Microsoft enables the main Azure OpenAI access setting by default, while provider access, two cross-region processing controls and storage remain separate. Approval should record five settings across four decisions.
Montréal and Toronto are now supported for conversational analytics. The deployment decision is permanent per agent, and heterogeneous source locations trigger a global fallback unless teams choose a location explicitly.
AWS, Snowflake, Google Cloud and Oracle expose different failure boundaries around data, tools, identity and execution destinations. A deployable agent version must name all of them.
Cortex Agents defaults to continuing with accessible tools, but its authorization pre-check does not cover SQL, code execution or custom functions. Production clients need an explicit completeness policy.
Microsoft’s sharing model means an agent user can query a semantic model without Build permission or workspace membership. Review item-level grants, not just workspace roles.
Databricks now exposes a user’s group memberships to Genie One. Treat that context as a presentation and routing signal; keep authorization in Unity Catalog.
Google’s Data Agent Kit separates where Spark code runs from where notebook results can go. Security reviews need to test both boundaries.
Microsoft’s ALM format makes instructions, examples and schema scope reviewable. The catch: table selection is enforced, while column-level selection is not yet supported.
Cube’s production notes show why zero rows, truncated results and failed saves must be different machine-readable outcomes.
Cross-schema reuse narrows who can edit an agent team, while every consumer still needs the database and tool permissions required at runtime.
Databricks, Google Cloud and Microsoft expose different shutdown boundaries for unattended analytics: code execution, orchestration identity and capacity consumption.
AWS’s internal Field Sales system treats data, tools, prompts and models as one releasable unit, then keeps testing after deployment.
Snowflake says semantic views and verified queries carry over unchanged. Teams should still regression-test routing, permissions and partial-tool access before switching production traffic.
Microsoft’s billing documentation separates model-token consumption from generated-query execution, while a worked example shows why teams should test both before rollout.
Promethium’s five-category framework is useful for procurement, but its own preferred category shows why buyers should separate architecture facts from vendor scoring.
Microsoft’s preview turns business definitions into searchable context. That makes retrieval quality a separate production test, not just a prompt-writing task.
A pending routing change improves sign-in locality, but enterprises with filtered public paths must discover, allow and test an additional account-specific hostname.
Teams have until November 6 to move to Fabric IQ, gaining multi-agent catalog discovery and MCP endpoints while keeping cross-geo and identity constraints.
The vendor’s new buyer guide is self-interested, but its proposed pilot checks are concrete: test identity passthrough, blocked writes, semantic grounding, concurrency and warehouse cost.
The new monitoring guide makes refresh status queryable through metadata and event logs. That solves observability, not the decision about whether an agent should answer from stale data.
Scoped OAuth secrets reduce API blast radius for unattended agents and analytics apps, yet migration is explicit: scopes are immutable, existing credentials keep all-APIs access, and one scope can mint broader tokens.
A real internal app moved from database-bound to application-bound as compute grew, making cache fit and worker count separate production decisions.
Google’s streamlined route removes copied A2A cards and hand-built OAuth credentials in the simplest topology—but cross-project deployments still need manual registration.
Google’s own deployment guidance treats BigQuery and Cloud Storage content as a prompt-injection surface and recommends network, identity and model-level containment.
Data Chat Mini shows the control stack around generated SQL: constrained MCP tools, bounded loops, query telemetry and a separate application-access boundary.
19c retains core natural-language querying and agent tools, while 26ai adds feedback, automatic schema selection and vector-backed retrieval.
CoCo now carries discovery through comparison, trial and installation. The important deployment detail is that conversation changes the interface, not the authorization model.
The 2026_08 preview targets 12 old functions only in accounts with no use during a fixed 90-day window, while several replacements change outputs or failure behavior.
Microsoft separates agent lifecycle management from runtime consumption, while its Responses migration changes conversation state, polling and trace inspection.
A new practitioner checklist turns Microsoft’s model-preparation guidance into an operational rule: fix relationships and measures first, then spend scarce description space on meaning, units and exclusions.
The new control split lets enterprises enable Explore conversations, dashboard agents and Python-backed advanced analysis independently—but permissions still determine who can reach each model.
Knowledge Assistant, Supervisor Agent and the older agent-server path are now legacy; the replacement depends on whether a team wants low-code analytics or code-first orchestration.
Prompt caching changes the economics of data-agent context: stable definitions belong at the front, volatile user questions at the back, and cache hits must be measured rather than assumed.
Apache Superset’s built-in MCP server combines semantic queries, chart and dashboard authoring, and RBAC-enforced SQL. Its own deployment guide also documents the operational limits teams must design around.
Microsoft’s integration preserves Fabric permissions in one mode and deliberately bypasses each user’s source access in another. The compliance boundary can move too.
Red Hat’s client-credentials proxy shows how unattended agents reach protected tools, and why machine authentication is not the same as per-user authorization.
Microsoft’s December 2026 deprecation creates a migration project around semantic models, verified answers, permissions and regression tests—not a simple visual swap.
Source links show where context came from; they do not validate metric selection, generated queries or chart transformations.
Exact match, execution accuracy, semantic tests, functional checks and speed answer different questions; a credible evaluation names the failure each metric can miss.
The new SQL-dataset materialization path can cut repeated warehouse work, but it changes freshness, permission and availability assumptions that teams need to test before rollout.
The release is production-labelled; its tool-selection and tool-execution scoring still carry private-preview labels. Procurement and test plans should separate those claims.
The Oct. 1 release note names a portable chat export; the linked operating guide still documents only pinning and renaming conversations.
A Microsoft consultant’s personal, open-source build shows a practical evaluation order: execute the reference and generated SQL in the same sandbox before asking an LLM to judge the prose.
Trinetro’s setup guide gets the most important control right: an analytics assistant should be unable to write, not merely instructed not to.
Spotter’s progressive rendering exposes a useful production boundary: deliver correct axes and data first, then let an isolated worker add labels, styling and annotations.
Decision provenance links the question, chosen meaning, analytical plan, evidence and reviewer disposition without pretending the result was permanently correct.
Db2 Recovery Expert 12.1 adds API-based archive-log access, KMIP support and tenant selection. For data-platform teams, the practical lesson is that managed-cloud recovery needs an explicit control path.
Default-deny is only the starting promise. The useful test is whether the system refuses questions that require blocked tables and cleans up connections after sign-out.
Trinetro’s personal-versus-company-default split suggests a practical acceptance test for keeping one user’s interpretation from becoming everyone’s metric.
The useful lesson in Domo’s case study is not “spreadsheet becomes chatbot.” It is that business aliases, permissions and user acceptance had to move with the data.