The assistant is getting an inbox of its own. On November 11, Databricks announced an always-on Genie designed to keep enterprise work moving and ask people for decisions. The important small print: the new agent is coming to preview and beta, not generally available today. That gap between a compelling demonstration and something you can safely delegate to is the thread running through AI news today.
If you build or buy these systems, yesterday’s useful question was less “How clever is the model?” and more “Where does it stop?” Google is narrowing AI down to bounded choices. ElevenLabs is putting voice agents inside government infrastructure. Adobe is keeping generated sound editable rather than flattening it into a finished answer. Meanwhile, Indonesia’s copyright debate asks who gets to set the boundaries in the first place. Here are five developments from November 11, with the shipping products separated from the promises.
AI news today: Databricks gives Genie work to do between questions
Databricks’ announcement describes a second-generation Genie that manages long-running goals, follows up on missing inputs and surfaces decisions through a priority inbox and email. Its business context comes from Genie Ontology: definitions, metrics, people and processes. The company also announced Genie Search across more than 100 enterprise sources, including Google Drive, Jira and Confluence.
Existing Genie and Genie Ontology are generally available. The new always-on agent, Genie Search and Ontology Manager are coming soon to preview and beta programs. Databricks says Unity Catalog and Unity Gateway govern permissions, approvals and budgets.
The buying test is straightforward: can it distinguish a useful follow-up from an unauthorized action? Ask for an audit trail of one complete task, including what the agent refused to do. More autonomy makes shared definitions and explicit spending limits prerequisites, not cleanup work after the pilot.
Google’s Gemini Decisions API trades essays for bounded choices
Google introduced the Gemini Decisions API on November 11, powered by Gemini Decision 1. Rather than writing an answer token by token, it scores candidate options. A request can contain up to 1,000 independent yes/no, multiple-choice or rubric questions against a shared multimodal input, with a context limit of 256K tokens.
Google lists a price of $0.10 per million input tokens and no output-token charge. Access is through the Gemini API in Google AI Studio. That is a materially different proposition from paying a conversational model to explain every routing decision; our recent model-pricing analysis provides a useful comparison of the broader cost question.
For builders, the attraction is a smaller decision surface: choose a tool, grade a document or escalate an ambiguous case. But a confidence score is not a guarantee. Before putting it in a control path, test false approvals and false rejections on your own examples. Watch whether cheaper decisions reduce the whole workflow’s cost, not merely one API line item.
ElevenLabs and NVIDIA take voice agents into sovereign infrastructure
In their November 11 collaboration update, ElevenLabs said NVIDIA Nemotron 3.5 Super VL is now available in ElevenAgents. It also outlined plans to add 15 megawatts of Vera Rubin capacity on top of existing Blackwell Ultra systems. That additional capacity is a plan, not a statement that the new infrastructure is already operating.
The more concrete deployment is Poland’s Centralna e-Rejestracja appointment system. According to ElevenLabs, its voice technology and NVIDIA infrastructure support an agent, delivered with EasyCall, that calls landline users to remind them about appointments and help reschedule. The company says it runs on-premises with NVIDIA Confidential Computing, keeping citizen data and audio within government infrastructure.
That makes voice AI a deployment and service-design story, not just a nicer synthetic voice. Public-sector buyers should ask about human handoff, accessibility and failed-call handling alongside data location. The announcement does not establish a measured reduction in missed appointments. Actual service outcomes are the next evidence to demand.
Adobe shows why generated sound needs separate tracks
Adobe Research’s November 11 technical write-up explains Generate Soundscape in Premiere Desktop: video understanding proposes synchronized audio layers directly in the editing timeline. Its examples separate background ambience from foreground effects, with individual sounds remaining editable. This is a new explanation of the technology and workflow, not a claim that every related Adobe MAX announcement happened yesterday.
The distinction matters if you make video professionally. A plausible composite soundtrack can still be wrong in one place. Separate tracks let an editor replace an effect, adjust its timing or change its level without throwing away everything else. Adobe describes training and listener evaluation, including feedback from professional sound designers; that is company-reported development work, not an independent quality verdict.
Watch the awkward cases: overlapping actions, subtle footsteps and sounds whose timing carries the scene. The useful benchmark is how much correction an editor needs before delivery. Generation speed alone does not tell you whether a tool saves production time.
Indonesia’s copyright hearing puts links and AI disclosure in dispute
Reuters reported on November 11 that Meta and TikTok joined Google in opposing proposed changes to Indonesia’s copyright law at a parliamentary hearing. The proposals concern payments for displaying news content and rules for AI. Google objected to compensation for news links; Meta distinguished referrals from republication. TikTok argued disclosure should distinguish AI assistance from fully AI-created content and urged preservation of safe-harbour protections. These are company positions on a proposal, not enacted requirements.
For publishers and platform teams, those distinctions change the practical question: is a system linking to an article, reproducing it, training on it or helping someone edit their own work? Treating those activities as interchangeable makes compliance planning less useful. Our earlier briefing on training opt-outs offers related context, not a statement of Indonesian law.
The next thing to inspect is the revised legislative text. Until then, map the content uses in your product rather than assuming yesterday’s objections settled the outcome.
Watch & Learn
Watch: “What is a Knowledge Graph?” — IBM Technology (5 minutes 36 seconds). Martin Keen’s lightboard lesson explains entities and their relationships, then shows why that structure makes data more understandable to machines. It is a useful starting point for leaders and builders trying to understand the context layer behind an enterprise agent. Watch for the difference between collecting documents and explicitly connecting concepts.
AI, Translated: Ontology
An ontology is an explicit model of the kinds of things in a domain and how they relate. In a subscription business, it might define customers, contracts and renewals, then connect each renewal to its contract and account owner. It can also encode the agreed meaning of a metric rather than leaving “revenue” open to interpretation. Think of it as shared business structure, not a pile of files. You should care because an agent can retrieve the correct record and still act on the wrong definition.
Try This Today: Make ChatGPT challenge your metric definitions
Goal: build a small, reviewable metric dictionary in about ten minutes. This uses an ordinary text conversation, not a special agent feature. OpenAI’s prompting guidance recommends specifying the goal, useful context, output and boundaries, then refining the result.
1. Start a new ChatGPT chat with three fictional metric definitions, or approved non-sensitive examples.
2. Ask for conflicting interpretations and concrete edge cases.
3. Correct the draft, then ask it to test the revised definitions against those same cases. Keep an “unresolved” column instead of letting it invent policy.
Copy-ready prompt: “Help me review a metric dictionary. Our fictional company uses booked revenue, recognized revenue and active customer. For each, propose a draft definition, one example, one counterexample and two questions for the metric owner. Label assumptions. Do not invent accounting policy or treat proposals as approved. Return a compact table, then test whether one customer could be counted differently under competing definitions.”
One thing to remember
The next useful AI advantage may be less about a model doing more and more about your organization specifying exactly what it means, what it may do and when a human must decide.