Automating the Full Customer Support Iceberg: How Gradient Labs Built a Multi-Agent Platform

This episode of Teresa Torres's show features Jack Taylor and Ibrahim Faruqi of Gradient Labs, an AI-native fintech startup automating customer support. They argue frontline answers are only the visible tip of support work, with hidden follow-up work like disputes, investigations, and outreach beneath it. Their platform coordinates three agents (inbound, back office, outbound) built on shared natural language procedures, modular skills, and configurable guardrails. The discussion covers how non-technical experts define agent behavior, how a state machine manages long-running conversations, and how guardrails and auto-evals are built and tuned. It matters for PMs building agentic products that must take action reliably in regulated domains. Note: the source provided is a show-notes page; the full transcript is paywalled, so this summary reflects the description and outline only.

01Key takeaways

• Look beyond the frontline interaction to map the hidden follow-up work that a customer issue triggers.
• Let domain experts author agent behavior in natural language to avoid engineering bottlenecks.
• Model long-running, asynchronous conversations with explicit state machines that track turns and triggers.
• Treat guardrails as classification problems and tune for high recall where regulatory compliance is critical.
• Build sampling-based auto-evals that route edge cases to humans and grow labeled datasets over time.
• For outbound agents, define clearly what 'done' means, since the customer may never end the conversation.

02Key sections

The iceberg of support work

Frontline support is framed as the small visible portion of automation potential, with dispute filings and investigations hidden underneath. Most AI support tools only address the surface.

Three coordinating agent types

Inbound, back-office, and outbound agents work together on complex workflows such as fraud disputes, sharing a common foundation.

Natural language procedures and skills

Subject matter experts define agent behavior in plain language, while modular skills are scoped deterministically per turn.

Orchestration and guardrails

A turn-based state machine manages multi-day async conversations, while guardrails are treated as binary classifiers tuned for recall on regulatory checks.

Human-in-the-loop and evals

An Ask-a-Human tool handles approvals and gaps, and an auto-eval pipeline samples conversations for review to build labeled datasets.

03From the post

“Listen to this episode on: Spotify | Apple Podcasts What happens when a customer reports a stolen credit card? The frontline answer is simple—freeze it. But underneath lies a cascade of follow-ups: dispute filings, fraud investigations, merchant communications, and proactive outreach to gather more details. Most AI support tools”

04Frameworks mentioned

Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.