AI inside the same account,
network and control plane

Teams evaluating where to build AI often compare model benchmarks. That comparison matters less than it appears, because model capability converges quickly and the operational questions do not. Where does the data go, who can reach it, what does it cost per query, and who is accountable when it is wrong?

The services we build on

Amazon Bedrock

Managed access to foundation models with a consistent API, so model choice becomes a configuration decision rather than an architectural commitment. Your inputs and outputs are not used to train the underlying models and are not shared with model providers — which is the point most privacy reviews turn on.

Knowledge Bases and OpenSearch Serverless

Managed retrieval with chunking, embedding and vector storage handled, or a custom retrieval pipeline where document structure demands it. We use the managed path where it fits and build custom where structure-aware chunking is required — which, for regulated document sets, is most of the time.

Bedrock Guardrails

Content filtering, denied topics, sensitive information redaction and contextual grounding checks applied at the platform level, so protection does not depend on every prompt being written correctly.

Amazon SageMaker

Custom model development, training and hosting for the problems foundation models do not solve well — forecasting, anomaly detection, structured prediction — plus the pipelines, feature store and drift monitoring that keep them honest.

The reference pattern we start from

Adapted per engagement — this is a starting point, not a template we impose.

Controlling cost per interaction

Inference cost scales with usage in a way most teams do not model until the bill arrives. Three techniques consistently cut it by half or more with no measurable quality loss:

• Model routing. Classify query complexity and send simple queries to a smaller, cheaper model. Most production traffic is simpler than the hardest case the system must handle.
• Prompt caching. Long, stable system prompts and retrieved context are re-sent constantly. Caching them removes a large share of input token cost.
• Context discipline. Retrieving twenty passages when five suffice multiplies cost on every single call, and frequently reduces accuracy through dilution.

Cost per interaction is an architecture decision. It is set in the first fortnight and very expensive to change afterwards.

Data residency and privacy

• Inference pinned to ap-southeast-2 where residency is required
• Customer data not used to train foundation models, and not shared with model providers
• Retrieval enforcing your existing permission model rather than re-implementing it
• Full request and response logging into your own CloudWatch and S3, under your retention policy

We document the complete data flow as a deliverable, because that document is what your privacy officer actually reviews — not a vendor datasheet.

Where AWS is not the right answer

If your organisation has no AWS footprint, no plans for one, and a narrow AI use case, the integration overhead may outweigh the benefit. We will say so. The account-boundary argument is compelling precisely because you already have the account — without it, most of the advantage evaporates.