not move the accountability to the vendor. What you cannot get disclosed, you may need to compensate with your own testing or contract terms.

What to do. Ask the business to state the use case's risk materiality tier and the full cost of governing it to that risk materiality tier, not the value alone. Then think about whether the value still makes sense once that cost is counted, and whether the firm can run the use case safely or should leave it to a third party.

In an ebook: Scaling AI Governance.

Decision 3 of 6

Is the organisation set up to run AI at scale

The spine, the appetite, the risk taxonomy, the inventory and the dashboard.

The firm wants to scale AI from a handful of use cases to many. Is the organisation set up to do that safely?

Approve one approach for the whole firm. The AIRG does not prescribe a structure. The operating model - whether a central team, hub-and-spoke, or full decentralisation across departments - is the firm's to choose. What's needed is one approach that governs AI risk consistently across the firm, with no gaps between risk types or parts of the business. I call that the spine: approve the spine.

Regardless of operating model, approve one governance approach across the firm.

Mapping to every framework multiplies the questions; a spine and a system hold as the use cases grow.

Fit AI into what you already run. Most firms would already have a risk appetite, and AI is not a separate thing to bolt a new one onto. It runs across many areas and cuts across many risks, so the existing appetite usually continues to apply. The focus should be to find the deltas or AI-specific risks that the existing risk appetites do not yet cover, and make someone accountable for each where that makes sense. For the same reason, think hard before naming just a single owner for all of AI. It looks clean on paper but may be a trap in practice, because many of the controls may sit in functions that one person does not own or run. Also remember that the three lines of defence already settle accountability for many other risks, and AI can also slot into them.

Choose how AI risk sits in the firm's risk taxonomy. MindForge sets out three approaches, which can be combined: update each existing risk category for AI (distributed), expand model or technology risk to hold AI risk (concentrated), or add a dedicated AI risk category (AI-specific). Any of these can work, as long as the choice is deliberate and clear across the firm.

Know where your AI is. You cannot govern, or scale, what you cannot see. An accurate, firm-wide inventory of AI use cases, including low-risk and third-party ones, is what lets risk be aggregated.

Use the inventory as a command centre, not just as a routine record to be maintained. Get a small set of AI key risk indicators from it, such as cases missing from the inventory, drifts or incidents. Ask for an executive view of the inventory against the risk appetite. Add a dashboard that is there to prompt questions from leadership. Measure the performance of use cases based on their tasks, not just the models they run on. After deployment, track these indicators regularly. Even without any change to the system, an AI system can drift away from the threshold it was approved against.

A complete inventory feeds an executive dashboard whose indicators prompt the right questions.

What to do. Make sure the firm has a stated AI risk appetite, a clear choice of how AI risk sits in the risk taxonomy, and a complete inventory of AI use cases. Then ask where the appetite is breached today, what AI is in the firm that is not in the inventory, and which indicator might show a problem first.

In an ebook: AI Risk Management for Directors.

Decision 4 of 6

Where do people stay in the loop, and the skills to do it

Human oversight matched to risk materiality, and the fluency to exercise it.

As AI does more of the work, where do people have to stay in the loop, and does the firm have the skills to be there?

Match oversight to the stakes. Human oversight is not all-or-nothing. It scales with how material the use case is: light, sample-based checks for a minor use case; review of every output, or a person in the decision, for a material one. Design the oversight into the use case from the start, sized to its risk materiality, rather than bolting a reviewer on at the end.

Human oversight scales with how material the use case is.

Oversight is only real if the person can act. A human in the loop who cannot realistically question or stop the AI is just a scapegoat. The reviewer needs the time, the information, the capability, and the authority to say no. The system also has to show enough for the reviewer to question what it produced.

Oversight is real only when the reviewer has the time, information, capability and authority to say no.

Build fluency by role, not a single course. MindForge sets skills against the five roles: Executives, Builders, Custodians, Use Case Owners and Business Users each need a different depth. The board and senior management need enough to challenge what they are told, not to build models. A baseline of AI literacy across the firm, and deeper skills where people govern or build, is what makes oversight work. And be clear what it takes: prompt engineering or prompt design is a shallow skill, not the AI literacy or fluency that AI governance and risk management requires.

Culture carries it. People should be able to raise a concern about an AI use case without it being career-limiting, and should know who owns what. Scaling AI safely is as much about culture and clear accountability as about controls.

What to do. Make sure human oversight is set according to how material each use case is. The people doing it need the time, the authority and the capability to challenge the AI. Then ask, for the firm's most material AI use case, who reviews it and whether that person can really say no.

In an ebook: AI Risk Management for Directors.

Decision 5 of 6

Can you govern an agent that acts on its own

Design-time versus runtime: mandates, institution-owned controls, dispositions and audit.

An AI agent can take actions on its own, not just suggest them. What has to be true before you let it?

You cannot sign it off once. A model that only produces an answer can be reviewed before it ships. An agent acts across steps, and the risk may arise from the path that it takes, not just the single output. So governance has to run while it acts, not only before it launches. This is the line between design-time governance, what you check before go-live, and runtime governance, which needs to be applied while it is working.

Design-time governance checks before go-live; runtime governance holds while the agent acts.

Give it a mandate, and own the controls yourself. Before an agent acts, it needs an explicit mandate: what it is allowed to do, within what limits, and what evidence it must record. The controls should be the institution's, not the vendor's, so that the firm, not the tool, decides what is permitted. The most reliable control is to make a forbidden action impossible, not to ask the model to avoid it.

Decide what happens at each action. For a given action the agent proposes, the firm's controls should decide its disposition: let it execute, send it for a person to approve, or reject it. High-impact actions sit behind approval; some should never be allowed at all. And every action and decision should be written to an audit log, so the path can be reconstructed after the fact and someone can be held to account.

For each action an agent proposes, the firm's controls let it execute, send it for approval, or reject it, and log everything.

Keep a hand on the brake. Oversight and key risk indicators scaled to what the agent can affect, such as actions outside scope, failed tool calls and escalations; a kill switch that someone owns and has tested; and controls that tighten with materiality. Not every agent needs the full intensity of controls, but one that can move money or perform consequential actions does.

A note on the state of the art. I helped a little with MAS's SAFR, a public industry reference for safeguarding agentic finance at runtime, and it relates to a paper I co-authored on runtime governance. The ideas above, mandates, institution-owned controls, action dispositions and tamper-evident logs, are still developing, and the specifics will keep changing.

What to do. Make sure every agent has an explicit mandate, that its controls belong to the firm and not the vendor, and that high-impact actions need a person's approval. There should be a tested way to stop an agent, and a record of its actions complete enough to reconstruct what it did. Then ask what this agent can do without a person, and what prevents it doing what it must never do.

In an ebook: AI Risk Management for Agents.

Any reuse requires attribution.

Want to go deeper?

IBF-accredited courses. The deep-dive and technical courses I designed and teach are IBF-accredited and listed on the IBF Programme Finder, delivered through the Association of Banks in Singapore and the Singapore College of Insurance: Banks (deep dive), Banks (technical) and Insurers (deep dive).

Courses I run directly. I also run sessions in-house for a single firm now, and maybe as open cohorts one day, including briefings for leaders and boards, an oversight course for the second and third line, and a technical hands-on day. See the courses.

Adopting generative AI and agentic AI, responsibly and at scale