AI LEADERS
Governance
Governance tops checklist for agentic autonomy change
A third of leaders say firms have already revisited frameworks or resourcing for agent management
It’s striking that leaders are almost twice as likely to prioritise strong output quality (57%) as potential efficiency gains (31%) when evaluating agentic genAI opportunities — and of course appropriate apparatus for oversight and control is the factor that features most prominently in leader discussions and decision-making. If a firm is removing people from direct execution of parts of a process it must be very confident in how such delegation is monitored, perhaps involving additional responsibility and review elsewhere as a result. A third say they have already made some changes to governance for this reason, and over half (57%) are reviewing the situation.
For example, agents may touch more systems and information repositories to complete tasks effectively than AI assistants — by extension they may need more controls. Agents can be restricted not only from accessing systems or sources, but from actions that are permitted where access is granted. Humans, of course, establish the parameters within which AI can operate and provide the supervision and points of potential intervention. At the same time, where deployment or use is not confined to a few individuals, firms may need to invest in new types of training and a communication programme to build awareness and appreciation of agent best practice.
2%
confident to grant agents fully autonomous process execution
Morgan says: “The fact agentic AI can take actions across multiple systems will likely create additional governance needs around confidentiality, client information, privilege, security permissions and auditability.
“But firms need to fully understand the workflows that underpin agentic automation and to have analysed where red flags are likely to arise. It also means that the human part of that equation needs to understand ‘what good looks like’ and be able to quickly identify irregular output before it goes in front of a client. Longer term, there are interesting questions around training of junior fee earners and how we inject that knowledge of what good looks like into them when automation has replaced some of the valuable experience-building work. Potentially agentic AI will lessen the need to be brilliant at prompting, but technical fluency, critical thinking, judgement and supervision of AI-enabled workflows will likely rapidly become core professional skills.”
Hardly any are prepared to let genAI agents take any actions without explicit approval at present — although two-fifths believe limited actions are possible where an approval process is in place. The question of extent of autonomy is clearly a live one, with almost a fifth of leaders currently unsure where the line lies.