How suitable is Agentic AI for real business world workflows and are the productivity gains promised real?

Course Q&A Agentic AI
, ,
1

  1. The chosen workflow in the final multi agent exercise (report writer , research agent etc) is very different from they types of workflow you would get in say financial services. Several years ago my company created a workflow and decisioning tool (Catalyst) which allowed non technical business users to define and configure workflows using a graphical interface as well as programatic interfaces (APIs). AN example area would be credit decisioning where a consumers details are loaded and processed by the work flow to give a accept/reject response to a consumers credit(eg loan) application. This relies on being certain that i get the same result every time if the same inputs go it…ie is deterministic. An LLM by contrast is not even if the temp is set accordingly. I would see needing to be deterministic as being critical for a lot of financial services workflows.

  2. There is a lot of ‘scaffolding’ code in the course examples…so lots of work still required

  3. Testing looks weak to me. I understand eval but in ‘normal’ software development you would have much more rigour. How would a test harness/framework deal with not deterministic LLM / agentic workflow results…it could be different in each execution of same workflow

2

Totally agree that in the real world agentic workflow is much more complex than what course explains, but one needs to keep in mind any agentic workflow planned based on the task or target in hand varies from industry to place to other confounding factors.

For example you mentioning a financial service using workflow selecting which user gets accepted or rejected for credit or loan, would again depends on the type of bank, or country the workflow is created.

Agentic AI courses explains more of generic aspect for learners tk understand how response are outcome based on the prompt features, parameters and context based refining.

If you can be specific on what specific feature you are looking for agentic ai to catch on, we can probably further elaborate.

Regards
Dr. Deepti

3

ok a more precise example would be a credit application for a consumer car loan in the UK(from what I have seen other countries in the western world are very similar). This would be classed as a credit decisioning workflow. The input to the workflow would be the consumers details eg Name, address, bank, Salary etc. These details would be processed by the work flow (eg verify address) and would involves calls to external facilities like credit ref agencies(eg Experian). The result of the workflow would be accept or reject or refer to underwriter with a list of reasons why. If I put teh same application in 1000 times I need to be sure I get exactly the same output results. To me this is deterministic. The agentic workflows in the course are not deterministic and that make it unsuitable for my example.

The overall point I am struggling with is given all teh hype about Agentic workflows being easy to set up(the are not in the real world), replacing existing legacy applications and humans etc doesn’t fit with what I see on the course. Also testing such applications doesn’t sound reliable

4

The only place where you have statistical results is when you are using generative tools (like LLM’s, etc).

None of your workflow sounds generative to me.