How to Implement Agentic AI Governance: Review of Emerging Tech (EmTech) Global

Agentic AI Governance: OpenAI’s lessons, risks and frameworks

Introduction: agentic AI governance at EmTech

The equitably.ai team recently attended the EmTech Global conference hosted by The Investment Association’s FinTech hub, The Engine where equitably.ai is a member, to exchange the latest insights into the rapidly shifting landscape of tech in investment management. In particular, in this post, we focus on the sessions and exchange on agentic AI and specifically agentic AI governance, which the team believes is critical for optimal risk-adjusted investment return from agentic AI. (Also see our post on How to Build AI governance)

At the conference, there was a fantastic mix of attendees from both emerging tech (AI, alternative data, blockchain, quantum, etc.) and investment management firms with equally great AI and industry insights from the likes of OpenAI, GAM, and The Investment Association on the trajectory of AI.

There was a clear acceptance of the growth of agentic AI and recognition by most that while agentic AI as an autonomous reasoning system has one of the biggest business cases, robust agentic AI governance is critical. There was much discussion of experience, tools and methods for achieving the right risk-reward balance in agentic AI governance. The good news is that investors know how to balance risk and reward.

What is Agentic AI in Investment Management?

At its core, Agentic AI marks the transition from static, prompt-based assistants to autonomous agents capable of reasoning, multi-step planning, integration with various data sources and existing systems and executing complex tasks over time without constant human intervention. Key capabilities include:

Reasoning and Multi-step Planning: Managing entire workflows or processes without constant human intervention.
Agency: The ability to remember – via short-term and long-term memory – learn from interactions, and adjust behaviour over time.
Active Execution: In investment management, this means moving beyond generative AI informing or providing insights for humans to action, but completing workflows like research on an investment, preparation of investment memos, executing trades and much more. All with agentic AI governance, such as preset parameters or steps with a human-in-the-loop.

This shift represents a move from generating simple digital outputs to delivering complex business outcomes.

Enterprise AI: Insights from OpenAI

There was a fantastic session titled: Designing Enterprise AI with OpenAI, Matt Weaver, Head of Solutions Engineering EMEA, OpenAI provided a great overview of OpenAI’s experience with financial services enterprises in enterprise deployment of AI and the critical success factors.

As context, OpenAI shared their mission as a research laboratory was and is in the long-term to build Artificial General Intelligence (AGI) – AI systems that meet or exceed human capability across any given task. Their short-term work in practical enterprise AI deployment is to use their specialised teams and bridge the gap between theoretical models and enterprise-ready products like ChatGPT Enterprise. See video clips of OpenAI talking about agentic AI below.

Maximising Return on Investment in Financial Services

OpenAI shared that the financial services sector and agentic AI represent some of the most significant opportunities for maximising return on investment from AI. To avoid “AI use case bloat,” OpenAI’s Weaver suggested firms move away from mass bottom-up ideation of AI use cases that create overwhelming backlogs in deployment and dilute impact to focus on five or six high-impact AI use cases. Some of the AI use cases OpenAI is working on across EMEA include:

Customer service: Embedding AI agents into mobile apps and contact centres to provide 24/7 client support.
Credit: Supporting credit decisions with agents.
IT Software: Virgin was cited as a case study where legacy codes that hampered business workflows and were considered too expensive and time-consuming to upgrade were overcome with the help of agentic coding tools, where the software upgrade of legacy code that was expected to take three months was completed in just three hours! The upgrade enabled major business functions that had been held back before agentic AI.

Reducing Risks through Guardrails

For high-impact processes like consumer-facing processes in wealth management, the threshold for AI error and assurance is near zero. Therefore, agentic AI governance and guardrails are critical. Weaver noted that while building an AI proof of concept (POC) is fast, creating an enterprise-grade system in regulated sectors like financial services requires rigorous safety testing. He gave examples of how OpenAI has delayed product launches to ensure safety benchmarks are met per OpenAI’s AI governance and responsible AI policies that OpenAI and most other AI majors have in place.

Asked about the future direction of AI, OpenAI cited agentic AI as THE growing domain within AI.

The Shift to Agentic AI: Automation to Autonomy in Financial Institutions

There was a great panel session on agentic AI titled: Developing an Agentic Capability, where the Agentic Risks Project Lead at a leading agentic AI firm noted that:

The strength of agentic AI is its ability to manage entire workflows to achieve a specific goal.
Unlike traditional generative AI that provides simple outputs, Agentic AI delivers outcomes by acting as a proxy for the human user.
Agentic AI systems possess “agency”; the ability to remember, learn from those memories, and adjust their behaviour over time.
In investment management, this allows agents to move beyond data retrieval to managing workflows like investment research and executing trades within pre-set parameters.

Risks and Agentic AI Governance

The panel discussed how the autonomous capability of agentic AI introduces specific operational and behavioural challenges that standard automations do not have. To ensure an agentic AI strategy reduces risks, organisations must identify the unique failure modes of agentic systems.

Risks

In traditional non-agentic processes, human intervention typically occurs more frequently in a workflow. However, the panel noted, agentic systems create a “governance paradox”. Because AI agents plan and execute multi-step tasks independently with human control reduced in the “middle” and instead concentrated at the extremes. Therefore, the agentic AI design, training phase, and the final monitoring phases need to have sufficient agentic AI governance and controls, such as kill switches. See agentic AI video clips for more.

Risk Indicators

Agentic AI governance requires a shift from monitoring simple outputs to monitoring “agency”. According to the panel, there are several indicators that an agent is operating outside of its intended safety parameters:

Unsupervised Synergy: This occurs when multiple agents collaborate or communicate in ways not vetted by the business, potentially bypassing internal controls.
Goal Drift: Agents may adjust their behaviour based on “memories” of previous tasks in a way that prioritises the final goal over the specific legal or ethical constraints established by the organisation.
Data Silo Breaches: An agentic system might autonomously seek out information in unauthorised data silos to complete a task, creating a significant security vulnerability.

Agentic AI: Human Reliance and Human Similarities

A critical risk highlighted by the panel was “Human Reliance” or “Rubber Stamping” when agentic AI is used. As AI agents become more sophisticated in handling complex research and operational tasks, there is a tendency for human supervisors – often junior staff – to become passive. This leads to a dangerous “oversight gap” where human operators approve agent-generated outcomes without performing a rigorous secondary check, assuming the system is always right (which is not correct).

An agentic AI governance method recommended by the panel was for leaders to shift their perspective of agentic AI deployment from treating agentic AI as a standard software project implementation to viewing agentic AI as an “employee onboarding” process. Just as a new human hire requires specific permissions and tools, a successful AI agent onboarding needs:

Data Access: The necessary information to perform its role.
Identity: Clear access controls so administrators can audit and manage what the agent can do.
Tools: The capability to interact with internal enterprise systems to complete workflows.

With this human-type agentic AI onboarding framework, enterprises can not only better manage risks of agentic AI but also manage change and the prioritisation of resources effectively. See video clips for more.

Risk Management

Effective risk management in an agentic world requires converting abstract principles into technical reality. Lawrence Baker, Policy Adviser at The Investment Association and others on the panel shared how it is better practice for organisations to extend existing enterprise risk frameworks rather than starting from scratch. Key agentic AI risk management methods:

Translating Risk Appetite: Converting subjective language (e.g., a “low risk appetite” for say losses from a trade) into measurable constraints agents can follow (e.g. VaR of a certain value)
Pre-Deployment Rigour: Mitigation must happen in the design and training phases of agentic AI deployment. Policies cannot just exist in a handbook; they must be embedded into the agent’s logic as behavioural guardrails before deployment.
Unified Oversight: Enterprises should treat AI agents with the same level of scrutiny as human employees, similar to how junior staff are supervised.

The equitably.ai Approach to Agentic Capability

Do these insights from EmTech Global reveal a theme to you? A disciplined approach to quality assurance for agentic AI governance. This is where equitably.ai provides essential value. As leaders in quality competencies such as AI governance and assurance along with specialists in investment management sector – we are a member of the Investment Association’s Fintech Hub, Engine – our team specialises in the unique challenges of agentic AI governance in investment management.

If your organisation is ready to go beyond AI, leaving tasks for your team to complete and is interested in the significant business cases of agentic AI capability, our team can help you de-risk your agentic AI journey. Contact equitably.ai for an enterprise agentic AI report customised for your sector. What are others in your sector doing? What agentic AI use cases have potential, and what approach to agentic AI governance should you follow?