Skip to content

SYSTEM Cited by 1 source

Slack AI

Slack AI is the LLM-powered feature suite inside Slack that covers AI-driven channel and thread summaries, Recap (catch-up of activity since last visit), AI Search (high-reasoning, context-aware answers across the workspace), and related generative / extractive surfaces. The product line was built starting in early 2023 to give enterprise customers "security, reliability, and performance our customers expect" (Source: sources/2026-05-28-slack-slack-ai-the-path-to-multi-cloud).

The wiki canonicalises Slack AI as the consumer surface whose serving substrate evolved through four phases:

Phase Period Substrate Wiki canonical
1. SageMaker era early 2023 AWS SageMaker with escrow VPC for Anthropic Multi-region; cross-region IAM; ODCR + cron scaling
2. Bedrock migration mid-2024 Amazon Bedrock (PT) Zero-incident migration; MUs as capacity unit
3. Bedrock On-Demand + Hybrid mid-2025 Bedrock PT + OD + spillover patterns/provisioned-throughput-with-on-demand-spillover
4. Multi-cloud early 2026 Bedrock + GCP Vertex AI systems/slack-intelligent-routing-layer + patterns/model-fallback-hierarchy-with-circuit-breaker

Feature workload shapes (disclosed)

The 2026-05-28 article names three workload shapes to motivate the hybrid + multi-cloud routing decisions:

  • High-volume, latency-sensitive features — channel summaries and similar surfaces that need a "snappy" feel. These were kept on Provisioned Throughput in Phase 3 specifically to guarantee consistent latency.
  • Asynchronous, bursty workloads — nightly Recaps is the canonical example, used to motivate the move to On-Demand capacity. Recap-class features can have 10× variance between peak and off-peak hours, the verbatim figure that justified OD over peak-provisioned PT.
  • High-reasoning features — AI Search is the canonical example named for the ~10% quality lift Phase 4 multi-cloud enabled — specifically by routing to "new high-reasoning models" available on different clouds.

Reported quantitative outcomes (Phase 4)

  • ~10% improvement in quality metrics for complex reasoning tasks (post: "more precise, context-aware answers").
  • ~67% reduction in latency for high-velocity, low-token workloads.

Engineering principles (canonicalised by the post)

  • "Measure first, migrate gradually, and monitor continuously." — solidified by the Phase 2 zero-incident migration.
  • "The abstraction layer is a core requirement" — the Intelligent Routing Layer dominates the model choice.
  • "Treat architecture as a living document" — provider- agnostic routing lets Slack adopt breakthroughs without a rewrite.
  • "Reliability requires provider agnosticism" — internal failovers within one cloud aren't enough.
  • "An LLM service that is 'up' but slow is effectively broken" — soft failures (p90 spikes, feedback trends) are first-class triggers for the routing layer.

Substrate composition

Slack AI sits on top of Slack's broader engineering substrate:

Compliance posture

  • FedRAMP Moderate maintained across all phases. Phase 2 Bedrock migration was specifically gated on Bedrock having "achieved FedRamp Moderate compliance" and matching SageMaker's security posture.
  • Escrow VPC in Phase 1 established the zero-knowledge property: "our data remained private to Slack, and the provider's proprietary model weights remained inaccessible to us."
  • Multi-cloud regional data boundaries — Phase 3 OD relied on Bedrock's cross-US-region routing "while adhering to our regional data boundaries." Phase 4 GCP integration required "Security, Risk and Compliance, Trust and Integrity, AI Quality, Legal, and Cloud Providers" alignment to ensure data boundaries remained ironclad.

Seen in

Last updated · 542 distilled / 1,571 read