Table of Contents
AI Agents in Investment Management: In brief
I had the pleasure of joining Lawrence Baker on the Investment Association’s IA Talks AI podcast to discuss AI agents in investment management – what makes agents different, the risks you need to manage, and how investment firms can adopt agentic AI safely.
Whereas generative AI produces outputs for your review, agentic AI creates outcomes in your name, meaning that AI agents are not just another AI model.
This page is your listener’s guide to the podcast: a summary, key takeaways, and a minute-by-minute map of the topics we discussed to our published frameworks and articles. By bringing them together here, you can explore how your firm can adopt and govern AI agents safely.
About the episode
The ‘IA Talks AI’ is the Investment Association’s podcast on artificial intelligence in investment management and wider financial services. In this episode, host Lawrence Baker and I discuss agentic AI and its impact – from what an agent is to the arrival of the Human-Agent Organisation.
The one thing to take away
AI agents are not just another AI model – and we should stop referring to ‘AI’ as a single phenomenon.
Generative AI is powerful but passive: you set it a task, and it creates an output for your review.
An agent goes further. It plans, pulls in the tools you allow it to use, takes actions, interacts with systems and other agents and then – crucially – creates an outcome in your name.
It can learn from its experiences, too, adapting how it performs the task next time.
Everything else in the episode follows from that distinction: new capabilities, a new class of risk, and a new discipline for AI agent governance in financial services.
A summary of the discussion
We start with what agents make possible for investment firms: research that navigates unstructured data, operations that no longer need a human to bridge unconnected systems, and people focused on the judgement tasks where they add most value.
From there we turn to control: how agents are plausibility machines, why your data should be fit for an autonomous consumer, and how our working dog analogy – centuries of delegating autonomy to non-humans – suggests three useful principles:
- Grant autonomy slowly.
- Trust the training and the controls rather than the agent’s ethics or the good nature of others.
- Keep accountability with the handler.
We work through the practical consequences for AI agent risk management in financial services: kill switches and incident management under DORA, the “need to know” principle of ‘least privilege’ reframed as the “need to do” to minimise the scope for an agent’s behaviour to drift, red teaming as a pre-deployment discipline, and a risk-based path from low to high-risk agents.
We close with where this leads: behavioural drift and the shift from static to dynamic governance, the human factors that deserve as much attention as the technology, and the advent of the Human-Agent Organisation.
We conclude with a positive message for regulated industries: accountability must remain with humans because an agent cannot accept it, so learn how to master this new technology.
Key takeaways
- AI agents are not just another AI model – generative AI produces outputs for your review; agentic AI creates outcomes in your name – and some agents can learn and adapt.
- Society has delegated autonomy to non-humans for centuries – working dogs teach us to grant autonomy slowly, conditionally, and only when supported by proven controls.
- Trust neither the agent’s ethics nor the good nature of others to treat your agent well – trust the training and the controls – and train the handler as well as the agent.
- Begin control assessments early – the controls you need will influence your platform choice, your build, your training data, and your testing requirements.
- Your incident management procedure will need to stop a run-away agent – and DORA already expects you to manage such incidents.
- Take a risk-based approach – start with low-risk agents, prove your capabilities, and move up the tiers only when your governance is ready.
- Accountability must sit with a human – an agent cannot accept accountability, which is why regulated industries will need people who can handle agents.
Listener’s guide: where to go deeper
Use the timestamps to jump to a topic in the episode, then click through to the article that covers it in full.
| Minute | What we discuss | Go deeper |
|---|---|---|
| 2:37 | AI agents are not “just another AI model” | AI agents and their benefits. |
| 5:24 | Use cases: investment research and operations | AI agents and their benefits. |
| 7:47 | Agents as a bias-free second opinion | - |
| 10:15 | Plausibility machines – the probabilistic nature of agents | Agentic AI Governance in Financial Services. |
| 10:52 | Data fit for an autonomous consumer | The Characteristics of Good Agentic AI Training Data. |
| 13:42 | The working dog analogy | AI Agent Ethics: How to Delegate Autonomy. |
| 20:50 | Defining agent behaviour precisely | Agentic Processes and the Human-Agent Operating Model. |
| 22:56 | Kill switches, runaway agents, and DORA | Incident Management for Agentic AI: Upgrade Needed. |
| 24:26 | ‘Least privilege’ becomes “the need to do” | - |
| 25:19 | Intelligent disobedience | AI Agent Ethics: How to Delegate Autonomy. |
| 27:16 | Red teaming – adversarial testing for agents | Agentic AI Red Teaming in Financial Services. |
| 28:50 | Low, medium, and high-risk agents | How to Design a Risk-Based Agentic AI Adoption Strategy. |
| 32:50 | Behavioural drift and dynamic governance | New Academic Research Finds Behavioural Drift To Be An Agentic AI Compliance Matter. |
| 33:40 | Multi-agent, security, and human-factor risks | Multi-agent risks. Agentic AI security. Human-factors. |
| 38:46 | The Human-Agent Organisation | The Advent of the Human-Agent Organisation. |
| 43:45 | Agent handlers, span of agency, and new skills | Agentic Processes and the Human-Agent Operating Model. |
| 47:38 | Against AI legal personhood | AI Agent Ethics: How to Delegate Autonomy. |
We publish new content on agentic AI all the time, so if you found this useful subscribe to our newsletter to stay up to date on the latest.
Where Agentic Risks can help with AI agents in investment management
If agentic AI is a live topic at your firm (or soon will be), our training sessions in the IA’s AI Academy cover ‘agentic AI 101’, executive briefings, half-day deep-dives, and use case surgeries.
If you would like to move faster, our range of services will support you through each stage of the journey.
Thank you
My thanks to Lawrence Baker and the Investment Association for a thoroughly enjoyable conversation. I hope listeners find it useful. Follow IA Talks AI for future episodes.
Frequently Asked Questions
Society has delegated autonomy to non-humans for centuries – guide dogs, sled dogs, sniffer dogs – but only after structured training (for agent and handler), clear boundaries, proven controls, and accountability that stays with the handler. The same principles should govern how organisations delegate autonomy to AI agents.
A Human-Agent Organisation is one in which humans and AI agents jointly perform work within a system of human governance, oversight, accountability, and intervention. Agents bring the first delegated workforce that is non-human – operating at machine speed and scaling elastically – making agentic AI an organisational change, not just a technology upgrade.
On YouTube and the Investment Association’s website.
