I wrote about AWS re:Invent back in November, covering the announcements that seemed most relevant for enterprise AI. Three months later, let’s check in on what’s actually available versus what was just demo-ware.

What’s Actually Shipped

Bedrock Agents: Fully available and being used in production by several enterprises I work with. The capability to build AI agents that can execute multi-step workflows is real and functional. Not perfect – there are edge cases and limitations – but production-ready for appropriate use cases.

Knowledge Bases for Bedrock: Also fully available. The managed RAG service works as advertised. Document upload, automatic chunking, vector storage, and retrieval are all functioning. Quality depends heavily on your document corpus and query patterns, but the infrastructure is solid.

Guardrails for Bedrock: Shipped and being actively used. Content filtering and safety measures are configurable and effective. Several organisations have used this to satisfy internal compliance requirements for AI deployment.

Claude 3 in Bedrock: Available and performing well. The Anthropic models have been the most popular choice among Bedrock users I’ve talked to, particularly for complex reasoning tasks.

What’s Still Preview or Limited

Amazon Q Business integrations: The third-party connectors (Salesforce, Jira, ServiceNow) that were demoed are available but not as polished as the demos suggested. Expect configuration effort and some gaps in functionality.

Trainium2 instances: Available in limited regions and quantities. If you’re not running massive training workloads, this doesn’t affect you directly. If you are, expect to wait for broader availability.

SageMaker HyperPod: Available but complex to set up correctly. This is for organisations doing serious custom model development, not typical enterprise AI consumers.

What Enterprises Are Actually Using

Based on conversations with AWS-native enterprises:

Most common: Bedrock with Claude or Titan models for document processing, summarisation, and code assistance. Straightforward use cases, well-supported.

Growing adoption: Knowledge Bases for internal document search and Q&A. Several organisations have replaced custom RAG implementations with the managed service.

Limited adoption: Agents for Bedrock for automated workflows. The capability is powerful but requires careful design. Most organisations are piloting rather than deploying at scale.

Very limited: Custom model training on SageMaker. Most enterprises are consuming pre-trained models rather than building their own.

The Reality Check

Here’s what I wish AWS had been clearer about:

Integration effort is substantial. Connecting Bedrock capabilities to enterprise systems requires significant engineering. The managed services reduce some complexity but don’t eliminate it.

Prompt engineering matters enormously. The difference between a well-engineered prompt and a naive one can be the difference between useful output and garbage. AWS provides the infrastructure but not the expertise to use it effectively.

Costs can surprise you. Token-based pricing for LLMs is straightforward until you hit scale. Several organisations have been surprised by bills when usage increased.

Multi-model strategy adds complexity. Having access to multiple models in Bedrock is valuable in theory. In practice, most organisations pick one and standardise. Switching between models requires adaptation.

Comparison with Azure

The question I keep getting: how does Bedrock compare with Azure OpenAI?

Bedrock advantages:

• Model choice (Anthropic, Meta, Amazon, Cohere)
• Better fit for AWS-native organisations
• Managed RAG that actually works

Azure OpenAI advantages:

• Deeper enterprise tooling integration
• More mature ecosystem
• GPT-4 access (still the benchmark for some tasks)

For organisations heavily invested in one cloud, stay in that ecosystem. The AI capabilities have converged enough that cloud relationship and existing infrastructure should drive the decision.

What to Do Now

If you were waiting to see whether re:Invent announcements were real:

1. Bedrock is production-ready for standard use cases. If you’ve been evaluating, you can proceed with confidence.

2. Knowledge Bases work and can replace custom RAG for many scenarios. Worth evaluating if you’re maintaining custom infrastructure.

3. Start small with Agents. The capability is promising but requires experimentation to get right. Don’t commit to large-scale agent deployments without piloting.

4. Budget for engineering. Plan for 2-3 months of integration work for meaningful Bedrock deployments. It’s not plug-and-play.

5. Monitor costs carefully. Set up billing alerts and usage tracking before scaling. Token costs add up.

The Bigger Picture

AWS has caught up. A year ago, Azure OpenAI had a meaningful lead for enterprise AI. That’s no longer true. Bedrock is a credible platform with some advantages over Azure for certain use cases.

Competition benefits customers. Pricing is more competitive, features are converging, and lock-in is reducing. Enterprises can make cloud AI decisions based on their specific needs rather than being forced into one ecosystem.

The announcements at re:Invent were mostly real. Three months later, the capabilities are available and being used in production. That’s actually a good outcome – too many conference announcements turn out to be vaporware.