What Happened
Recent Google Cloud announcements highlight three connected requirements for production AI: asynchronous orchestration, governed data access and explicit unit economics.
- Orchestration at creative-production scale: Mito built an asynchronous Google Cloud platform to coordinate more than 5,000 assets, including 1,000 videos, across roughly 65 rounds of iteration for a 120-second Movistar Plus+ television spot. Virtual queues, bounded retries and defined job outcomes helped manage third-party API limits. A Gemini-powered evaluation harness checked outputs against user intent. [1]
- Enterprise analytics through managed connectors: Data Cloud connectors in Gemini for business, currently in preview, enable natural-language access to supported enterprise databases and analytics platforms. Queries execute under the user’s identity, with read-only connections and existing row- and column-level permissions. Google distinguishes these exploratory connectors from data-team-built Conversational Analytics agents for governed, multi-step analysis. [2]
- Lower-priced video generation: Gemini Omni 1.2 Flash is generally available, with generation and editing of clips up to 15 seconds. Listed pricing is $0.07 per second of 720p video, 30% below Omni 1.1 Flash. It supports lower-resolution drafts and upscaling to 1080p or 4K. [3]
Why It Matters to Businesses
Cheaper models do not automatically produce cheaper business outcomes. As generation becomes more accessible, teams can create more variants, run more evaluations and request more revisions. Mito’s workflow illustrates why managing intermediate assets and iteration becomes a platform problem—not simply an API integration. [1]
At the listed rate, 1,000 generated clips of 15 seconds each would cost $1,050 for 720p generation alone. That is an illustrative workload, not Mito’s reported spending. It excludes any additional charges for upscaling, evaluation, storage, orchestration and human review. Teams should track cost per approved deliverable, not just cost per generated second. [3]
For analytics, managed connectors can reduce custom authentication and integration work. However, permission-respecting SQL is not necessarily correct SQL. Business definitions, join logic and interpretation still require governance. Google’s distinction between ad hoc exploration and purpose-built analytics agents is therefore an important architecture boundary. [2]
Kimbodo Engineering Perspective
The central design choice is to separate model execution, workflow control and business approval. A model produces a candidate result; the platform controls execution, permissions and spending; an acceptance process determines whether that result is usable.
For long-running generation, asynchronous jobs are preferable to holding application requests open. Queues absorb bursts, but they do not create provider capacity. Concurrency limits and workload priorities must reflect actual quotas and delivery deadlines. Mito’s bounded retries and explicit outcomes are useful production patterns because unresolved jobs otherwise become both operational debt and unpredictable spending. [1]
For enterprise data, prefer managed connectors when supported engines, identity integration and access controls meet the requirement. Use custom MCP servers or other integration services when necessary capabilities are missing—not merely to reproduce managed functionality. Keep repeatable, decision-critical analysis behind governed agents with tested metrics and query behavior. [2]
Model-based evaluation can prioritize review, but should not be the sole acceptance gate. Creative quality, factual accuracy and business suitability require different checks.
How We Would Implement It
1. Establish a durable execution layer
On Google Cloud, we would place an authenticated application API in front of a durable job store, Cloud Tasks queues and Cloud Run workers. The API would return a job identifier immediately. Workers would invoke model providers and persist outputs in Cloud Storage, with metadata linking each asset to its prompt, model version, parent assets and approval state.
- Use separate queues and concurrency caps for each provider and workload class.
- Define explicit queued, running, succeeded, failed and cancelled states.
- Retry transient failures with backoff and attempt limits; route exhausted jobs to review.
- Persist provider operation identifiers and reconcile uncertain outcomes before resubmitting potentially billable work.
2. Put budget controls before generation
Estimate cost before submission, enforce project budgets and cap variants per request. Use draft resolutions where suitable, then generate or upscale approved candidates. Record generated duration, retry spending, approval rate and cost per accepted asset. Validate the economics of draft-to-final workflows against actual provider billing. [3]
3. Separate exploration from governed analytics
Configure read-only connectors with user-scoped identity and test permissions using representative roles. Define business terminology and approved metrics in catalog and semantic layers. Route recurring, high-impact questions to governed agents; apply query timeouts, scan limits where supported, and audit logging. [2]
4. Make evaluation and observability operational
Combine deterministic checks, model-based assessment and human approval according to risk. Trace each request through queue wait, provider execution, evaluation and review. Version prompts and evaluation criteria so regressions can be investigated rather than attributed vaguely to model behavior.
Risks, Costs and Security
- Retry amplification: A timeout does not prove generation failed. Blind retries can create duplicate assets and charges.
- Data exposure: Read-only access prevents writes, not inappropriate disclosure. Results exported into documents, email or slides need destination-sharing and retention controls. [2]
- Preview dependencies: Confirm connector availability, support terms and limitations before making preview capabilities part of a critical production path. [2]
- Query expense and correctness: Authorized queries can still scan excessive data or return misleading conclusions. Test both resource consumption and semantic accuracy.
- Media rights and provenance: Omni outputs include SynthID watermarks, and audio references require allowlisting. Neither replaces rights clearance, consent checks or brand review. [3]
- Platform overhead: Include evaluation, storage lifecycle, network transfer, observability and human review in the business case—not only model API charges.
The practical objective is a controlled path from request to approved outcome: durable execution, least-privilege access, bounded spending and evidence that the result meets business requirements.
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our AI Infrastructure & MLOps practice, or Estimate My Infrastructure.
