GenAI: the model is raw material; customization determines the result

The Databricks Genie and Amazon Quick case for automating inventory replenishment shows the right choice of lever: demand forecasting once required manually tuning the model for every SKU, covering tens of thousands of items and becoming outdated faster than planners could reconcile supplier spreadsheets. Here, a foundation model forecasts demand across the entire catalog without item-level tuning—the task is solved through RAG and data orchestration, without fine-tuning.

At the other end of the spectrum is Abnormal AI, which protects email for more than 25% of Fortune 500 companies.

Their agents process billions of operations daily, and real-time phishing detection requires more than the model’s semantic reasoning: it needs a computational scratchpad—a sandbox for aggregating data, verifying information, and analyzing code on the fly.

Here, the right level of customization is infrastructure around the model: Bedrock AgentCore Code Interpreter gives the agent an executable coding environment separate from the prompt text.

The Ninth Wave case in open finance is a third example of the same principle.

Each bank exposes an API with its own field names and deviations from the FDX standard; reconciling formats once took specialists weeks of work with correspondence and spreadsheets.

The problem is solved by a data normalization layer built on top of the model—integration engineering.

The model remains constant in this task; the surrounding infrastructure does the work.

One category of tasks is not directly covered by the customization spectrum but consumes just as much engineering time: managing agent access.

When an agent must act in GitHub or Slack on a user’s behalf, it needs an OAuth grant tied specifically to that user (session binding).

Teams used to build this themselves: token storage, access revocation, and agent activity auditing. AgentCore Identity provides ready-made infrastructure for this layer.

The key point is that this layer is mandatory for production, not optional—a fact every company moving an agent beyond a demo environment should remember.

The same logic underpins LLM & Security Gateway in an AI-native integration environment: without it, every agent gets direct access to internal systems without a central point for auditing or revoking permissions, turning a CRM connection into a perimeter vulnerability. An agent that works safely and predictably with corporate data does not appear by itself—it requires authorization, logging, and scope-limiting layers designed separately from the prompt and the model.