Feedback to Action
An AI voice agent that collects customer feedback by phone, and turns thousands of calls into one report per location.
- Year
- 2026
- Time to build
- 2 months
- What we did
- Voice agent, Telephony, Web app, Reporting, Infrastructure
- Built with
- TanStack, Twilio, ElevenLabs, LangGraph, Postgres, AWS
- Alongside
- Dan Gingiss
The problem
Surveys get answered by the people least representative of your customers, and the answers arrive as numbers with no reason attached. A phone call gets the reason. A phone call also does not scale, and that has always been the end of the argument.
How it reaches people
Every brand gets an agent with its own greeting and closing, and every location its own number and its own QR code for the counter, the receipt or the door. A customer rings, or scans, and talks for about a minute.
Two fixed questions, and one that is not
The agent asks everyone the same two questions. That is what makes a thousand calls comparable to each other — change the wording and you are no longer measuring the same thing. The third question is written on the spot, out of what the person just said, so the call follows the one thing actually on their mind. Fixed enough to count, loose enough to be a conversation rather than a survey read aloud.
The recording
Every location scored on its own
The score itself is Dan Gingiss’s — the Experience Maker method, which Humach brought to the project and we implemented. Every call comes out of the pipeline with a score of its own, and a location’s score is those calls. A brand with three shops is really three businesses, and the average across them hides which one has the problem. So each location carries its own number and its own line on the chart, and a branch sliding away from the others shows up before anyone rings to complain about it.
What to fix, in their words
The pipeline does not put a sentiment score on a call and stop. It pulls out the points the person actually made — the seating, the price, the person at the register — and attaches a sentiment to each one separately, which is how a single call can be warm about the staff and scathing about the queue. Those points are then counted across every call and sorted into what to keep doing and what to fix. Each one carries the customer’s own phrasing rather than a summary of it, and opens back to the calls it came from, so a claim about a branch is never something you have to take on trust.
Where it landed
Two months from the first line of code, and we built all of it: the numbers and the call handling on Twilio, the agent that speaks and listens through ElevenLabs, the LangGraph pipeline that reads every conversation afterwards, the dashboard a manager works from, and the AWS infrastructure under all of it.
In July 2026 Humach took it to the UnitedHealthcare and Optum Global Innovation Challenge, the seventeenth one, and it placed third.
The agent answering the phone is not the interesting part. Plenty of things answer phones. The interesting part is the pair either side of it: a call structured tightly enough that a thousand of them can be compared, and a pipeline that turns each one into scored points rather than a number.
What exists today is the measurement. The next phase is the part that spends it — a report that says what to do about the seating at Hawthorne, and in what order, instead of leaving a manager to work it out from the evidence. Everything above was built so that report has something true to stand on.
Built for Humach, who partnered with Dan Gingiss and hired us to build it. The method is his — the Experience Maker Score the product reports is his framework, implemented. Everything that shipped is ours end to end: the voice agent and the telephony behind it, the analysis pipeline, the web app, the reporting, and the AWS infrastructure it all runs on.