Contents
IVRy is shifting reliability improvement work to product teams, while SRE transitions to providing the expertise and shared infrastructure each team needs. Its 2026 activity report shows practical progress, including completing the on-call handover for one team, expanding SLO reviews, detecting silent failures, and reducing dependency risks associated with external AI.
Key Points#
- IVRy uses a Hub & Spoke model combining embedded SREs with a shared Core Reliability function.
- The AI dialogue engine team took over first-line on-call response in September 2026, while other areas are still transitioning.
- Through weekly Production Meetings, product teams turn error budget consumption, incidents, and alert trends into concrete actions.
- Fallbacks for external AI and business-level monitoring address failures that infrastructure metrics alone cannot reveal.
Embedding SRE in Product Teams#
IVRy’s “Reliability by Default” strategy aims to make reliability part of everyday product development. Ultimately, product teams should be able to manage SLOs, handle on-call response, and conduct postmortems without ongoing SRE involvement.
The organizational model has two components. Embedded SREs work across the telephony, AI dialogue engine, and Rails application product areas. Core Reliability handles shared observability standards, incident response processes, EOL management, and cross-team improvements. Staffing across these areas is not yet complete, and the central function is still taking shape.
SRE involvement also varies with product maturity. Early-stage products may need direct implementation support, while others may need coaching or shared infrastructure. IVRy uses product stages and engagement models to explain priorities and derive project goals. The practical lesson is to choose the support a team needs now, then define what it should eventually operate independently.
Turning SLO Reviews into Action#
IVRy introduced SLI/SLO reference materials and templates across its engineering organization. Adoption remains uneven, and many teams still need SRE support to select useful metrics. Publishing templates alone does not establish operational practices.
Weekly Production Meetings provide a place to put this into practice. Each team reviews error budget consumption, recent incidents, and alert trends, then decides what to change. These meetings are already running in the telephony and AI dialogue engine areas and began in the Rails area in October 2026. Developers have started raising reliability concerns themselves, but independent operation remains limited to some teams.
To decide where to start measuring, IVRy prioritizes functions using four questions.
- Would a failure stop customers’ business operations?
- Does the function have a financial impact?
- Do other functions depend on it?
- Does it involve external integrations?
Teams also factor detection gaps into these assessments. High-impact functions whose failures remain invisible take priority over well-monitored, lower-impact functions.
Preparing for the On-Call Handover#
The AI dialogue engine completed its handover to developers in September 2026. IVRy organized the transition into four reusable steps.
1. Ask each developer about their concerns and capability gaps. 2. Document investigation and response procedures for every alert sent to the alert channel. 3. Establish a shadowing period in which SRE leads the response and explains its decisions. 4. Transfer first-line response to developers, with SRE handling escalations.
Other areas require different preparation. The telephony area is separating immediate alert response from the coordination and decisions needed after escalation, while working with HR to account for working-hour constraints. The Rails team aims to establish a rotation within an engineering group of around seven people after agreeing on SLIs and adding instrumentation.
Developer on-call duties also carry a cost. More operational work means less capacity for feature development. IVRy treats “You Build It, You Run It” as an arrangement to revisit as workloads change. Completing a handover is therefore a checkpoint, rather than proof that the division of responsibility should remain permanent.
Detecting Failures Customers Experience#
Even when Rails requests return HTTP 200, notifications may fail to arrive, billing workers may exhaust their retries, or transcription jobs may never complete. IVRy calls these semantic failures. Measuring request success alone cannot detect them.
The team created a Function Catalog that records operational units beneath the existing product catalog. Each function lists its owner, dependencies, detection methods, links to SLIs/SLOs, and criticality. IVRy assessed 13 functions and started with incoming-call notifications and notification integrations. The overall effort proceeds through defining SLIs, agreeing with product managers, implementing instrumentation, and establishing group-level on-call rotations, with an expected timeline of six months to two years.
External AI requires the same attention to customer outcomes. When a transcription API slowed significantly, IVRy concluded that it had been relying on observed performance without any processing-time guarantees. It is now evaluating fallback providers against datasets, quality, and operational requirements, while separating processing capacity for production calls from other LLM usage. The fallback providers’ capacity also needs verification.
Allocating Product Resources to Reliability#
In October 2026, IVRy consolidated much of its external AI and Rails reliability work into Dialogue Product Reliability. A product manager owns the project, an SRE serves as technical lead, and backend and AI engineers participate. This gives work spanning multiple disciplines a clear place in product prioritization.
Incident response also has shared support. A Slack bot creates incident channels, assigns roles, records timelines, and generates postmortem templates. The remaining weakness is the quality of learning: reviews sometimes end with instructions to “double-check.” SREs now participate in postmortems for SEV2 and above, steering discussions toward systemic improvements.
Several plans remain unfinished. Core Reliability is expected to start with 20% of each embedded SRE’s working time. Disaster recovery scope and incident metrics need further consideration, and LLM-assisted incident investigation remains a future topic.
Next Step#
Choose one function that matters to customers and, before the next weekly reliability review, document its owner, business success criteria, current detection gaps, and first SLI.
Source: https://zenn.dev/ivry/articles/70d4f05ce4ae5c
Contents
