Octave operates in 180 countries and serves more than 60% of the Global Fortune 500, according to an AWS case study. Its Reliance quality management system supports over 650 enterprise customers through more than 40 applications and hosts more than two billion documents. Each customer is a separate tenant whose data must remain within its boundary.

Octave built a shared AI foundation before individual features, choosing complaint summarization and triage first. The design addressed tenant separation, execution under the requesting user's permissions, untrusted content and document ingestion at scale. The company stated a commitment to onboard even its largest tenant within one working day.

Preparing documents for retrieval

The foundation combines Cohere Embed v4 on Amazon Bedrock, per-tenant Amazon OpenSearch Service vector stores, AgentCore execution and Amazon Bedrock Guardrails. The article cites Bedrock's common API for multiple models, AWS Region residency, IAM governance and managed content controls as reasons for the choice.

Octave evaluated embedding models on its own corpus. According to the article, Embed v4 places text and images in one vector space and supports multilingual retrieval without translation preprocessing. In internal testing, 512-dimensional embeddings retained target accuracy while cutting storage roughly 66% compared with 1,536 dimensions.

Document events pass through SNS and SQS to Step Functions, coordinating parsing, chunking and embedding across concurrent Lambda executions in a dedicated VPC. Processing preserves semantic structure in Word files and spreadsheets, applies OCR to scanned PDFs without text layers and handles embedded images. Chunks carry tenant, document type, source application, language and content-hash metadata.

Octave packed approximately 4K-token chunks into requests using the model's 128K-token context window. It reports roughly tenfold fewer API calls and an internal result of 28 million tokens per minute. After batch tuning with Cohere, the pipeline sustained 1.4 million documents per 24 hours per tenant workload. That throughput does not establish onboarding time for every customer.

Permissions and untrusted content

According to AWS, each user session receives a dedicated microVM with its own CPU, memory and filesystem, stopped and sanitized afterward. Per-user OAuth token vaults and session-limited agent credentials propagate identity. AgentCore Gateway authorizes access at tool, operation and parameter levels, with automated reasoning checking defined rules.

The ApplyGuardrail API checks attachment text before embedding, retrieved documents and tool outputs before agent context, and completed briefs before delivery. Attachments are tagged as data, and personal information is masked during ingestion. Extraction runs in a Code Interpreter sandbox without network egress or persistence beyond the session. Detection and masking rates are tracked per tenant.

From attachment to recommendation

The Complaint Triage Assistant uses the Strands Agents SDK for graph workflows in isolated AgentCore sessions. Code Interpreter turns PDFs, images, email attachments and videos into structured content. A Claude-generated summary becomes the retrieval query. A separate model call scores matches and rejects cases with different root causes. Claude Sonnet then assembles facts, sentiment, urgency indicators, historical outcomes and a recommended priority.

Completed briefs are cached in DynamoDB with an environment-specific expiry. Portal access uses private API Gateway and Lambda; document access uses roles and pre-signed URLs, while attachments remain in S3. Traces capture model calls, retrieval and extraction. Bedrock logging and CloudWatch support tenant-level records of requests, responses and costs.

Results reported by Octave

After prompt, model or retrieval changes, AgentCore Evals reruns workflows against expert-annotated ground truth. It measures summary faithfulness, retrieval precision at K and output completeness. After three months of production operation, the article reports that the assistant absorbed tasks Octave measures at 4–6 hours per investigation. The quality handler retains the final decision.

Octave reports 50% shorter deployment time per AI use case, development cycles reduced from months to weeks and semantic search delivered in Q2 instead of its planned Q4 arrival. Audit readiness monitoring, incident narrative generation and predictive quality risk scoring are further extensions. These are findings from a single case study, without independent corroboration of the measurements.

Source: AWS case study.