Queries on AI Workflow/Usecase

Vasu ch · 2 hours ago

Hello,

I have been exploring Now Assist and AI Workflows, and I have a few questions. I’m hoping to get them answered here.

I have created my AI Workflow with all the instructions and agents. How can I train this workflow to make it more efficient?
While testing the AI Workflow, I sometimes observed that the AI hallucinates and provides incorrect outputs. Is there a way to provide feedback in such scenarios to let the AI know that the output is incorrect?
I see that we can allow our AI Agents to learn from past executions. Can this training data be moved across instances?

Regards,

Vasu

rpriyadarshy · an hour ago

@Vasu ch here is my 2 cents on this.

I have created my AI Workflow with all the instructions and agents. How can I train this workflow to make it more efficient?---->You do not “train” an AI Workflow in the traditional ML sense. Efficiency is improved through design-time optimization, data grounding, and instruction refinement. Since LLM is owned and controlled by Servicenow so no training we can impart. But based on Ai workflow output you can mark is thums up or down (sys_gen_ai_log_metadata). This goes as a feedback on LLM response to Servicenow Team. You can improve efficiency - Use Case Instruction refinement, Agent Instructions,Roles , Scope of work , lesser no of agents to reduce orchestration complexity

While testing the AI Workflow, I sometimes observed that the AI hallucinates and provides incorrect outputs. Is there a way to provide feedback in such scenarios to let the AI know that the output is incorrect?--> There is currently no direct user-feedback loop that teaches the LLM that an output was wrong. You can refine prompts, instructions, and grounding data based on incorrect outputs to reduce hallucination.

Some URLs to refer

I see that we can allow our AI Agents to learn from past executions. Can this training data be moved across instances?--> NO

Regards

RP

Yassine6 · 2 hours ago

Hello Vasu,

I've heard about a feature that performs tests on AI agents: Agentic Evaluations.

I haven't tried it yet, but I hope to do so ASAP.

Dr Yassine

Vasu ch · 2 hours ago

Thanks for sharing the article Yassine. I will check it out.

Regards,

Vasu

rpriyadarshy · an hour ago

@Vasu ch here is my 2 cents on this.

I have created my AI Workflow with all the instructions and agents. How can I train this workflow to make it more efficient?---->You do not “train” an AI Workflow in the traditional ML sense. Efficiency is improved through design-time optimization, data grounding, and instruction refinement. Since LLM is owned and controlled by Servicenow so no training we can impart. But based on Ai workflow output you can mark is thums up or down (sys_gen_ai_log_metadata). This goes as a feedback on LLM response to Servicenow Team. You can improve efficiency - Use Case Instruction refinement, Agent Instructions,Roles , Scope of work , lesser no of agents to reduce orchestration complexity

While testing the AI Workflow, I sometimes observed that the AI hallucinates and provides incorrect outputs. Is there a way to provide feedback in such scenarios to let the AI know that the output is incorrect?--> There is currently no direct user-feedback loop that teaches the LLM that an output was wrong. You can refine prompts, instructions, and grounding data based on incorrect outputs to reduce hallucination.

Some URLs to refer

I see that we can allow our AI Agents to learn from past executions. Can this training data be moved across instances?--> NO

Regards

RP

Vasu ch · an hour ago

Thanks for the response @rpriyadarshy . I got my queries answered now.

Regards,

Vasu