Simplify Your Life

Agentic AI is the latest advancment in the world of AI. It’s based on Generative AI (GenAI) but it’s different from GenAI because it has the ability to reason, make decisions and take actions (beyond just info retrieval). The reasoning ability is what makes it different from an AI workflow also. It tries to figure out the steps needed to get a task done and tries to execute the task. 

This post tries to answer 3 most common questions related to Agentic AI

1. When should we use Agentic AI?

We already have process re-engineering, automation, rule engines, machine learning and GenAI and we can just combine them so when should we use Agentic AI. 

For the current crop of AI Agents, the use case needs to meet 3 conditions for it to be worth the effort. The task has to be Repetitive, Valuable and Non-standardised.

• Repetitive – A Company’s and the function’s context is important and current crop of AI agents can’t automatically get the context. Maybe AGI will get there but not right now. A lot of context is in non-digital format. Effort is needed to make AI understand the context correctly. 
• Valuable – The use case has to be of business value and that means it may also be taking significant effort today. Current crop of AI is expensive in terms of energy/ water consumption so best not to use it for low value tasks. If the use case does not move revenue, risk or cost, don’t use Agentic AI.
• Non-Standard (i.e. edge cases are common)- most important factor. If the job can be stadardized and it’s valuable, it’s best to automate it. The task has to have element of judgement/ reasoning in the task. An examples of non-standardised task is when dealing with unstructured data – If you are dealing with structured data then transitional AI i.e. good old ML may work. 

One way to identify the use case in a organisation is to think through where humans are spending time “thinking” ie. using “judgement”. 

2. So what is the biggest challenge to using Agentic AI today?

The biggest challenge is to find the ideal use case that meets the above criteria. Agentic AI is going through what happened in the early days of dot com bubble where there were a lot of bad ideas (eg. Online pet supplies stores with valuations of 400M) or “too early” ideas (food delivery before mobile phone became ubiquitous). Today like then, it’s primarily driven by FOMO (Fear of missing out).

Just like the Internet is ubiquitous now and produced companies like Amazon, Meta, AI will also be ubiquitous as everyone will learn the correct use case for AI.

The focus today is still on use cases that drive productivity instead of driving new business models etc., especially because for the current crop of AI we still need HIIL (Human in the loop).

3. If GenAI is smarter than an average graduate and can perform tasks, are Agentic AIs today more capable than a graduate when applied in work related use cases?

Both Agentic AI and a graduate have basic general knowledge + reasoning ability and can figure out what to do. The key difference is that for a graduate what she learns in the organisation is forming her training data and she is learning from the hands on experience.

For the current crop of AI, it’s not learning from doing. The knowledge from the use case is not going back to its training data. It’s acting like a graduate who knows how to find the policy and steps to get to get a job done (based on documents given to it) but then immediately forget about it. 

The memory is only within the context window (i.e. set of instructions given to it) and in general it cannot use the memory from all other previous tasks it has performed or learn from it.  There are many ways this limitation is trying to be overcome. Bigger context window, RAG (Retrival augmented generation i.e. retrieve the right set of context to address the query or action needed), MCP (Model context Protocol – for agents to work with storage systems), Dedicated Agentic storage or knowledge layer are some of the solutions being explored.

Unless all the outputs generated by the Agentic AI can be used to do reinforcement learning to continuously train the model, this gap will be hard to bridge.