Amazon Bedrock

What it is

Bedrock is AWS's front door to large language models and other foundation models, exposed through a single set of APIs and the AWS SDKs. Rather than committing to one model vendor, teams can switch between models by changing a model identifier. It sits alongside the rest of AWS, so identity, logging, networking and billing work the way the rest of an AWS estate does.

Why it matters

For marketing and discovery work, Bedrock is often the platform underneath internal AI features such as site search, product question answering, content drafting tools and support assistants. Those features shape how customers find and evaluate things on your own properties, which is a growing part of the discovery picture alongside public AI assistants. Choosing a platform also decides how easily you can test a new model when a better one appears.

How it works

Practitioners call Bedrock from application code or a workflow tool, usually with a prompt plus context retrieved from their own data. Bedrock Knowledge Bases can connect a document store to a vector index so answers are grounded in your content, Agents can chain tool calls, and Guardrails can filter topics and content the business does not want generated. Teams typically start on on demand pricing, then move high volume workloads to provisioned throughput.

When it applies

It applies when an organisation already runs on AWS and wants model access under existing security, data residency and procurement arrangements. It is less relevant if you only need a chat interface for individuals rather than an API behind a product or internal tool.

Examples

• A retailer builds an on site product adviser using Bedrock Knowledge Bases over its product manuals and returns policy.
• A B2B SaaS team drafts first versions of help centre articles through a Bedrock API call, then routes every draft to a human editor.
• A publisher runs the same prompt set against two Bedrock models to compare answer quality and cost before committing.

How it is measured

• Cost per thousand input and output tokens by model and workload
• Answer groundedness or citation accuracy in retrieval based features, sampled and scored by humans
• Median and p95 response latency for user facing features
• Guardrail intervention rate, split by blocked topic and false positives