As a Product Manager, I do not view Sierra as just a "chatbot." I view it as an agentic orchestration layer. It sits between the unstructured intent of the user and the structured systems of record within the enterprise.
The architecture manages the tension between probabilistic reasoning (where LLMs are creative but potentially unreliable) and deterministic execution (where APIs must be accurate). Here is my understanding of how the platform functions.
"One Agent" does not fit all. The architecture remains consistent, but the Data Sources and Guardrails shift drastically between verticals.
Below is how I map this architecture to three distinct customer use cases to visualize the logic flow.
How does the agent decide whether to offer retention incentives or just process cancellation?
I would start by partnering with the Netflix retention team to define exactly what they want in clear rules. This is best captured in a simple Excel or Google Sheet. For example: "If engagement score is below 30 and no offer has been used in the last 6 months, show the ad tier trial. If already offered twice, skip retention and cancel."
This sheet becomes the single source of truth. We then implement it two ways. First, turn the rules into deterministic guardrails or routing logic. This is fast, cheap, and predictable. Second, feed it to the LLM as structured examples or context, which is better for handling unusual cases.
Later, we review real outcomes like offer acceptance and post offer churn, then update the sheet. This can be done manually by the retention team or with light automated ranking based on logged data. The result is that the business has full control while the agent executes consistently and safely.
How do we verify the user before changing subscription status?
Any state changing tool call like cancel, downgrade, or credit must pass supervisor gated verification. The flow starts with low friction checks like "What show did you watch last?" and escalates to OTP via email or SMS if needed. Device signals are used when available. Confidence must hit a high threshold, say 95%, before execution. Anything ambiguous routes to human support. This keeps fraud low without adding too much friction for real users.
How do we know the system is actually working well?
On the quantitative side, I would track resolution rate, handle time, containment percentage, retention lift, upsell acceptance, and supervisor rejection rate. On the qualitative side, CSAT and NPS after interactions along with sampled session reviews for edge cases. The real test is ROI. Does the agent meaningfully improve lifetime value through retained subscribers and reduced churn? Does it lower support cost without creating bad experiences? A/B tests on reasoning prompts, offer logic, or verification steps would help us iterate.
I am not within the company, but looking at the broader competitive landscape from the outside, I see a clear divide in how competitors position themselves.
In the US, Giga.ai markets heavily on speed, claiming agents can go live in just two weeks. Their strength is rapid deployment. In Asia (relevant given this role's Singapore base), Yellow.ai focuses on volume, boasting 150+ pre-built integrations and massive scale for BFSI.
The implicit claim from competitors is that speed and quantity matter most.
Winning Strategy: Sierra wins by proving that reasoning quality and safety result in better long-term ROI. A fast bot that hallucinates a refund policy creates liability. A safe bot that reasons correctly creates trust.
To execute this strategy effectively, I would prioritize the following technical considerations: