ROOT FORWARD / NOVEMBER 12, 2026
Where AI is going—and what you can do with it.
Two coding models get a new place to run, and Databricks brings business data into more assistants. DXC shows where it puts approval, IBM asks whether companies can prove they could switch providers, and Google makes it easier to try speech offline.
An exit plan belongs in the design review, while the system still works.
Root Note
01 / MODELS / DEPLOYMENT
Two Ornith coding models land in JumpStart
AWS added DeepReinforce AI’s Ornith-1.0-9B and Ornith-1.0-35B to the SageMaker JumpStart catalog on November 11. The models were already available as weights under the MIT license. The new option is a managed deployment path through the SageMaker console or Python SDK. Check the model card in your account for supported instance types, regional availability and quota before planning a deployment.
The 9B is a dense model. The 35B uses a mixture-of-experts architecture, activating a subset of its parameters for each token. Both were post-trained from Qwen 3.5 for coding workflows. The publisher supplies benchmark results for particular agent and model configurations. Those scores depend on the surrounding tools and test conditions, so they give you a starting point for comparison.
Root Access Take · The smallest model that gets a change through review is often the better service. Measure the cost of running an endpoint through quiet hours as well as the cost of inference under load. An idle GPU can outspend an impressive benchmark result.
Try next · Deploy one model at a time in a sandbox account. Use the same small set of documentation and unit-test tasks, tool permissions and acceptance checks. Track review time, passed tests and endpoint cost, then tear down the endpoint.
Sources · AWS availability notice · Ornith 9B model card · Ornith 35B model card
02 / INTEGRATIONS / BUSINESS DATA
Databricks Genie brings the same data into more assistants
Databricks announced native Genie integrations for ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot on November 11. Other compatible clients can connect through Genie One’s Model Context Protocol server. The point is to let someone ask a business question from the assistant they already use while drawing on the data and business context managed in Databricks.
The setup depends on the client. Databricks documents administrator setup followed by user sign-in for ChatGPT and Claude, and notes that restricted workspaces may also need an ingress policy that allows the client. Teams with an existing custom MCP client have a separate deadline. Databricks plans to retire the beta endpoint on November 30. The generally available Unity Gateway endpoint uses the ai-gateway scope for on-behalf-of OAuth, and callers need execute permission on the server.
Root Access Take · Decide what “open ticket” means before adding four new places to ask. A shared data source is useful when the metric definition, business context and allowed data travel with it. Test those details from the user’s seat.
Try next · Pick a trusted count from one Genie space. Compare the answer from two supported clients under the same user, then try a test account with narrower data access. Include one question with an ambiguous term and check how each client handles it.
Sources · Databricks integration announcement · Genie One MCP setup and migration
03 / MSP PRACTICE / SERVICE DESIGN
DXC puts authority outside the model
AWS and DXC published an architecture walkthrough for DXC OASIS on November 11. The platform uses Amazon Bedrock as a foundation for model access and a separate layer for orchestration and operating decisions. DXC’s layer adds context from runbooks and service-management data, directs work to specialized agents and tracks actions with execution IDs.
The case study describes four authority levels, from observing and proposing to acting with approval or within a predefined scope. New personas start with read-only access. The walkthrough follows a fictional airline in a demo environment. One example reviews service-ticket patterns, while another summarizes cloud costs. The illustrations show how DXC designed the workflow. A pilot in another environment still needs its own evidence about performance and fit.
Root Access Take · Give the agent authority through the service layer, where you can enforce it and record what happened. A model switch should leave a technician’s approval requirements intact. The same applies to customer boundaries and to the point at which the job returns to the queue.
Try next · Take one read-only weekly service review. Write down the inputs it may see, the output format and the person who accepts it. Run it with a complete ticket export and again with missing fields, then check that the trace explains the difference.
Sources · AWS and DXC architecture case study
04 / GOVERNANCE / PORTABILITY
IBM’s sovereignty survey makes exit a testable question
IBM released its Digital Sovereignty Index on November 11. Its research included 2,003 senior decision-makers at large enterprises across seven countries, plus qualitative interviews. It looked at how organizations demonstrate control across AI, technology, operations and data. One dimension asks whether a team can show it could change providers or recover operations when required.
IBM reports that about one-third of survey respondents regularly or continuously test their ability to demonstrate sovereignty. That describes the surveyed enterprises, using IBM’s methodology. For a service provider, the useful move is to make an exit test concrete. A data export matters, as do the permissions, audit history, configuration and people needed to make that export usable.
Root Access Take · Ask the vendor to show the shape of an exit before you add the next dependency. A working copy of the data is only one piece of a service you can restore. The support team also needs to know who can act and what evidence survives the move.
Try next · Pick one AI workflow and rehearse a provider outage with synthetic data. Time how long a second operator takes to find the inputs, export the required records and run the manual fallback. NIST’s voluntary AI RMF and ISO/IEC 23894:2023 offer risk guidance. ISO/IEC 42001:2023 sets AI management system requirements.
Sources · IBM index findings and methodology · NIST AI RMF · ISO/IEC 23894 · ISO/IEC 42001
05 / LOCAL AI / ACCESSIBILITY
Google AI Edge Gallery adds offline speech in both directions
Google added on-device speech recognition and text-to-speech to AI Edge Gallery on November 10. With the needed models or language packs in place, the local pipeline can transcribe speech and read a response aloud offline. The gallery lets people try voice-enabled chat and shows a read-aloud interface that highlights the current sentence.
For developers, Google describes two paths. On supported Android hardware, ML Kit can use system-managed models. MediaPipe Tasks and LiteRT provide a path with more control over the speech models and target devices. The gallery lists examples such as Parakeet for recognition and PocketTTS for playback. Model choices affect which languages and voices are available. The project’s example code and each model’s terms are worth checking independently before reuse.
Root Access Take · Offline is a useful failure mode to design for. A field tech or a person using a voice interface still needs a way to edit a transcription, repeat a sentence and fall back to text. The app gets more dependable when those controls work in the same places as voice.
Try next · Set up the required model and language pack on the actual device, then switch to airplane mode. Try a quiet room and background noise. Check recognition of names and ticket numbers, playback timing and how easily someone can correct the text.
Sources · Google AI Edge speech announcement · Google AI Edge Gallery project
Explore the Root Forward archive, AI Radar, Governance guide and The Root Shelf.