Skip to content
Illustrative scenario

Not based on a specific customer deployment. The company, figures, decisions, and outcomes below are a constructed example that shows how Intended’s mechanism applies to this kind of problem. They are not measured results from a named customer. Where we publish a real, attributed customer outcome, we will say so explicitly.

← All case studies
Technology

Gating Unauthorized Production Deployments by AI Agents

Consider a large engineering org running many daily deployments across dozens of microservices, where AI agents drive CI/CD pipelines, rollbacks, and infrastructure scaling decisions.

01 · the challenge

The kind of problem this addresses.

Incidents
production outages from AI-initiated deployments that shipped without review
$$$
cost per incident including downtime and remediation
0%
of AI deployment decisions tied to a policy-evaluation audit trail

02 · how it works

See the difference.

Code Push

Build

Test

Intended Gate

Deploy

Monitor

03 · the solution

What they deployed.

  • Installed the SDLC domain pack (OI-100) with 44 deployment and infrastructure intent classifications
  • Integrated Intended as a GitHub Action in every deployment workflow
  • Deployed Kubernetes admission controller for infrastructure changes
  • Configured thresholds: staging auto-approve, production requires risk < 40 or human approval
  • Enabled Slack + PagerDuty escalation for high-risk production deployments

04 · implementation

From zero to governed.

Week 1

Connect

Added Intended GitHub Action to all 40 microservice repos. Installed Kubernetes admission controller.

Week 2

Configure

Installed SDLC domain pack. Defined staging vs. production policies, risk thresholds, and team-based approval chains.

Week 3

Shadow

Ran in shadow mode across all pipelines. Tuned risk scoring to match team expectations. Zero false positives in final 3 days.

Week 4

Enforce

Enabled enforcement. All 200+ daily deployments now flow through Intended authority gates.

05 · illustrative outcomes

What this is designed to deliver.

Modeled figures for this scenario — what the workflow above is built to achieve, not measured results from a named customer.

0%

Production deploys gated

No deploy without an authority token

<0ms

Latency added per deploy

Design target — imperceptible in the pipeline

0%

Deployment decisions auditable

With cryptographic proof

0

Deployment intents classified

Via the SDLC domain pack

06 · decision replay

Example decisions, full trace.

Sample decision records that show the shape of the evidence Intended produces. Illustrative, not drawn from a live customer’s logs.

2026-03-15 08:22:11sdlc.deploy.stagingRISK: 12/100ALLOW34ms

Deploy cart-service v2.14.3 to staging environment

Resolved by: Policy: auto-approve staging deployments

2026-03-15 09:45:33sdlc.deploy.productionRISK: 78/100ESCALATE28ms

Deploy payment-service v3.8.1 to production (database migration included)

Resolved by: Lead Engineer (approved in 6m 44s after review)

2026-03-15 10:12:07sdlc.deploy.productionRISK: 22/100ALLOW31ms

Deploy search-service v1.22.0 to production (config change only)

Resolved by: Policy: auto-approve config-only deploys with risk < 40

2026-03-15 14:33:18sdlc.infra.scale-clusterRISK: 65/100ESCALATE26ms

Scale checkout cluster from 8 to 24 nodes (Black Friday prep)

Resolved by: SRE Lead (approved in 2m 11s)

2026-03-15 16:01:44sdlc.deploy.rollbackRISK: 45/100ALLOW22ms

Rollback inventory-service from v4.1.0 to v4.0.9 (error rate spike)

Resolved by: Policy: auto-approve rollbacks to known-good version

the takeaway

The point of a deployment gate is proof: every production deploy is reviewed against policy before it ships, and that review leaves a signed, auditable record. Authority tokens turn an implicit trust assumption into something an SRE or a CISO can verify.

Why this pattern matters — not a customer quote.

Start governing AI deployments

Free to start. See every AI decision from day one.

Intended — Intent Verification Infrastructure for Autonomous Agents