Artificial intelligence news.
Straight from the original announcements.

ArisGlobal cuts AI inference costs by caching shared prompts on AWS

November 17, 2026, 16:57By the AI News Bites Editorial TeamSource: AWS
ArisGlobal cuts AI inference costs by caching shared prompts on AWS
Image: AWS

ArisGlobal changed how its NavaX AI engine uses Amazon Bedrock, caching a shared prompt prefix instead of each adverse-event document. The company says the redesign reduced effective input costs and helped it support more tenants without proportional cost growth.

Key points
  • NavaX extracts around 130 fields from adverse-event reports.
  • The shared prompt grew from approximately 72,000 to 180,000 tokens over 6 months.
  • Per-case inference costs increased by more than 200% before the redesign.
  • ArisGlobal now caches the shared prompt prefix before sending each case document.
  • NavaX has processed over 1 million real transactions, according to ArisGlobal telemetry.

A growing prompt created a cost problem

ArisGlobal provides life sciences technology across pharmacovigilance, regulatory, medical affairs, clinical, and quality functions. Its LifeSphere NavaX engine uses generative AI to automate adverse-event case processing, a workflow in which pharmaceutical companies assess reports about patients’ reactions to medicines.

When a report arrives, NavaX’s Dynamic Data Extraction agent processes a typically multipage PDF and extracts around 130 fields. These include patient demographics, suspect drugs, reporter information, event timelines, seriousness criteria, and regulatory classifications. The system also applies controls such as Amazon Bedrock Guardrails to filter content and detect or mask sensitive information.

Each case involves multiple sequential model requests. Every request carries system instructions, customer-specific rules, an extraction schema with few-shot examples, and the source document. As ArisGlobal added pharmaceutical customers, the shared prompt prefix grew from approximately 72,000 to 180,000 tokens over 6 months, a 150% increase. According to the source, per-case inference costs increased by more than 200%.

ArisGlobal reversed the cache order

Amazon Bedrock prompt caching reuses a contiguous prefix from the start of a request through a cache checkpoint. Later requests with the same prefix can read cached tokens at a reduced rate. Because the system caches a prefix rather than separate blocks, the order of content determines what can be reused.

ArisGlobal’s first design placed the 10, 200 page source document first and set the checkpoint after it. That allowed the document to be reused for later requests belonging to the same case, but the shared prompt remained outside the cache. A new case brought a different document, so the document cache could not provide reuse across cases or tenants.

The production design puts the shared prompt prefix first, followed by one explicit checkpoint. The per-case document comes after the checkpoint and is processed on each request. This means ArisGlobal gives up limited document reuse within one case to reuse the growing prompt across requests, cases, and tenants using the same agent.

The change supports wider enterprise use

ArisGlobal selected Amazon Bedrock for managed access to foundation models, EU geographic cross-Region inference, and managed prompt caching. The company productionized its first generative AI solution on Amazon Bedrock with Anthropic’s Claude in March 2025.

According to ArisGlobal’s internal production telemetry, NavaX has processed over 1 million real transactions in regulated pharmaceutical environments, handling billions of tokens across multiple enterprise tenants. The entire LifeSphere system runs on AWS.

The caching change matters because the shared prefix contains the rules, schemas, examples, and customer-specific logic that expand as more pharmaceutical companies are onboarded. AWS says fleet-wide prompt reuse supports sublinear cost growth, allowing ArisGlobal to add tenants without costs rising in proportion to the growing prompt and transaction volume. EU routing remains unchanged.

Related news

Airbnb Uses Cost per Task to Tune Coding Agents on Amazon Bedrock
For developersNovember 18, 2026, 22:48

Airbnb Uses Cost per Task to Tune Coding Agents on Amazon Bedrock

Airbnb is using cost per task, alongside model evaluations and production monitoring, to select models and manage coding agents running through Amazon Bedrock. The…

CloudKeeper reaches Select tier in Anthropic’s Claude Partner Network
BusinessNovember 16, 2026, 14:58

CloudKeeper reaches Select tier in Anthropic’s Claude Partner Network

CloudKeeper has achieved the Select partner tier in Anthropic’s Claude Partner Network Services Track, expanding its role in helping businesses implement and manage…

Amazon Bedrock adds structured outputs for all Anthropic Claude models
Products and toolsOctober 20, 2026, 01:32

Amazon Bedrock adds structured outputs for all Anthropic Claude models

Amazon Bedrock now supports structured outputs for all Anthropic Claude models, letting developers receive JSON responses that match a defined schema. According to AWS…