Skip to main content

Where would you add an AI agent to an existing automation?

  • Nov 18, 2026 9:26 PM UTC
  • 2 replies
  • 9 views

Hi Pathfinders! If you could add an AI agent to one of your existing automations, which step would you pick?

In a bank chargeback workflow, for example, an agent could review incoming documents, summarize the case, and flag missing evidence for an analyst before they make a decision.

What’s one step you’d give to an agent, and where would you keep a human review?

2 replies

Shreya.Kumar
Pathfinder Community Team
Forum|alt.badge.img+4

Hey ​@asaudi , great question!

we talk about this in depth and shed light on other best practices in our AIAE certification; 

as a general rule of thumb, you could add AI Agents wherever the input comes from multiple sources, and context from one step in the process changes what information is available in further steps. a human in the loop step should ideally be added for reviewing or taking action before critical information has been published, saved in systems , or communicated to downstream systems. 

 

curious to know more opinions


KhaledMostafaMe
Most Valuable Pathfinder
Forum|alt.badge.img+2
  • Most Valuable Pathfinder
  • Nov 20, 2026 1:59 PM UTC

@asaudi  Your chargeback example suits an agent well: the documents arrive messy, and the output is a judgment call.

I'd give the agent the "read the evidence, summarize the case, flag what's missing" step. The analyst keeps the chargeback decision, and the review gate goes before anything is saved or sent downstream.

The pattern is deterministic bot ➔ agent ➔ human review ➔ deterministic bot:

  1. The step for the agent. Agents suit reasoning and flexible decisions. Scripted steps stay as bots or API Tasks, and the docs say not to default to agents for everything. Keep the agent narrow, with one clear job. For a bigger case, a supervisor agent can delegate to specialists, for example a Claim Details Agent and a Fraud Check Agent.

  2. Feeding it cases. Loop outside the agent, so a process or bot sends one chargeback at a time. The agent then reasons over a single case's documents.

  3. Flagging missing evidence. Give the agent a tracker or validation tool that echoes back what it found and shows what's absent. Agents are more likely to notice gaps and escalate to a human when that tool is present.

  4. Where the human review goes. Put it right after the summary and before the decision is written to a system. In the Process editor, a human step assigns a review to a user or group before the process continues: Bot (Intake) ➔ AI Agent (Summarize + Flag) ➔ Human Task (Analyst) ➔ Bot (Update Systems). Start at 100% review, then drop to 50% and 10% as the agent earns trust. Eventually only exceptions reach the analyst.

The AI Agent best practices page covers this in more detail.