Over recent years, we have seen generative AI (genAI) develop with impressive speed. In just three years, the use of LLMs has gone from being a novelty, to become an accepted part of many professional sectors; many business leaders have moved from cynicism and wariness, to asking how they can implement genAI and how it might help their business to grow and stay ahead of the curve.
Our FutureProof series charts how AI has developed, where it is being used by professionals and the benefits and risks to both businesses and their PI Insurers. However, while many are still grappling with the impacts of LLMs and genAI, a further development is already here: agentic AI – which looks likely to take genAI to the next level.
What is agentic AI?
The precise definition of agentic AI may be hotly debated in some circles, but one way to think of it is genAI, turbo-charged. Essentially, it is a number of AI systems, steps or sub-tasks that have been programmed to work together to undertake a more complex task autonomously. Typically, a genAI “Project Manager” orchestrates various other software tools to perform various functions and provide an overall final output.
Whereas previously, genAI was deployed to carry out a specific task on a specific user instruction (think of a conversation with a chatbot), agentic AI can identify, carry out and even delegate tasks autonomously. Where genAI functions as an assistant, agentic AI could be seen more as a very efficient co-worker. Not only can it identify which tasks need to be completed, it can also identify which other tools in its “team” have the skills and/or access to undertake those tasks and the agent can then delegate the tasks accordingly. The idea is that once the “AI Agent” is pointed in a specific direction, limited further input should be required for it to perform the service required.
The benefits
As with genAI, the obvious benefit of agentic AI is that it can boost productivity and efficiency – even more so than the simpler conversational LLM-based AIs that many are now starting to implement. Agentic AI will in theory be able to take over lower-level tasks and perform them autonomously, freeing up human employees to focus on high level work and strategy.
Repetitive and time-consuming tasks that are subject to a risk of error or oversight (such as a document review for lawyers) could now (at least in theory) be entirely undertaken by agentic AI, resulting in a quicker and arguably more consistent output.
Given the potential to save both time and costs, firms are under pressure to implement AI systems to their full potential and for the benefit of their customers and clients, or risk being left behind.
The downsides
However, while agentic AI is expected to magnify the benefits of AI (potentially providing even more efficient systems), it will also magnify its risks. Even agentic AI ultimately relies on its initial instructions. The more complex the AI, the more time-consuming and challenging it is to prompt and instruct. Firms will need to invest to make the most of its potential. Having said that, once the tool is properly set up, and as AI becomes more powerful and responsive, the danger of asking slightly the wrong prompt, or unintentionally sending an AI tool on an unhelpful tangent may become less of a concern.
Once instructed, it must also be carefully monitored because there is a risk that an unforeseen quirk in the instructions or the information that the AI can access might result in unexpected consequences. In our theoretical legal document review workflow mentioned above, what would happen if an unexpected jurisdiction is mentioned, or a contract term usually only seen in a very narrow specialist field is included in the very different document under review? How will a fully agentic system handle hard-to-predict and or exceptional cases?
Users will need to ensure that appropriate restrictions and governance are in place to keep the AI on track. This will be particularly key in terms of data protection. AI should not be granted uncontrolled access to all data held by a firm. The appropriate safeguards must therefore not only be built into its instructions and permissions but also maintained while it is being used.
It will also be important to monitor and assess the outputs, as with any generative system, there are risks of error, bias and hallucination. Security risks might be exacerbated by the use of multiple tools collaborating via an orchestration AI, without human intervention. If an error is made early on and is not identified, there is the potential that this could be repeated or extrapolated, making any output unreliable or worthless.
How it is already being used
While many firms are still trying to find their feet in the first wave of genAI, those with larger R&D budgets and early adopters are already testing the potential of what Agentic AI has to offer. Some firms have already begun to embrace the use of agentic AI for tasks such as matter management and billing, while exploring where else it might be utilised.
Conclusion
It seems likely that genAI is here to stay, and it will be interesting to see how quickly this next wave is incorporated into professional service and what the professional indemnity ramifications of that might be for professionals, their insurers and industry regulators. Whilst we can see the obvious potential of these systems, careful implementation will be required to ensure that firms (and their clients) are protected.
This article is a part of our FutureProof series. If you're interested in finding out more about how AI is impacting professional life, you can follow that here.
Our content explained
Every piece of content we create is correct on the date it’s published but please don’t rely on it as legal advice. If you’d like to speak to us about your own legal requirements, please contact one of our expert lawyers.