Skip to content
SID Global Solutions is a Google Cloud PartnerCareers
Google Cloud

Ten principles · API platform and AI agent readiness

What Gets AI to Production: the expanded principles

These ten principles come from the field. They draw on production Apigee deployments, where governance made the difference between a demo and a live system. They are written for the architect who has to make the argument in the room, not just in the build.

Printed from the poster at the Google Cloud Apigee Roadshow, New York. Get the PDF

01

A pilot that cannot be audited is a demo.

Field noteA governance reviewer does not ask whether the model performed well in testing. They ask whether a specific decision made six months ago can be reconstructed from logs. Systems that cannot answer that question in production do not pass review. They restart. Build the audit trail before the model, not after it ships.

02

Pick the workflow, not the use case.

Field noteUse cases run in demos. Workflows run in production. The difference is ownership: a workflow has a named human owner at each handoff, a deadline that the business tracks, and an audit trail the compliance team can read. Attaching AI to a named workflow produces return. Attaching it to a capability produces a proof of concept.

03

Put the gateway in front of the agents before you build the agents.

Field noteApigee AI Gateway is not an add-on. It is the point where rate limiting, token quotas, authentication, and observability are enforced across every agent that runs through it. Teams that retrofit the gateway after the agents are built spend more time on integration than on capability. Set it up on day one and the rest of the build works toward it.

04

Measure cost per completed task, not per model call.

Field noteModel calls are a purchasing metric, not a performance metric. What the business cares about is how many tasks the agent completed and what each one cost. Teams that optimize for call volume end up with fast agents that do not finish things. Define task completion first, then build the agent that hits it at a defensible cost.

05

Run the model where the data lives.

Field noteThe most common architectural mistake is designing the agent first and solving data access second. Vertex AI and BigQuery are built to analyze data in place, which removes the data movement problem before it starts. Design the data access layer in sprint one. The model follows the data, not the other way around.

06

Design for the review cycle before you write the first prompt.

Field noteGovernance and legal review are not a phase at the end of delivery. They are a constraint that shapes every design decision from the first sprint. Architects who treat review as a final gate find that the gate changes the design. Architects who treat review requirements as inputs build systems that pass on the first submission.

07

Audit the infrastructure before you build the business case.

Field noteMost enterprise AI budgets compete with existing infrastructure spend. The fastest path to approval is to show that the AI workload can be funded by redirecting capacity that is already approved. An audit of middleware, on-premise compute, and redundant tooling can surface capacity that covers a pilot without a new budget line.

08

Name one person who owns the next decision, or nothing ships.

Field noteStalled AI programs share one pattern: the next decision has no named owner. This is a meeting structure problem more than a culture problem. At the end of every sprint review, before the session closes, confirm one name, one decision, and one date. If that answer is not available, the sprint has not ended.

09

The platform is ready. Design the integration layer to match it.

Field noteGemini Enterprise, Vertex AI, and BigQuery are production-grade systems. What delays enterprise readiness is the integration layer around them: access control that was not designed for agents, schema governance that was written for batch systems, and data lineage that stops at the API boundary. Address those and the platform delivers what it promises.

10

Governed AI reaches production. Ungoverned AI reaches review.

Field noteThe speed argument for ungoverned development is real in the short term. An ungoverned pilot ships in days. It also stalls at the first compliance checkpoint and restarts under a governance requirement that was always going to be there. Building to governance from sprint one adds discipline to each sprint and removes the reset at the end. The governed route is the faster route to production.

Take it with you

Send me the expanded version as a PDF

Name and work email is all we ask. You will get the ten principles with their field notes in one document you can hand to the people who sign off on your roadmap.

More on how we work with Google Cloud.

Request the PDF

We will email it to you. No newsletter unless you ask for one.


    SID Global Solutions · Google Cloud PartnerGoverned AI. Built for Production.
    Contact us

    Get answers to your questions

      Upload file

      File requirements: pdf, ppt, jpeg, jpg, png; Max size:10mb