By Rajesh Beri | July 26, 2026
On June 30, AWS published a service availability update that reads like a product obituary. Amazon Kendra, Amazon Q Business, and Amazon Bedrock Agents — now renamed "Bedrock Agents Classic" — all moved to maintenance mode. New customers are locked out starting July 30. Feature development is over. The services are on life support.
These aren't decade-old legacy products being gracefully retired. AWS shipped Bedrock Agents in November 2023. Q Business launched less than a year later. Kendra was the newest of the trio in its current form. The company is now placing AI services into maintenance faster than most enterprises complete a single procurement cycle for the same products.
That's the real story. Not that AWS is cleaning house — every cloud provider does that. The story is that the clock between "Generally Available" and "maintenance mode" has collapsed from a decade to under three years. And the enterprises that trusted first-generation AWS AI services are now carrying forced migration debt they never budgeted for.
The Scale of the Purge
The June announcement goes far beyond three headline services. AWS moved roughly 20 services and features into maintenance mode simultaneously, including:
AI and ML services:
- Amazon Kendra (enterprise search)
- Amazon Q Business (enterprise AI assistant)
- Amazon Bedrock Agents Classic (agent orchestration)
- 10 Amazon SageMaker AI features: Ground Truth, Clarify, Debugger, GeoSpatial, Model Monitor, A2I, Mechanical Turk, Role Manager, Studio Lab, and Profiler
Supporting services:
- Amazon Cognito Sync
- AWS Directory Service (Simple AD)
- AWS IoT Device Defender Detect
- AWS Service Catalog Application Registry
- AWS Systems Manager Application Manager
A March 2026 update had already pushed AWS App Runner, CloudTrail Lake, and Audit Manager into maintenance, while sending Amazon WorkMail and RDS Custom for Oracle into full sunset.
Taken together, these represent the broadest coordinated pruning in AWS catalog history. And the message is unmistakable: AWS is collapsing a sprawl of point solutions into three anchor platforms.
The Three Anchors: Where AWS Wants You to Go
Every retired service has a designated successor. The migration map reveals AWS's consolidated architecture:
| Retired Service | Successor | What Changes |
|---|---|---|
| Amazon Kendra | Bedrock Knowledge Bases | Managed RAG with hybrid search, agentic retrieval, document ACLs |
| Amazon Q Business | Amazon Quick Suite | Agentic modules, BI integration, custom agents, desktop app |
| Bedrock Agents Classic | Bedrock AgentCore | Serverless runtime, policy enforcement, multi-agent orchestration |
| SageMaker Ground Truth | Bedrock evaluations | Automated eval frameworks, A/B testing, batch assessments |
| SageMaker Clarify/Monitor | AgentCore Observability | Real-time quality scoring, behavioral monitoring, CloudWatch spans |
AWS is betting its AI future on three pillars:
- Bedrock — Models, retrieval (Knowledge Bases), and the foundation layer
- AgentCore — Agent execution, governance, identity, and observability
- Quick Suite — Business user experience (replaces both Q Business and QuickSight)
This mirrors what Microsoft and Google did earlier — Microsoft consolidated under Copilot, Google under Gemini Enterprise. The difference: Microsoft and Google built one flagship assistant from the start. AWS shipped Kendra, Q Business, and Bedrock Agents as separate products, then unwound the portfolio in public.
Why This Matters More Than Normal Cloud Pruning
Cloud providers retire services constantly. What makes this different:
1. The lifecycle has collapsed.
Traditional AWS service lifecycle: launch → 5-7 years of active development → 2-3 years in maintenance → sunset. Total: ~10 years of reliable operation.
New AI service lifecycle: launch → 18-24 months of active development → maintenance mode. Total: under 3 years.
An enterprise that started evaluating Bedrock Agents when it launched in late 2023, completed procurement in mid-2024, deployed in early 2025, and reached production scale by late 2025 now has roughly 6-12 months of stable production before needing to plan migration. That's not a lifecycle — it's a sprint.
2. Vendor trust is now a procurement criterion.
The Gartner Security & Risk Summit 2026 identified vendor trust as a formal procurement criterion for the first time. When a platform retires AI products within three years of launch, it teaches customers to discount the next launch. Every enterprise buyer now has to ask: "Will AgentCore still exist in 2029?"
3. Switching costs are invisible but massive.
Research from VaasBlock documents how AI vendor switching costs accumulate through data pipelines, workflow automations, employee habit formation, and institutional knowledge — none of which appear on a balance sheet. The direct migration cost is the smallest part. The real cost is the productivity loss, retraining, and lost customizations that enterprises built on the retired platform.
4. The migration paths aren't clean.
AWS's own Kendra migration guide carries a dedicated section on feature gaps and workarounds. Some Kendra data source connectors lack a native equivalent in Bedrock Knowledge Bases — AWS recommends routing unsupported sources through S3 as an intermediary. The Q Business migration steers customers toward Model Context Protocol integrations for connectors that Quick Suite doesn't natively support — but those MCP integrations can't serve as knowledge base data sources for document indexing.
The Competitive Context
AWS isn't operating in a vacuum. This consolidation happens against a backdrop of intensifying platform wars:
- Microsoft has 20 million paid Copilot enterprise seats and a unified agent platform (Agent 365) shipping at $15/seat/month
- Google consolidated under Gemini Enterprise with usage-based pricing and deep Workspace integration
- Salesforce has Agentforce running in 30,000+ enterprise accounts
- ServiceNow just reported $1 billion in AI ACV in Q2
Meanwhile, Gartner forecasts that $234 billion in enterprise application software spend is at risk from agentic AI disruption through 2030, and expects more than 40% of agentic AI projects to be canceled by end of 2027 due to cost, unclear value, and weak risk controls.
The enterprises getting burned by AWS's forced migration are exactly the early adopters that every platform needs to succeed. If those buyers lose confidence in first-party cloud AI services, they'll increasingly build model-agnostic abstraction layers that treat any single cloud's AI services as replaceable commodities.
The numbers tell the story. Public cloud end-user spending will reach $850 billion in 2026 according to Gartner — a 21.3% jump from 2025. AWS captures roughly 31% of that market. When the leading cloud provider signals that AI services have a 2-3 year lifecycle before forced migration, it doesn't just affect AWS customers. It reshapes how every enterprise evaluates every cloud AI investment. The procurement question shifts from "which platform has the best features today?" to "which platform will still support these features in 2029?"
This is why 81% of Global 2000 firms now use 3+ model families, according to TechJack Solutions research. Multi-vendor isn't just about capability diversity — it's risk diversification against platform rotation.
Framework #1: Enterprise AI Platform Durability Assessment
Before committing to any cloud AI service — AWS, Azure, or GCP — score it against these 8 durability signals. Each factor is scored 1-5 (5 = most durable).
| Durability Factor | What to Evaluate | Score (1-5) |
|---|---|---|
| 1. Anchor Position | Is this service one of the vendor's declared strategic pillars, or a point solution? | ___/5 |
| 2. Revenue Criticality | Does this service appear in the vendor's earnings calls with named revenue metrics? | ___/5 |
| 3. Ecosystem Integration Depth | How many other services depend on this one? (More dependencies = harder to kill) | ___/5 |
| 4. Time Since GA | Services <2 years old carry 3x retirement risk vs. 5+ year services | ___/5 |
| 5. Successor Overlap | Does the vendor already have a newer service covering 80%+ of the same use cases? | ___/5 |
| 6. Migration Guide Existence | If the vendor already published a migration path, the clock is ticking | ___/5 |
| 7. Competitive Differentiation | Is this service unique to this cloud, or commoditized across all three? | ___/5 |
| 8. Customer Count Signal | Is the vendor still actively marketing this service, or has marketing shifted to the successor? | ___/5 |
Scoring:
- 35-40: High durability — safe to build deep dependencies
- 25-34: Moderate durability — build abstraction layers at integration points
- 15-24: Low durability — treat as temporary, invest in portability now
- Below 15: Migration risk imminent — plan exit within 6 months
Applying this to AWS's current portfolio:
| Service | Score | Assessment |
|---|---|---|
| Bedrock (foundation) | 38/40 | Strategic anchor — safe |
| AgentCore | 32/40 | New but anchor-positioned — moderate confidence |
| Quick Suite | 29/40 | Young, combines two predecessors — watch closely |
| SageMaker (core training) | 36/40 | Deep integration, revenue-named — safe |
| Any "Classic" labeled service | 8/40 | Migration imminent — move now |
Framework #2: 90-Day Cloud AI Migration Readiness Scorecard
For enterprises currently running Kendra, Q Business, or Bedrock Agents Classic, here's the operational migration framework. Score each dimension to identify your readiness and gaps.
Phase 1: Discovery (Days 1-15)
| Assessment Area | Questions | Status |
|---|---|---|
| Service Inventory | Which retired services are in production? How many workloads depend on each? | ☐ |
| Connector Audit | Which data source connectors are in use? Do equivalents exist in the successor? | ☐ |
| API Surface | How many unique API calls are made to retired services? Are they abstracted behind an internal interface? | ☐ |
| Data Volume | How much indexed data exists? What's the re-indexing timeline? | ☐ |
| Access Patterns | How many users/agents actively query each service? Peak QPS? | ☐ |
Phase 2: Architecture (Days 16-45)
| Decision Point | Options | Risk Level |
|---|---|---|
| Kendra → Bedrock Knowledge Bases | Direct migration (fastest) vs. S3 intermediary for unsupported connectors | Medium |
| Q Business → Quick Suite | Native migration vs. MCP-based connector bridge vs. custom agent on AgentCore | High |
| Bedrock Agents Classic → AgentCore | Re-deploy with Harness (managed) vs. bring-your-own framework (LangGraph/AutoGen) | Medium |
| Abstraction Layer | Build internal interface to survive future migrations (adds 2-4 weeks) | Low (long-term payoff) |
Phase 3: Execution (Days 46-90)
| Milestone | Target | Verification |
|---|---|---|
| Successor services provisioned | Day 50 | IAM, networking, observability confirmed |
| Data migration complete | Day 65 | Index parity validated, access controls tested |
| Shadow traffic routing | Day 70 | Both systems running, comparing outputs |
| Cutover with rollback capability | Day 85 | <5% latency regression, zero data loss |
| Decommission old service | Day 90+ | Cost savings realized, monitoring stable |
Migration Risk Calculator
For each workload, calculate your Migration Complexity Score:
Migration Complexity = (Connectors × 2) + (Custom Integrations × 3) + (Users ÷ 100) + (Data TB × 5)
| Complexity Score | Timeline | Resources Needed |
|---|---|---|
| 0-10 | 30 days | 1 engineer |
| 11-25 | 60 days | 2-3 engineers |
| 26-50 | 90 days | Small team + AWS SA support |
| 50+ | 120+ days | Dedicated migration squad, executive sponsorship |
The Abstraction Imperative
The deeper lesson from AWS's forced migration isn't about AWS specifically. It's about the structural reality of cloud AI in 2026: AI services are rotating faster than enterprise deployment cycles.
The enterprises that will carry the least migration debt in 2028 are the ones building internal abstraction layers now. That means:
1. Separate orchestration from execution. Don't couple your agent logic to AgentCore's specific APIs. Use framework-agnostic patterns (LangGraph, AutoGen, or custom) behind an internal interface that can target any runtime.
2. Own your retrieval layer. Whether you use Bedrock Knowledge Bases, Vertex AI Search, or Azure AI Search, put your own retrieval interface in front. When the next consolidation hits, you swap the backend — not the application.
3. Treat connectors as adapters. Every data source connector should be wrapped in your own adapter pattern. AWS recommends MCP integrations for unsupported connectors — MCP's stateless spec landing July 28 makes this increasingly viable as a portable standard.
4. Budget for platform rotation. Add 15-20% to every cloud AI project budget for "platform evolution contingency." This isn't waste — it's insurance against the new reality that AI services have a 2-3 year expected lifespan before significant refactoring.
The model-agnostic architecture pattern we covered earlier applies at the platform layer too. Just as you shouldn't couple to a single model, you shouldn't couple to a single cloud's AI service abstractions.
What Happens Next
AWS hasn't set a hard end date for existing Kendra, Q Business, or Bedrock Agents Classic workloads. Maintenance mode means security patches continue. But the writing is on the wall: services in maintenance mode at AWS historically move to sunset within 18-24 months.
That means:
- July 30, 2026: New customer cutoff (4 days away)
- H1 2027 (estimated): Sunset announcement with 12-month countdown
- H1 2028 (estimated): Full service termination
Enterprises on these platforms have roughly 18 months of operational runway. But the smart money starts migrating now — while AWS Solution Architects are actively supporting transitions and before the successor services hit their own scaling growing pains.
If AWS holds this consolidated architecture (Bedrock + AgentCore + Quick Suite) through the next two re:Invent cycles without another disruptive reorganization, the June pruning will read as a calculated risk that traded short-term migration pain for a coherent platform. If they shuffle the deck again in 2028, the platform trust crisis will become permanent.
For now, the action is clear: assess your exposure, build your abstraction layer, and start moving. The July 30 cutoff is 4 days away. The migration clock is already running.
Rajesh Beri is Head of AI Engineering at Zscaler. Views are his own.
