From Survey to SMS in Seconds: How an AI Agent Handles Customer Complaints

In our recent user group session, our Community Captain and Agentic Bounty Challenge winner Akbar Shaikh, RPA Principal Architect at Novatio Solutions, showcased one of the agents he built — a live demo of his Customer Complaint Routing Agent.

Akbar's client was receiving hundreds of survey responses daily with no efficient way to triage them. Teams spent hours manually reading free-text feedback, assessing severity, and routing complaints — often too late to make a difference. The goal was simple: have an AI agent read the feedback, classify how serious it is, and alert the right person immediately.

What He Built

The agent pulls customer feedback from SurveyMonkey via API and classifies each response into one of three tiers: normal, complaint, or critical. Normal covers positive feedback and minor suggestions; complaint flags clear dissatisfaction; critical catches aggressive language or sentiment that demands an immediate response.

Four tools power the workflow: a Survey Tool to retrieve responses as JSON, an AI Skills-based Severity Check Tool that loops through each response, a Human-in-the-Loop approval step for critical complaints, and a Twilio SMS Alert Tool that fires a real-time notification once approved.
There are no hardcoded loops or branching rules. The agent reads the JSON structure and iterates autonomously, deciding which tool to call next based on what it finds.

Watching It Run

Akbar shared a live SurveyMonkey link and invited attendees to fill it in. It pulled all five responses, classified them, and surfaced three as critical. For each critical complaint, the agent paused for human approval before escalating. After Akbar confirmed all three, Twilio sent three SMS messages live. He shared the screenshot: three real alerts on his phone, generated during the session.

How It Was Built

The agent lives in Automation Anywhere and is structured around four tools, each responsible for one part of the workflow.

• The Customer Survey Tool makes a GET request to the SurveyMonkey API and returns all survey responses as a raw JSON payload. No parsing, no key-value extraction — the full response is passed directly to the agent, which figures out the structure itself.
• The Severity Check Tool wraps an AI Skill that reads each response and applies classification rules written in plain language. The rules define what normal, complaint, and critical each look like — positive feedback vs. clear dissatisfaction vs. aggressive language. These rules live in the skill configuration and can be updated without touching the automation.
• The Human-in-the-Loop step surfaces an approval prompt for any response flagged as critical. The agent pauses, presents the complaint, and waits for a business user to confirm before escalating. This step is intentionally temporary — once the team is satisfied with classification accuracy, it can be removed and escalation becomes fully automatic.
• The SMS Tool calls the Twilio API with a dynamically generated message body built from the complaint content the AI Skill identified. The sender and recipient numbers are fixed; the message itself changes based on what the agent found.
• The process wrapper tying all of this together is deliberately thin — start, call the agent, close. There is no branching logic sitting above the agent. It decides the order of operations itself, based on how each tool is described and what inputs and outputs each one expects. That is what makes it genuinely agentic — the developer sets the destination, the agent figures out the route.

This agent is available in the Automation Anywhere Agentic App Store — download, import, and adapt it to your own feedback source.

What the Room Asked

Someone in the session brought up a question that a lot of developers would relate to: if you're passing the full JSON from SurveyMonkey, how do you keep token usage in check, and what stops the agent from processing the same responses twice? On the token side, the SurveyMonkey free tier gives you around 500 API calls a day — enough for a POC, with paid plans scaling up from there. For deduplication, SurveyMonkey lets you mark responses as closed via API once they've been handled. Set that up, and the agent naturally skips anything it's already dealt with on the next run.

The Takeaway

Akbar pointed out that SurveyMonkey is just the input for this POC — the same agent works with any feedback tool. What the demo showed is that AI Skills inside an agentic workflow can replace manual triage with something that reads language contextually and acts before the moment passes. If you have a feedback backlog and a manual triage process, this is a good place to start.

Sessions like this are what the community is for — someone solves a real problem, shows their work, and everyone leaves with something they can use.

The agent is available on the Agentic App Store. Join the conversation in the Pathfinder Community forums or check our events page for upcoming sessions.