ROOT FORWARD / NOVEMBER 9, 2026
Where AI is going—and what you can do with it.
A hosted model opens a test lane, Google gives Workspace MCP clients a deadline, and Taxbit shows how support answers can follow the caller’s role. A draft evaluation framework and an analog audio kit round out the issue with two ways to measure what matters in use.
A good evaluation tells you who can act, what happens when the input changes, and what evidence is left behind.
Root Note
01 / MODELS / HOSTED INFERENCE
LG opens a hosted test lane for K-EXAONE 2.5
LG AI Research opened a no-fee hosted API for K-EXAONE 2.5 on November 9. LG describes it as a technology-validation model that adds training to its August release. An EXAONE API Platform key gives developers access to a text-only model via an OpenAI-compatible chat-completions endpoint. The platform enforces request limits. That makes a comparison possible without setting up inference hardware.
LG reports its own results on coding, browsing and tool-use evaluations, including SWE-bench Pro and BrowseComp. Those numbers are starting points for a comparison. In an agent workflow, the tool definitions, context, retry policy and criteria for accepting an answer all influence how much usable work comes back.
Root Access Take · Free calls lower the cost of starting a test. Count the time a technician spends checking and repairing the result, too. A model that needs a second pass on every ticket may be inexpensive to call and expensive to operate.
Try next · Run the same small set of synthetic ticket summaries and code-review prompts against two approved endpoints. Record correct suggested tool choices, parseable output, rate-limit responses and minutes of human rework. Check the provider’s current usage and data terms before introducing client information.
Sources · LG AI Research release · K-EXAONE 2.5 API documentation · Rate-limit documentation
02 / INTEGRATIONS / AGENT IDENTITY
Google Workspace MCP clients have a November 11 deadline
Google’s October 12 Workspace release notes set November 11 as the start of enforcement for a dedicated OAuth client setting. User-delegated clients calling Google Workspace Model Context Protocol servers need to be marked for use by an AI-powered agent. The notice says clients left unmarked by November 12 will need a new marked client to call those MCP servers. The affected servers are part of Google’s Developer Preview program.
Google’s setup guide, updated November 6, advises splitting the client if it also serves regular app features. REST-only integrations are outside this requirement. Service accounts using domain-wide delegation follow a separate administrator-controlled path, so first identify the authentication flow your connector actually uses.
Root Access Take · A shared OAuth client is convenient until two workflows need different consent, scope or change windows. Giving the agent its own client lets the service desk trace a failed request to the right configuration and plan a rollback without touching routine integrations.
Try next · Inventory preview MCP integrations for each managed tenant. Record the owner, OAuth client, redirect URIs, scopes and a low-impact test call. Stage a separate marked client for any mixed-use app, confirm consent and access in a test tenant, and verify the ordinary REST workflow still succeeds.
Sources · Workspace developer release notes · Google’s OAuth and MCP setup guide
03 / MSP PRACTICE / SERVICE DESIGN
Taxbit makes the help desk answer for the caller’s role
Taxbit and AWS described an embedded support assistant on November 6. Taxbit Assist uses Amazon Bedrock AgentCore and a curated knowledge base to answer questions about the Taxbit product with citations. The calling service binds the authenticated organization and dashboard role into the tools. An organization entitlement check and a dashboard role check can keep the assistant from directing a user to a workflow they cannot open.
Taxbit also adds document titles and scope context to retrieved chunks. That preserves the distinction between passages that look alike but apply in different jurisdictions or product configurations. The case study describes product-knowledge answers today. Querying a customer’s reporting data is identified as future work.
Root Access Take · The pattern travels well to an MSP help desk. A cited answer can still be wrong for the caller. A runbook fragment should travel with its client or product scope, prerequisites and escalation route, and the application should supply the caller identity from its authenticated session.
Try next · Pick one support topic with different instructions for an admin and a standard user. Run both identities against the same question in a sandbox. Check that each answer cites the correct runbook and that an unsupported role gets a clear handoff to a person.
Sources · Taxbit and AWS implementation case study
04 / GOVERNANCE / EVALUATION
NIST’s draft shows how to build an evaluation around a job
NIST’s August 7 initial public draft of AI 200-2 describes a four-stage way to design custom testing, evaluation, verification and validation. Its TEVV-Athlon organizes an assessment around what an organization needs to measure, the events and tools that produce evidence, and the data those events capture. The framework is designed to apply across systems, including statistical models, multimodal models and agents.
The draft comment period ran through October 6. This is draft evaluation guidance. NIST’s voluntary AI Risk Management Framework provides broader context. ISO/IEC 42001:2023 specifies AI management system requirements, and ISO/IEC 23894:2023 offers risk management guidance. These documents serve different roles in a supplier review.
Root Access Take · A generic score is useful for sorting a shortlist. The deployment decision needs evidence tied to the service you intend to offer. A ticket summarizer that performs well on short requests can still omit the exception that changes which technician should act.
Try next · Write three test events for an intake assistant. Give it a clean ticket, a ticket with contradictory notes and one with no reliable answer. Set the expected outcome, the reviewer and the evidence to keep for each event. Record the model version and configuration so the team can rerun the same exercise after a change.
Sources · NIST draft announcement and scope · NIST AI RMF · ISO/IEC 42001 · ISO/IEC 23894
05 / EDGE HARDWARE / AUDIO
Blumind puts analog audio inference on the test bench
Blumind announced availability of the BM110 Evaluation Kit on November 9. It includes an analog microphone, a board with the BM110 chip, software tools and reference models for voice activity detection and keyword spotting. The board includes a way to measure current during inference. Blumind’s product page directs qualified teams working on high-volume consumer or industrial applications to contact the manufacturer.
The chip stores neural-network weights in the analog compute fabric and runs a small model near the sensor. Blumind describes a workflow that starts with training in PyTorch or TensorFlow and uses its translator to prepare weights for the chip. That makes the kit relevant to engineers evaluating a battery-powered device that must listen for an event for long periods.
Root Access Take · A low-power detector only pays off if it wakes the rest of the system at the right moments. A sensitive detector can trade a small idle load for a surprising number of expensive wakeups. Real room noise belongs in the test plan.
Try next · For a design evaluation, compare the existing digital baseline with the kit on identical audio. Log system current, missed triggers and false wakes in quiet and noisy settings. Include the microphone and downstream processor in the power budget, and ask which parts of the model conversion and deployment flow your team can reproduce.
Sources · Blumind’s dated kit announcement · BM110 product and evaluation kit
Explore the Root Forward archive, AI Radar, Governance guide and The Root Shelf.