AWS Machine Learning Blog, read from an operations desk

Latest

AWS Batch adds quota shares for shared GPU training on SageMaker

AWS Batch has a new scheduling capability called quota management. Admins split GPU capacity into quota shares per team or project, and idle capacity can be lent to other shares automatically. Preemption reclaims lent capacity and runs higher-priority SageMaker Training jobs first, so nobody coordinates by hand.

What changes for operators — This matters only if your company trains or fine-tunes models on GPUs that several people or projects share. If a 10-200 person team fits that description, the quota shares replace the Slack threads about who holds the GPUs: give each project a share, mark which ones may lend or borrow, and let the scheduler reclaim capacity. If you only call hosted models through an API, there is nothing to change. The source gives no pricing or regional availability, so check the AWS Batch user guide before you plan around it.

AWS adds team access and health metrics to Nova Act QA Studio

If you run a web product and have been piloting AI-written tests, this update addresses what breaks once a second and third team join: nobody knows whose tests are failing, and everyone can edit everything. A 10-200 person company would map each product to one application and one owner, give developers Editor, give release managers Executor and leave management Viewer. The flaky-test ranking shows which tests to fix first. Note that it is a reference solution you deploy and maintain with AWS CDK in your own account, so it only suits you if you already run on AWS and have someone to own the deployment.

Amazon Polly upgrades Generative voices, adds Hindi and Finnish

If your support line, order-status IVR or voice agent already uses a Polly generative voice, the new engine arrives on its own and costs nothing extra, so the work is checking that your own scripts still sound right. Listen to the calls where the voice reads back order numbers, addresses, totals and confirmation codes, since that is where AWS says accuracy improved. If you serve Finnish or Hindi speakers, or Korean, German or Brazilian Portuguese ones, the catalog now has generative options to test against what you use today. If your flow depends on speech marks, those are still not available on generative voices, so check that before you switch. Before deploying across Regions, confirm the voice you picked exists in the Region you run in.

AWS SageMaker adds training session APIs for custom LLM loops

For a 10-200 person B2B company, this matters only if someone on the team already fine-tunes models, for example to train a support-reply model against your own quality checks or a sales-qualification model against your own scoring rules. In that case you no longer have to choose between a fixed managed recipe and running a GPU cluster, and per-token billing makes experiments cheaper to compare. If you only call hosted models through an API, nothing changes for you. The first step is to run the GSM8K sample with your own reward function and see what a run costs before committing to anything.

Amazon Bedrock adds TwelveLabs video model Pegasus 1.5

For a 10-200 person B2B company, this matters only if you hold long video: recorded demos, support sessions, training material or webinars. The source's own examples come from media and sports, so any fit with sales or support recordings is our reading, not the vendor's claim. If you do have such a library, you can define a segment once, such as a pricing discussion or a product shown on screen, and get consistent JSON for every file. That feeds a CRM, a ticket queue or a search index without anyone scrubbing through footage. If you have no long video, there is nothing here for you to act on. Check the Region and the separate AWS Marketplace billing before you plan around it.

AWS adds Claude Haiku 5.5, cuts subagent costs 75%

For a 10-200 person company running support ticket routing, document classification or knowledge-base Q&A through Bedrock, this is a direct cost lever: the same volume of quick, repeatable tasks now runs at roughly a quarter of the previous Haiku price, and the new effort controls let you dial down cost on simple lookups while keeping more reasoning budget for harder cases. Paired with Opus 5.5 as the planner, Haiku 5.5 becomes the cheap execution layer for many parallel subagents, which is the architecture pattern most relevant to teams automating high-volume, low-judgment work like inbound triage, data extraction or bulk document review.

Claude Models Now Run Inside India via Amazon Bedrock

For a 10-200 person B2B company serving Indian customers, or bound by contracts that require in-country data processing, this removes a real obstacle to using Claude for support or sales automation: previously, routing through Bedrock could mean your prompts and outputs left India. Now you can select the IN inference profile in the console or API and get the same models with inference confined to ap-south-1 and ap-south-2. Billing, quotas, and CloudWatch/CloudTrail logs stay attached to your source region, so monitoring and cost tracking don't fragment across geographies. Teams building a Claude-based support bot or sales assistant for an Indian entity can now do so without a separate compliance workaround or a different model vendor.

Claude Sonnet 5.5 lands on AWS Bedrock

For a company already running Bedrock-based coding assistants, alert triage, or document drafting, this is a straight cost-and-latency lever, not a new capability to evaluate from scratch: routing well-defined, high-volume tasks (SQL generation, first response to alerts, spreadsheet edits, routine document work) to Sonnet 5.5 while reserving Opus 5.5 for release debugging, security review, or contract redlining should lower per-task spend without touching the existing IAM, CloudTrail, or Guardrails setup. Teams with fixed monthly AI budgets for IDE coding agents or support-ticket triage get the clearest immediate win, since Sonnet 5.5 is pitched specifically at continuous or at-scale workloads with a fixed spend cap.

AWS Ships a Ready-Made Container for Speaker-Labeled Transcripts

For a support or sales team drowning in call recordings, this removes a real chunk of the engineering work needed to get transcripts that say not just what was said but who said it and when, down to the word. That's the difference between a transcript you can search for compliance review and one you can actually build automated QA, coaching, or sentiment scoring on top of. The catch: this is still an AWS infrastructure component, not a finished product, someone still has to wire up the SageMaker endpoints, choose real-time versus asynchronous deployment based on call length, and manage GPU costs, autoscaling, and S3 security. Teams without in-house ML engineering will still need a systems integrator or an existing vendor that has already built this layer in.

Anthropic's Opus 5.5 lands on AWS Bedrock with lower cost per task

For a 10-200 person B2B company running agentic coding assistants or document-heavy knowledge work on Bedrock, Opus 5.5 offers a direct lever: lower average cost per task and clearer step-by-step reporting during long-running sessions, which matters for anyone reviewing agent output before it reaches a customer or a contract. The catch is the new safety classifiers, which refuse more requests than earlier Opus versions in areas like biology and cybersecurity — teams with support or ops bots that occasionally touch security-adjacent topics (vulnerability triage, incident response drafting) should test refusal behavior before rolling this model into production, since a blocked response mid-workflow is worse than a slightly more expensive one.

Grok 4.6 Lands on Amazon Bedrock With Guardrails

For a 10-200 person company running sales or support agents on AWS Bedrock, this is a real build decision, not just a new model to try. Guardrails, invocation logging and prompt caching now work on the bedrock-runtime endpoint, which is the one to use if you need a policy boundary around an unattended agent and an audit trail of what it actually did. But structured JSON output and server-side tool use stay on bedrock-mantle only, so a workflow depending on strict schema output has to pick that endpoint instead. Pricing adds another lever: routing through the cheaper Global cross-region profile versus the US-only Geo profile, or dropping to the Flex service tier at half the standard rate for non-urgent batch work, changes the unit economics of an agent more than the model choice itself.

AWS Reworks Bedrock AgentCore to Cut Idle Memory Costs

If your support or sales agents run on Bedrock AgentCore, this changes two numbers you actually pay attention to: the AWS bill and how long a customer waits when an idle agent wakes back up. Previously, a long-running or bursty agent kept paying for its memory peak the whole session, and cold starts got worse as your container image or concurrency grew — a real problem for agents that sit quiet most of the day and spike during business hours. The new runtime bills closer to actual usage and keeps cold starts flat at roughly 2 seconds no matter the image size, so teams running several agents (a support triager, a lead-qualification bot, an internal ops assistant) can leave them scaled to zero between requests without the old latency penalty when a customer or rep hits them cold. It's an infrastructure change, not a new capability, but it lowers the operating cost of exactly the pattern most 10-200 person companies use: several purpose-built agents that are mostly idle.

Amazon Bedrock Adds Prompt Caching, Cutting AI Support Tool Costs Sharply

If your support chatbot, internal knowledge assistant, or sales copilot runs on Bedrock and sends the same system prompt or product documentation with every request — which most retrieval-augmented tools do — this reduces your per-query cost and speeds up response times without any change to the model itself. Teams running high-volume support automation (think hundreds or thousands of tickets a day) should see the caching applied automatically or configure it explicitly, since AWS notes it works best when a large portion of the prompt — like a knowledge base excerpt or tool definitions — stays identical across calls. For a 10-200 person company already paying per-token for AI support or sales workflows on Bedrock, this is a direct cost lever worth checking this quarter, not a future consideration.

Ninth Wave Automates Partner Onboarding with Bedrock AI Agents

For a 10-200 person B2B company, client and vendor onboarding is often the slowest, most manual part of the sales-to-delivery handoff — contracts, compliance checks, integration specs, and account setup all reviewed by hand. Ninth Wave's approach shows how a Bedrock-based agent can read onboarding documents, flag exceptions, and generate next-step guidance automatically, which is the same architecture an operations team could apply to shrink a two-week onboarding process to days without adding headcount.

AWS Adds Real-Time Quality Scoring for Production AI Agents

If your company has deployed an AI agent to handle support tickets, qualify leads, or trigger workflows, you likely have no visibility into whether that agent's answer quality is degrading over time — a model update, a new edge case, or a documentation change can silently break it. AgentCore Evaluations lets you set quality thresholds (accuracy, relevance, safety) and get alerted automatically when a production agent starts drifting, rather than discovering it three weeks later in a customer escalation. For a 10-200 person company without a dedicated ML monitoring team, this is the difference between catching a broken support bot in hours versus finding out from an angry client.

AWS SageMaker Adds Smarter Routing to Cut Self-Hosted LLM Latency

Most 10-200 person B2B companies call a hosted API like OpenAI or Anthropic and this change does not touch them directly. But if your ops or support automation runs a self-hosted or fine-tuned model behind SageMaker — common when handling sensitive customer data, ticket histories, or proprietary sales scripts that need to stay in your own VPC — this routing update is a free latency and cost reduction. Support bots and agent-assist tools that reuse the same system prompt across thousands of tickets per day will see faster first-token response and lower GPU spend simply by upgrading to the new routing strategy, with no changes to the prompts or application logic themselves.

Claude 5.1 Lands on Amazon Bedrock

If your support ticketing, sales-enablement, or internal copilots already call Claude through Bedrock, this is a low-friction upgrade: change the model ID in your existing integration rather than re-platforming. Before flipping the switch on a production workflow — a support triage bot, a CRM summarizer, a contract-review assistant — run the new version against a sample of real tickets or deals and compare output quality, latency and per-call cost side by side with the model you're currently paying for. Anthropic and AWS have not published independently verified benchmark deltas for this release as of writing, so treat any capability claims as unconfirmed until you've tested against your own data. Companies not yet on Bedrock gain another reason to consolidate model access through AWS if they're already paying for EC2, S3 or other AWS services, since it simplifies billing and IAM permissions compared to managing a separate Anthropic API key.

AWS Shows How to Build a WhatsApp Ordering Bot on Bedrock AgentCore

For a B2B company that takes repeat orders over WhatsApp — distributors, wholesalers, food and beverage suppliers, spare-parts sellers — this removes a chunk of manual order-entry work: staff no longer have to read incoming messages or photos and key them into an order system. The catch is that this is a developer-facing reference architecture, not a packaged product, so it requires AWS engineering time to adapt to a specific catalog, ERP or CRM, and someone still needs to own exception handling for unclear photos, out-of-stock items or pricing disputes. Companies already running WhatsApp as an order channel should treat this as a build-vs-buy signal: the underlying capability is available on Bedrock now, so the cost of automating this workflow just dropped, but only for teams with cloud engineering capacity, not as a plug-and-play tool.

AWS Adds Access Controls for AI Agents Calling External Tools

If your company has connected an AI agent to your CRM, ticketing system, or internal APIs to automate sales outreach or support triage, that agent likely has more access than it needs and no audit trail of what it actually did. AgentCore Gateway lets you set per-tool permissions (e.g., an agent can read customer records but not modify billing) and get a log of every call, which matters the moment a customer asks what data an AI touched or a security review asks the same question. For a 10-200 person company without a dedicated security team, this shifts agent governance from a custom-built afterthought to a configuration you turn on, provided you're already on AWS or willing to route agent traffic through Bedrock.

AWS Shows How to Cut RAG Token Costs on Bedrock by Trimming Irrelevant Context

If your support bot, sales assistant, or internal knowledge search runs on a retrieval-augmented pipeline through Bedrock (or a similar architecture), the token bill scales with how much irrelevant context gets stuffed into every prompt — long documents, boilerplate, and near-duplicate passages you retrieve 'just in case.' Query-aware compression addresses that by filtering retrieved chunks against the actual question before they reach the model, which is the same lever that determines whether a 20-person support team's AI assistant costs $200 or $2,000 a month at scale. Teams already running RAG in production should treat this as a concrete cost-reduction checklist item, not a future upgrade — it requires no model swap, only a compression step inserted into the existing retrieval-to-generation pipeline.

AWS Lets AI Agents Pay Vendors Directly, No Human Click Required

For a 10-200 person B2B company, this closes a gap that has kept procurement and billing workflows partly manual: an agent handling vendor renewals, ad spend top-ups, or SaaS subscription changes can now execute the payment itself instead of routing to a person for card entry or approval. The practical move is not to hand agents a blank checkbook — it's to define hard spending caps, vendor allowlists, and transaction logging before connecting any payment-capable agent to a live account, then start with low-risk, recurring spend (subscription renewals, small supplier invoices) rather than open-ended purchasing.

AWS Lets Agent Builders Restrict Web Search to Approved, Recent Sources

For a 10-200 person B2B company running a support or sales agent that pulls live web results to answer customer questions, this closes a real gap: until now, an agent grounded in open web search could just as easily surface a three-year-old blog post or a competitor's page as your own documentation. Teams building on AgentCore can now lock search to a whitelist (docs.yourcompany.com, trusted partner sites, industry standards bodies) and require content published within a set window, which matters for anything involving pricing, compliance, or product specs that change often. It also gives ops and legal teams a concrete control to point to when a customer or auditor asks how the agent decided what to cite, rather than an unverifiable 'it searched the web.'

AWS Adds Cross-Region Routing for GPT-5.6 on Bedrock

If your support bot, lead-qualification agent, or ops automation calls GPT-5.6 through Amazon Bedrock, this removes a real operational headache: capacity crunches in a single region that cause dropped or delayed responses during peak hours. Instead of writing and maintaining your own retry-and-failover logic across regions, Bedrock now handles that routing for you, which means fewer 3am pages when a customer-facing AI workflow starts throttling. Teams running lean ops (10-200 people) rarely have spare engineering time to build resilience infrastructure themselves, so this is a case where the cloud provider absorbing that complexity is a direct, if modest, win for uptime of any AI-driven sales or support pipeline built on Bedrock.

AWS shows AI agents that can actually pay for things, not just recommend them

For a B2B company running 10-200 people, procurement, subscription renewal, and vendor payment tasks currently sit in someone's queue as an approval step because no automation layer was trusted to move money. This integration gives ops teams a concrete pattern for agents that can complete the transaction itself, e.g. renewing a SaaS subscription, paying a recurring vendor invoice, or restocking supplies, inside defined spend limits and authorization rules, collapsing a multi-step approval workflow into a monitored autonomous action.

AWS Shows How to Build Multi-Step AI Agents Without Custom Orchestration Code

For a 10-200 person B2B company, this matters less as a coding tutorial and more as a signal of what's now buyable versus what still needs building. If you're running sales development, tier-1 support, or order-to-cash operations, agentic workflows that check a CRM, pull an order status, escalate to a human, and remember context across a session are exactly the kind of task these tools target. The practical takeaway isn't "go build this yourself" — it's that the underlying primitives (session memory, tool invocation, identity-aware agents) are now standardized enough that a consultancy or vendor can assemble a working agent for a specific process in weeks rather than months. Ops leaders should ask any automation vendor pitching "AI agents" whether they're using managed infrastructure like this, since it affects reliability, security boundaries, and how fast changes can be made later.

AWS Lets AI Agents Click Through Old Web Apps That Have No API

Most 10-200 person B2B companies carry at least one legacy system with no API: an old order-management tool, a supplier portal, an internal ticketing app, or a vendor's dated admin console. Until now, automating around these meant either brittle custom scraping scripts, a costly system replacement, or accepting that someone on the team manually re-keys data between systems every day. AgentCore's Browser Tool gives a managed, sandboxed way for an AI agent to operate that old interface directly, essentially automating the human clicking-and-copying step without touching the underlying application. For an ops or support lead, this matters for a specific class of task: pulling status updates from a legacy tracking system into a CRM, filing renewals through an old vendor portal, or reconciling records across a system nobody wants to migrate. It doesn't replace a proper integration, but it closes the gap where integration isn't available or isn't worth building, and it's a capability worth flagging to whoever owns your process automation roadmap.

NVIDIA's Fast, Cheap Nemotron Model Lands on AWS SageMaker

If your ops or engineering team already runs on AWS, this matters less as a "new AI model" story and more as a procurement and latency story: one-click deployment inside SageMaker JumpStart cuts the integration overhead of adding a fast, lower-cost model to sales chatbots, support triage, or internal workflow automation. For a 10-200 person company, that's the difference between a two-week engineering sprint and an afternoon's work testing whether a lighter model handles ticket routing or lead qualification well enough to replace a pricier one. The catch: this is an AWS-specific convenience, not a universal capability shift — if you're not on AWS, or you don't yet have infrastructure to A/B test model swaps safely, there's nothing to act on here today beyond noting the option exists.

AWS Extends AgentCore Observability to On-Prem and Multi-Cloud

If your sales, support or ops team has AI agents running in different places — a chatbot hosted on AWS, an internal automation on a local server, a vendor tool on another cloud — you've probably had no single view of what's actually happening across them. This update means a 10-200 person company can now get one dashboard showing which agent handled which ticket, how long it took, and where it failed, regardless of where that agent lives. For lean ops teams without a dedicated platform engineer, that's the difference between debugging blind and having an actual audit trail when a customer complains an automated response was wrong or slow.

AWS Lets Developers Write Custom Reward Rules for Multi-Turn AI Agents

Most sales, support and ops teams don't train models from scratch, but many now run AI agents that handle multi-step interactions — qualifying a lead across several messages, resolving a support ticket through back-and-forth, or executing a multi-stage internal workflow. The core problem this AWS post addresses is real for those teams too: a single-turn "was this response good?" check misses whether an agent actually got the customer to a resolution, followed policy the whole way through, or avoided going in circles. If you're evaluating vendors or building custom agent logic, ask specifically how success is measured across the full interaction, not just per message — that distinction is exactly what reward function design is trying to fix, and it maps directly onto how you should be scoring your own agents' performance internally.