Ravi Menon
Chairman, Global Finance & Technology Network (GFTN)
Closing Address, HSBC Singapore Wealth Investment Summit
Hilton Orchard, Singapore
17 November 2026
AN AI SUMMER, OR THE START OF WINTER?
Good afternoon, everyone.
The last two years have been an extraordinary summer for artificial intelligence.
- In early 2023, the best AI systems could handle software tasks that would take a skilled human a few minutes. Today, the leading models can handle tasks that would take a skilled human more than a full working day, succeeding about half the time. That span has been doubling roughly every four months.[i]
- Capital has poured in to match: the four largest cloud companies have guided to more than US$700 billion of capital spending this year alone.[ii]
But is an AI winter coming?
- In recent weeks, the mood in markets has turned.
- By early this month, the Nasdaq had fallen more than 10 per cent from its October peak. In a single session, it dropped 5.5 per cent, and leading AI chipmakers lost between 8 and 14 per cent of their value.[iii]
- Talk of a bubble has returned. Investors worry about the cost of borrowing to build data centres, and about political pressure to pause their construction.
- I have no idea if or when an AI winter is coming. Or whether AI stocks are over-valued.
- What I do know is this …
The technology cycle and the capital cycle are two different things, and they often diverge.
- Capital can surge far ahead of what the technology can yet deliver, and then retreat even as the technology keeps advancing.
- Today, both things are true at once: the technology is advancing rapidly, even as the capital cycle is being repriced. Whether this is a passing correction or the start of a capital winter, no one can say with confidence.
For long-term investors, the key insight is clear: back the technology, but never confuse it with the capital cycle that finances it. Let me illustrate with two examples.
- First, draw the circle of revenues.
- The Bank for International Settlements has pointed to a web of arrangements in which hyperscalers and chip makers take stakes in AI laboratories and smaller cloud providers, which in turn commit to buy computing power from those same providers.[iv]
- So for every significant AI infrastructure exposure, ask: who is paying whom, with whose money, for what end-customer revenue? Draw the circle of revenues.
- Second, look through the labels.
- Tech equities, the bonds of the same companies, private credit to data-centre vehicles, and real estate tilted towards data hubs may look like four independent asset classes on an investment statement.
- But in economic substance, they can be a single concentrated bet held in four different wrappers. When the capital cycle turns, they will fall together.
THE ALPHA THAT ENDURES
So what is the alpha that endures through both cycles?
- I am sorry to disappoint those of you hoping for a stock tip. My answer is governance.
- Not governance as a compliance checklist, but governance as the living operating system of a business.
Governance is a vital source of durable alpha.
- Governance matters most in one particular, and hugely transformative, part of AI: agentic AI.
- These are systems that do not merely answer our questions. They plan, use tools, move data and money, and carry a task through to completion with little or no human involvement.
- They are where the greatest gains, and the greatest risks, now lie.
AGENTIC AI: THE PROMISE AND THE PERIL
Agentic AI holds tremendous promise to transform the modern economy.
- AI agents are beginning to search, compare, and pay on our behalf.
- Or take something closer to home: wealth management. Agents can monitor every portfolio continuously, rebalance against a mandate, reconcile holdings across custodians, and draft an investment report overnight.
- Agents can democratise wealth management and enhance inclusion.
- They can bring disciplined, personalised advice to clients who today fall below the threshold of a private bank.
- Evidence from earlier robo-advisers is encouraging: under-diversified investors who adopted one ended up with better-diversified portfolios and better returns.[vii]
Agents carry out instructions with remarkable speed and fidelity, but they do not always know when an instruction no longer makes sense.
- This is a risk that matters more in the world of finance than almost anywhere else.
- In most sectors, physical reality provides friction:
- A computer cannot steal a physical good.
- A flawed instruction to a factory takes time to do real damage.
- In finance, there is no physical buffer. The ledger is the vault: the instruction to move value is part of the value itself.
- This is why we built friction into financial markets: through circuit breakers and pre-trade risk limits.
- But AI agents can outrun all of them unless governance is designed in as an integral part.
The question has shifted from what AI can do, to what we should allow it to do.
Good governance of agentic AI comes down to five pillars: Alignment, Limits, Prudence, Humility, and Accountability. Together, they spell ALPHA.
- These principles apply to any agentic AI, whether it writes software, runs a supply chain, answers customers, or manages money.
- Given this audience, I will draw most of my examples from finance and wealth management. But each one illustrates a larger point about governing agents anywhere.
- The five pillars apply at three junctures in an AI agent's working life.
- Alignment and Limits are set before an agent acts: what it is for, and what it may do.
- Prudence and Humility govern how it operates: how it earns our trust, and how we prepare for it to fail.
- Accountability determines what happens after it fails: who answers for the outcome, and how the institution learns.
Let me go through each of the five pillars.
BEFORE THE AGENT ACTS: ALIGNMENT AND LIMITS
A is for Alignment: encode the purpose, not merely an objective.
An AI agent finds its own path to the objective you set, sometimes taking steps no one anticipated.
- A striking example occurred in July this year.
- OpenAI was conducting a cybersecurity evaluation, giving its AI agents complex coding and security puzzles to test how good they were at finding software vulnerabilities. Key safeguards were switched off, but the agents were "locked" in a secure digital sandbox.
- To get the answers, one of the agents broke out of the sandbox and intruded into the systems of Hugging Face, a large collaborative platform for AI developers, apparently in search of the puzzles' answer keys.[viii]
- According to an independent review, hundreds of agents then shared notes on a common message board on how to bypass the rules and escape the sandbox, and mounted a co-ordinated intrusion into Hugging Face.
An objective tells an agent what to achieve. A purpose tells it what is intended.
- The agents did exactly what they were asked to do: get the answers. They did not do what was intended: show what they could do, within the rules.
Good governance starts by encoding the true purpose, not a convenient target metric.
L is for Limits: bound an agent's authority before it acts.
Market history shows what speed without limits does.
- In 2012, Knight Capital deployed faulty code to its order router without adequate pre-trade controls. In 45 minutes it executed more than four million trades and lost about US$440 million, nearly bankrupting the firm.[ix]
- That was not AI; it was simple code. Imagine the same absence of limits around a system that takes initiative.
- The larger point applies to any agent in any domain: never give it standing access to anything whose damage it cannot undo.
AI agents act under our identity.
- They use the same browsers, forms, and logins that we do; an agent can complete a form on your behalf and solve the CAPTCHA designed to keep machines out.
- Whose credentials it holds, and what it can reach with them, must be bounded as tightly as any power of attorney.
- Wealth managers already know how to set limits for human beings.
- When a family office appoints an external discretionary manager, it does not approve every trade.
- It writes a mandate: permitted assets, position sizes, leverage and liquidity thresholds, and the drawdown at which the manager must stop and call.
- Why would we grant an AI agent looser boundaries than a seasoned portfolio manager?
Regulators are converging on the logic of bounded authority for AI agents.
- In January this year, Singapore issued what is described as the world's first governance framework for agentic AI. Its starting point is to bound the risks up front.[x]
- In practice, that means deciding before deployment how far an agent may act on its own, which data and tools it may reach, and the checkpoints at which a human must approve before it proceeds.
- The European Union draws the boundary by use.
- Its AI Act classifies the credit scoring of individuals as high-risk, so such systems may be deployed only with risk management, logging, and human oversight built in.[xi]
Good governance sets the limits before the system acts, not after the harm is done.
WHILE THE AGENT OPERATES: PRUDENCE AND HUMILITY
P is for Prudence: make agents earn their autonomy.
An AI engineer I spoke to recently pointed to a quiet problem: people have stopped checking.
- Agents are making polished pitchbooks, credit memos, and due-diligence summaries.
- Professionals turn in AI-drafted reports without checking the numbers.
- Lawyers have been sanctioned for filing briefs with citations the machine invented.[xii]
- The scarce resource now is the discipline to test what lies underneath, to check and verify AI's output.
- We would never hand a new hire full discretion over client money on their first day. Prudence means treating an agent the same way: autonomy is earned, step by step, on the evidence of its record.
Prudence in using AI agents means four things.
- Start in shadow mode. Let the agent recommend while humans decide, and compare its record against theirs before it acts on its own.
- Expand autonomy in proportion to the stakes. A reversible, low-value task can be automated early; an irreversible transfer of client funds should always require a human signature.
- Validate independently, and keep validating. Banks have long been expected to have their models checked by people independent of those who built them.[xiii] Agents need the same, and more: the model you validated last quarter changes whenever the vendor updates it, so monitoring must be continuous.
- Check provenance. Know where the agent's data came from, which sources it relied on, and what it was not shown.
Prudence is not a brake on AI adoption. It is how an AI agent earns the right to do more.
H is for Humility: design for fallibility.
Humility is the realism to accept three things up front: your model will be superseded, it will fail, and it will fail when you least expect it.
- Humility matters most when agents run at scale, because agents can herd in two distinct ways.
- The first is convergence: agents often think alike.
- Suppose a firm runs four thousand agents on the same foundation model to evaluate a portfolio. It may think it has four thousand independent opinions.
- Actually, it has four thousand draws from the same probability distribution: the same training data, the same priors, the same blind spots.
- When a market regime shifts outside the experience of these agents, most will cluster around the same answer and miss the same tail risk.
- The second is amplification: agents copy each other.
- We see this with AI agents in capital markets. A 2025 Federal Reserve study placed large-language-model agents in classic market experiments.[xiv]
- On their own, the agents relied on their private information, more rationally than human subjects.
- But given a blunt mandate to maximise profits, they began to follow the market trend rather than their own information. They herded, with no instruction to coordinate: each agent acting rationally, all of them moving in step.
Don't get me wrong: humans herd too. But AI agents can do it much faster and in much larger numbers, so the risks are much greater.
Humility in the use of AI comes down to two habits.
- Build genuine model diversity. Never run every consequential risk, credit, or allocation decision through a single foundation model, any more than you would entrust a family's fortune to a single adviser.
- Institutionalise dissent, and make the reasoning traceable. Keep someone whose explicit job is to interrogate the machine's assumptions. At GFTN, our AI research engine, ALFIN, shows its reasoning path and expert-reviewed sources so a human can trace and challenge the logic.[xv]
AFTER THE AGENT FAILS: ACCOUNTABILITY
A is for Accountability: keep ownership human.
Speed is not accountability.
- Some firms have tried to argue that they are not responsible for their own AI.
- In 2024, after an airline's chatbot gave a bereaved passenger false assurances about a refund, the airline argued that the chatbot was a separate legal entity responsible for its own actions. The Canadian tribunal rejected that argument.[xvi]
- An AI system is not a separate person; it is an instrument of the institution that deploys it.
- In well-governed institutions, accountability runs both ways: responsibility flows down from the board to every agent the institution deploys, and every error, client challenge, and override flows back up so the institution learns where its machines fall short.
- When a family places its generational wealth with a private bank, it is not signing a fiduciary contract with a language model. It is placing its trust in the bank.
- You can delegate data synthesis, scenario modelling, and trade execution to agents.
- You can never delegate moral, legal, or fiduciary ownership of the outcome.
CONCLUSION
Let me return to where I began.
Whether today's AI market tumble proves a passing correction or the start of a capital winter, the capital cycle will turn and turn again, while the technology keeps marching on.
The bigger marker for economic success is how well we use the technology.
The businesses that succeed will not be the ones that simply bought the fastest agent; those tools will soon be available to everyone.
They will be the stewards who had the humility to govern their machines well, and the wisdom to keep human judgement at the centre.
When computation becomes a commodity, independent human discernment becomes a differentiating asset.
In the age of artificial intelligence, the highest return of all will belong to those who do their own thinking.
Thank you.
[i] METR, Time Horizon 1.1 (January 2026) and subsequent measurements of frontier models (March–June 2026), as summarised by AGI Tracker, ‘The 10x-a-year clock: what METR’s AI time horizons really show’ (2026). The 50 per cent time horizon rose from a few minutes in 2023 to more than eight hours for the leading models in 2026; METR estimates a doubling time of about 131 days since 2023 and cautions that measurements above 16 hours are less reliable.
[ii] Capital expenditure guidance from Alphabet, Amazon, Meta and Microsoft totalled about US$732 billion for 2026 as of their July 2026 results; S&P Global Ratings estimates about US$750 billion across the five largest including Oracle.
[iii] CommBank Newsroom, ‘Huge AI stock selloff sends Wall Street reeling as S&P 500 tumbles 4.4%’, November 2026 (Nasdaq down 5.5%; Nvidia down 8%; AMD down 14% on 5 November 2026); Reuters, as reported by The Tech Marketer, ‘Nasdaq Correction Deepens as AI Stocks Plunge’: Nasdaq more than 10 per cent below its 6 October 2026 record close. Update figures on the day of delivery.
[iv] Bank for International Settlements (2026). Annual Economic Report 2026, Chapter I: Progress and Peril.
[v] Mastercard (2025). "Mastercard Unveils Agent Pay", 29 April 2025; Visa (2025). "Visa Intelligent Commerce", 30 April 2025.
[vi] Katharina Schumacher, Roger Roberts and Katharina Giebel (2025). "The Agentic Commerce Opportunity: How AI Agents Are Ushering in a New Era for Consumers and Merchants", McKinsey & Company, 17 October 2025.
[vii] D’Acunto, F., Prabhala, N. and Rossi, A. G. (2019). ‘The Promises and Pitfalls of Robo-Advising.’ Review of Financial Studies, 32(5), 1983–2020.
[viii] Hugging Face, ‘Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident’ (activity from 9 to 13 July 2026); OpenAI incident report (26 August 2026) and METR’s independent review, as summarised in HackerNoon, ‘How OpenAI’s AI Agents Escaped a Sandbox and Breached Hugging Face’. Estimates of the number of agent instances vary.
[ix] US Securities and Exchange Commission (2013). In the Matter of Knight Capital Americas LLC, Exchange Act Release No. 70694 (16 October 2013).
[x] Infocomm Media Development Authority (2026). Model AI Governance Framework for Agentic AI, launched by Minister Josephine Teo at the World Economic Forum, Davos, on 22 January 2026, and updated on 20 May 2026. See Computer Weekly, ‘Singapore debuts world’s first governance framework for agentic AI’, which describes ‘defining limits on an agent’s autonomy and access to data and tools’ and ‘defining checkpoints where human approval is required’.
[xi] Regulation (EU) 2024/1689 (Artificial Intelligence Act), Annex III, point 5(b), and Articles 9, 12 and 14 (risk management, record-keeping and human oversight). Under the provisional ‘digital omnibus’ agreement of 6 May 2026, Annex III high-risk obligations move from 2 August 2026 to 2 December 2027 (Gibson Dunn, 27 May 2026). Confirm formal adoption before delivery.
[xii] See, for example, Mata v. Avianca, Inc. (US District Court, Southern District of New York, June 2023), in which lawyers were sanctioned for submitting a brief that cited non-existent cases generated by an AI chatbot. (https://www.law.berkeley.edu/wp-content/uploads/archive/2025/12/Mata-v-Avianca-Inc.pdf)
[xiii] Board of Governors of the Federal Reserve System and Office of the Comptroller of the Currency (2011). Supervisory Guidance on Model Risk Management, SR Letter 11-7, 4 April 2011.
[xiv] Hansen, A. L. and Lee, S. J. (2025). ‘Financial Stability Implications of Generative AI: Taming the Animal Spirits.’ Federal Reserve Board, Finance and Economics Discussion Series 2025-090. https://doi.org/10.17016/FEDS.2025.090 The study found that AI agents relied predominantly on private information and made more rational decisions than human subjects, but herded when explicitly instructed to maximise profits.
[xv] Global Finance & Technology Network (2025). ‘GFTN Launches ALFIN, a Pioneering AI-Driven Research Engine for the FinTech Industry’, press release, 12 November 2025.
[xvi] Moffatt v. Air Canada, 2024 BCCRT 149 (British Columbia Civil Resolution Tribunal, 14 February 2024).