Agentic AI - Whoa, Pump the Brakes
The current AI Agent problem is reliability, not IAM
Lately I frequently come across articles that say the biggest obstacle to Agentic AI progress is “IAM” - “Identity and Access Management”. For example, if one worker has 5 or 10 AI Agents acting on his behalf, those agents require access to systems (and therefore permissions and passwords), to execute the tasks they’re responsible for.
First, I don’t know anybody that has deployed 5 or 10 AI Agents that are truly AI Agents as opposed to simple software automations such as a Github cron job. The reason people have not widely deployed AI Agents is because LLMs are fundamentally unreliable. IAM is a future problem. The current AI Agent problem is reliability.
Let’s say I task an AI Agent to write a summary report and send it downstream to my equity traders each morning. I don’t know when or where the underlying LLM - and therefore the AI Agent it underpins - may hallucinate. Therefore, I have to check every stitch of content the AI produces to make sure there are not any errors. What is the ROI in that example? Probably negative.
I do believe there is a tangible ROI in text-centric use cases where the AI Agent acts as the first line of defense. Examples include writing an initial summary of a long-form report and looking for syntax errors in a block of software code.
I’ve previously written that LLMs can be well leveraged across the Software coding function, including writing new code and maintaining existing code bases- albeit while under the supervision of humans. One may set up agents to monitor agents that write code. Yet, we are far away from fully autonomous AI Agents that write, maintain and analyze enterprise code because AI Agents are too unreliable and too dangerous to allow to run autonomously across a company’s enterprise data. Sure, one may set-up guardrails to limit potentially catastrophic actions, but that does not deal with the reliability issue which is a function of LLM’s propensity to hallucinate, to not interpret instructions correctly, and the fundamental lack of commonsense or real world heuristics.