The Role of AI in Azure and AWS DevOps

Hijab's profile picture
Hijab e FatimaLast Updated on
18-19 Min Read TimeAdd as preferred on Google

Key Takeaways

Predictive scaling is the lowest-risk first AI step on Azure or AWS, while several AWS code review tools are retired or ending support by 2027.

  • AWS CodeGuru Reviewer has taken no new repository associations since November 7, 2025, and Amazon Forecast is closed to new customers.
  • Amazon Q Developer is also ending, with GitHub app support gone on January 30, 2027 and IDE plugins and paid subscriptions ending on April 30, 2027.
  • Azure and AWS predictive scaling both scale out only, leaving scale-in and unexpected spikes to standard rules.
  • AWS predictive scaling accepts custom metrics and runs in GovCloud, while Azure's version covers only Percentage CPU on virtual machine scale sets in the commercial cloud.
  • Both clouds offer a forecast-only mode for a two-week check of predicted against actual load.
  • Azure machine learning alert thresholds start from 10 days of history; weekly seasonality takes three weeks.

AWS teams should choose a code review replacement before January 30, 2027.

AI in Azure and AWS DevOps pipelines helps teams predict infrastructure demand, detect unusual activity, and automate parts of the development and deployment process. Azure uses services such as Azure Machine Learning and Azure Monitor for model deployment, predictive scaling, and intelligent alerting, while AWS offers EC2 predictive scaling and CloudTrail Insights for capacity forecasting and anomaly detection.

 

However, these capabilities come with different requirements, limitations, and service availability. With several AWS developer tools being phased out or approaching end of support in 2027, choosing the right AI-powered DevOps tools requires careful evaluation.

 

Where does AI help in Azure and AWS DevOps pipelines?

AI is worth adding to a DevOps pipeline when a model has enough history to learn from and can take over a decision engineers would otherwise make by hand. That happens at four stages.

 

  • Less manual tuning. Machine learning sets alert thresholds and scaling triggers that engineers otherwise adjust by trial and error, as part of wider software development automation.
  • Fewer defects in production. Automated code review and scanning flag security and quality issues before a release ships.
  • Earlier warning of unusual activity. Anomaly detection learns a baseline and flags behavior that departs from it. It complements dedicated cybersecurity services.
  • Capacity that follows demand. Predictive scaling adds instances ahead of a forecast peak and can reduce over-provisioning.

 

Be wary of headline figures like "100 times faster deployments". Results depend on your own traffic patterns and on how much clean history your metrics hold.

 

Unsure which platform is the right fit for your DevOps needs?

Download our AI in Azure vs. AWS DevOps Cheat Sheet to make an informed decision for the right platform!

 

Which AI features does Azure DevOps offer?

Most of the machine learning on the Azure side lives in three services spread across Azure Monitor and Azure Pipelines: Azure Machine Learning jobs, predictive autoscale and machine learning alert thresholds. Bicep sits underneath them, declaring the infrastructure those features act on. It predicts nothing.

Bicep and AI settings in infrastructure code

Bicep is a domain-specific language that deploys Azure resources with declarative syntax, and it integrates with Azure Policy, template specs and Azure Blueprints. It doesn't forecast demand. Its job in AI-assisted DevOps is to keep AI-related settings in code that gets reviewed.

 

Microsoft's predictive autoscale documentation enables the feature through a Resource Manager template with a predictiveAutoscalePolicy setting whose scale mode is ForecastOnly or Enabled, and any property valid in a Resource Manager template is valid in a Bicep file. A pull request can then move a scale set from forecast-only to live predictive scaling, and the what-if operation previews the change before deployment.

Azure Machine Learning jobs in Azure Pipelines

There's no separate "Azure DevOps ML Services" product. Instead, teams install the Azure Machine Learning extension for Azure Pipelines from the Visual Studio Marketplace and submit jobs with the Azure Machine Learning Python SDK v2 and the Azure CLI ml extension v2. From there, Azure Pipelines can automate the machine learning lifecycle, including data preparation, model training, model deployment as web services and monitoring deployed models for performance or data drift. Models then get the same reviews, approvals and audit trail as application code.

 

Two limits affect pipeline design. The AzureML Job Wait task can wait for at most two days. That's a hard limit in Azure DevOps pipelines.

 

Microsoft's documentation also notes that an issue with the completion notification this task relies on is under investigation. Our guide to using Azure DevOps for CI/CD in web projects covers the pipeline setup itself.

Predictive autoscale in Azure Monitor

A virtual machine scale set whose CPU load repeats on a schedule can scale out before demand arrives, because predictive autoscale uses machine learning to forecast the set's total CPU load from past usage. Five limits shape how a team plans for it.

 

  • Virtual machine scale sets only, with Percentage CPU (Average) as the only supported metric.
  • It scales out only, and standard autoscale rules must exist first to handle scale-in and spikes the model didn't predict.
  • At least seven days of history, with the best results from a rolling 15-day window.
  • Azure Commercial cloud only. Azure Government clouds aren't supported.
  • Prelaunch between 5 and 60 minutes before the predicted load.

 

A forecast-only mode shows predicted CPU against actual CPU without scaling anything, which is the safe way to test the model.

Machine learning alert thresholds in Azure Monitor

Hand-picked alert values are mostly educated guesses. Azure Monitor's machine learning thresholds learn the history of metrics and log query results, find hourly, daily or weekly patterns and calculate the thresholds themselves. They start from 10 days of history and need three weeks before they detect weekly seasonality, while a new rule won't fire until it has three days and at least 30 samples of data.

 

There are gaps, too. These thresholds can't be used in rules with multiple conditions or in log search alerts that run every minute, and slow changes may not trigger an alert at all. Thresholds of this type for PromQL query-based metric alerts were still in preview in November 2026.

Example of predictive scaling for an e-commerce platform on Azure

Traffic peaks every weekday evening for an online store whose web tier runs on an Azure virtual machine scale set. For two weeks, the team runs forecast-only mode and compares predicted CPU with actual CPU. Once the forecast tracks reality, they switch on predictive autoscale with a 30-minute prelaunch, and the new instances are already warm by the time the evening rush starts.

 

Black Friday needs its own plan. Microsoft recommends schedule-based or metric-based autoscale for monthly or yearly patterns, and a once-a-year sale gets a scheduled profile.

 

Which AI features does AWS DevOps offer?

On AWS, the AI pieces still open to new projects are Amazon EC2 Auto Scaling predictive scaling and AWS CloudTrail Insights, with AWS CloudFormation as the infrastructure layer. Code review is the area that changed most. CodeGuru Reviewer and Amazon Forecast are closed to newcomers, and Amazon Q Developer reaches end of support in 2027.

Infrastructure as code with AWS CloudFormation

AWS CloudFormation provisions and configures AWS resources from a template that describes them. A template might describe an Auto Scaling group, a load balancer and a database, and because it's a text file, teams track infrastructure changes in version control and can return to a previous template when an update causes problems. Like Bicep, CloudFormation has no prediction logic. It's where a predictive scaling policy gets declared and reviewed.

AI code review on AWS after CodeGuru Reviewer

New repositories on AWS need a code review plan that avoids two retiring services. CodeGuru Reviewer used program analysis and machine learning to find defects in Java and Python code, and AWS's documentation for it states that new repository associations can't be created from November 7, 2025. AWS pointed teams to Amazon Q Developer and Amazon Inspector.

 

The first of those is now winding down as well. According to the Amazon Q Developer end-of-support announcement on the AWS DevOps Blog (2026), new account and subscription creation was blocked from May 15, 2026, and the IDE plugins and paid subscriptions reach end of support on April 30, 2027.

 

Its GitHub app, Amazon Q Developer for GitHub, accepts no new installations from November 18, 2026 and loses support on January 30, 2027. AWS points to Kiro and the Kiro GitHub App for similar capabilities. How similar is still an open question, and teams should confirm which review checks they cover before relying on them.

 

Amazon Inspector Code Security is the steadier AWS-native choice. It discovers GitHub and GitLab repositories and scans first-party source code, third-party dependencies and infrastructure as code for vulnerabilities. It focuses on vulnerabilities. Teams that want general code review comments as well, the sort a human reviewer would leave on a pull request, should check exactly what Inspector reports before they commit to it.

Predictive scaling in Amazon EC2 Auto Scaling

Daily and weekly load patterns are what Amazon EC2 Auto Scaling looks for in historical data, and once it finds them it adds instances to an Auto Scaling group before the forecast load arrives. It starts with 24 hours of metric data. Two full weeks of data gives the best results, and each forecast covers the next 48 hours in hourly steps, refreshed every six hours.

 

  • A new policy starts in forecast-only mode, which lets teams check accuracy before it changes capacity.
  • It scales out only. Other scaling policies handle scale-in and unexpected load.
  • Custom metrics are supported, including metrics aggregated across old and new Auto Scaling groups.
  • The pre-launch setting starts instances early, ahead of the forecast load.
  • Mixed instance groups can produce inaccurate forecasts, because the model assumes instances of equal capacity.
  • AWS GovCloud (US-East) and AWS GovCloud (US-West) are among the supported Regions.

 

One setting needs care. The max capacity behavior option can raise a group's maximum capacity automatically when the forecast approaches it, and that raised maximum stays until someone changes it by hand.

Unusual API activity with AWS CloudTrail Insights

Every action by a user, role or AWS service lands in CloudTrail as an event. Insights learns a baseline of API call rates and API error rates from those events and generates an Insights event when activity departs from it, which is how it detects unusual account activity. It can analyze management events and data events, though Insights on data events work only on trails and can't use event data stores. Insights events cost extra.

Example of predictive scaling for a SaaS platform on AWS

A B2B SaaS product on AWS runs its API servers in an Auto Scaling group, and load climbs every weekday morning. Instances take several minutes to boot and warm their caches. With reactive scaling alone, the first users of the day wait.

 

The team leaves a predictive scaling policy in forecast-only mode for two weeks, compares forecast capacity with what the group actually used, then switches to forecast and scale with a pre-launch buffer. A target tracking policy covers unplanned spikes and scales in overnight. Teams that built demand forecasts in Amazon Forecast need another route, because the service is no longer available to new customers. Predictions that depend on promotions or bookings call for a custom model built on that business data, which is work for predictive analytics solutions.

 

AI test selection, build failure prediction and deployment risk checks in CI/CD

Inside the CI/CD pipeline, AI-assisted delivery usually means two jobs: choosing which tests a commit needs and checking a release's risk before it reaches every user. Test selection on Azure works from a map of which tests touch which code. Deployment risk checks on AWS can use machine learning baselines.

Test selection with Test Impact Analysis in Azure Pipelines

Azure Pipelines can cut a long regression run down to the tests a commit affects. Test Impact Analysis, switched on in version 2 of the Visual Studio Test task, runs impacted tests, previously failing tests and newly added tests. When it can't reason about a change, such as an edit to an HTML or CSS file, it falls back to running everything. It builds its selection from a dependency map of managed code.

 

The supported scope is narrow. Microsoft's documentation lists .NET Core, multi-machine setups where tests exercise an app on another machine, and data-driven tests among the unsupported scenarios. Teams that qualify should also schedule a periodic full run, which the task supports, to catch any test the selection missed.

Build failure prediction in CI/CD pipelines

Predicting a failed build before it runs takes a model trained on the team's own build history, and none of the Azure or AWS services covered in this article offers one as a built-in feature. A team with enough labelled run history could train its own classifier as an Azure Machine Learning job in the pipeline. Diagnosis from pipeline logs and execution data is the more practical starting point, because it works on the failures a team already has.

Deployment risk checks with anomaly detection alarms and gates

CodeDeploy and Azure Pipelines can stop or hold a deployment when a monitored signal turns unhealthy. On AWS, that signal can come from a machine learning baseline.

 

On AWS, CodeDeploy stops a deployment when any CloudWatch alarm associated with its deployment group is activated, with up to 10 alarms per group, and it can roll back automatically to the last known good revision. Basing those alarms on CloudWatch anomaly detection swaps a fixed threshold for a band of expected values. The model trains on up to two weeks of metric data and accounts for hourly, daily and weekly patterns, and anomaly detection alarms cost extra.

 

One setting matters for release pipelines. CloudWatch lets teams exclude time periods from model training, and excluding past deployment windows stops a bad release from teaching the model that its error spike is normal.

 

On Azure, a release pipeline can add a Query Azure Monitor alerts gate before or after a stage, which watches the configured alert rules for active alerts. Gates are re-evaluated at a set interval, and the release is rejected if they don't all pass in the same interval before the timeout.

 

How AI speeds up incident detection, triage and root cause analysis

Much of the time in an incident goes on correlation, before anyone starts on a fix. A failure in one microservice can set off alerts across monitoring, logging and application performance tools. Someone on call then has to work out which alerts describe the same problem. AI-assisted incident response, often called AIOps, takes on three jobs here: grouping related alerts into one incident, triaging it and proposing a likely root cause.

 

Triage is the step a model learns from your incident history. It classifies the new incident as an outage or a performance problem, sets its severity against learned baselines and routes it to the team that resolved similar issues before. Root cause analysis goes further, linking the alert to logs, traces and recent changes such as a deployment. On Azure and AWS, that work now sits in two investigation features.

Incident stage

Azure Copilot Observability Agent in Azure Monitor

Amazon CloudWatch investigations

Getting started

Chat and deep investigation need no setup; background alert correlation (autonomous operations, preview) needs an Observability Agent resource

Investigation starts from an alarm, a metric or a Logs Insights query with no setup; accepting hypotheses, runbook suggestions and reports need an investigation group

Signals analyzed

Application, infrastructure and Azure platform signals; tuned for AKS, Application Insights and virtual machines

Metrics, logs, deployment events, AWS Health events, CloudTrail change events and X-Ray traces

Root cause output

A report with findings and recommended next steps, saved as an Azure Monitor issue

Root-cause hypotheses that engineers accept or discard, plus a generated incident report

Remediation

Doesn't restart resources, change configuration or resolve issues on its own

Can suggest a Systems Manager Automation runbook for the team to run

Limits to plan for

9 Azure regions, English only, conversations can't continue past 24 hours, no customer-managed keys

One investigation group per account, 2 concurrent active investigations, 150 enhanced investigations a month

Neither feature fixes production on its own, and that is a sensible place to start.

 

Consider an illustrative case. A build fails, and the investigation traces it to a stale cache on the CI agent instead of a code change. Cleaning the workspace and rerunning on a fresh agent is a known, low-risk fix, which makes it a good first candidate for an approved runbook. A memory leak that would need a pod's limits raised in production should stay a human decision until that runbook has a track record.

 

Both features reason over telemetry the cloud already holds. The groundwork is getting logs, metrics, traces and deployment events into Azure Monitor or CloudWatch before switching anything on. Measure mean time to resolution (MTTR) for a few weeks before and after, broken into detection, triage, analysis and action, and the team can see which delay the AI shortened. AWS teams starting now begin with CloudWatch, because Amazon DevOps Guru no longer accepts new customers and AWS points to CloudWatch for anomaly detection and to AWS DevOps Agent for AI-assisted investigation.

 

2027 outlook for Azure and AWS DevOps AI services

Most of the dates that affect 2027 planning are on AWS. As of November 2026, these are the documented changes.

 

  • November 7, 2025. CodeGuru Reviewer stopped accepting new repository associations. Existing associations still work.
  • May 15, 2026. New Amazon Q Developer accounts and subscriptions were blocked.
  • November 18, 2026. New installations of the Amazon Q Developer for GitHub app stop.
  • January 30, 2027. Support for the Amazon Q Developer for GitHub app ends.
  • April 30, 2027. Amazon Q Developer IDE plugins and paid subscriptions reach end of support.

 

Amazon Forecast is closed to new customers, but existing customers can keep using it as normal, which at least buys some time for the teams that already depend on it. Amazon Q Developer in the AWS Management Console is outside the announced sunset. On Azure, the documented constraint is regional. Predictive autoscale still doesn't run in Azure Government clouds.

 

Amazon DevOps Guru also stopped accepting new customers on October 29, 2026, and support ends on September 30, 2027, after which its console, API and past insights are no longer available.

 

The date reaches into infrastructure code too. Templates in CloudFormation, the AWS CDK or Terraform that declare AWS::DevOpsGuru::ResourceCollection or AWS::DevOpsGuru::NotificationChannel will fail to create or update once AWS withdraws those resource types. One failed resource can block a whole stack operation, which makes removing them the most urgent part of a DevOps Guru migration.

 

How do Azure and AWS DevOps AI services compare?

Azure and AWS cover the same four stages of AI-assisted DevOps with different services and limits. The table reflects vendor documentation as of November 2026.

Capability

Azure

AWS

Best for

Infrastructure as code

Bicep, declarative, no forecasting

AWS CloudFormation templates, no forecasting

Keeping scaling and alert settings in reviewed code

AI in the pipeline

Azure Machine Learning jobs through the Azure Pipelines extension

Amazon Inspector Code Security scanning; Amazon Q Developer review until its 2027 end-of-support dates

Azure for teams shipping ML models, AWS for repository vulnerability scanning

Predictive scaling

Azure Monitor predictive autoscale (scale sets, Percentage CPU, scale-out only)

EC2 Auto Scaling predictive scaling (Auto Scaling groups, custom metrics, scale-out only)

Cyclical traffic and slow-starting applications

History needed

7 days for predictive autoscale; 10 days for ML alert thresholds

24 hours for predictive scaling, best with 14 days

Estimating how soon a pilot can start

Anomaly detection

ML alert thresholds on metrics and log queries

CloudTrail Insights on API call and error rates, extra charge

Azure for performance alerts, AWS for unusual account activity

Status to check

Predictive autoscale not available in Azure Government clouds

CodeGuru Reviewer and Amazon Forecast closed to newcomers; Amazon Q Developer support ends in 2027

Planning a toolchain for new projects

Which cloud's AI DevOps tooling fits your stack?

Start on the cloud that already runs your workloads and holds your data and source code. Every scaling and alerting feature above learns from data that platform already collects, and moving clouds just to get one AI feature rarely pays for the disruption it causes. Four differences then decide where AI helps first.

 

  • Azure fits best when models are trained in Azure Machine Learning and code ships through Azure DevOps, since model releases can share the application pipeline.
  • AWS predictive scaling is broader. It accepts custom metrics and runs in AWS GovCloud, while Azure's version covers only CPU on scale sets in the commercial cloud. Both scale out only.
  • AWS teams using CodeGuru Reviewer or the Amazon Q Developer GitHub app need a replacement chosen before January 30, 2027.
  • Azure teams with noisy alerts can switch existing metric alerts to machine learning thresholds and stop picking values by hand.

 

Our view is that predictive scaling is the lowest-risk first step on either cloud, as long as the forecast is checked against real traffic before it controls capacity. Both clouds now offer a forecast-only mode built for that check.

 

Frequently asked questions

Is Amazon CodeGuru Reviewer still available?

Only for existing repository associations. AWS stopped new associations on November 7, 2025, so new repositories need another option, such as Amazon Inspector Code Security for vulnerability scanning.

Should a new project adopt Amazon Q Developer for code review?

It's a poor choice for a new project. New Q Developer subscriptions were blocked from May 15, 2026, the GitHub app loses support on January 30, 2027, and the IDE plugins follow on April 30, 2027. AWS names Kiro as the replacement, so teams should evaluate Kiro or another reviewer alongside Amazon Inspector scanning.

Can Bicep or CloudFormation scale infrastructure using AI?

No. Neither one forecasts demand. The machine learning that decides when to scale lives in Azure Monitor predictive autoscale and EC2 Auto Scaling predictive scaling, while Bicep and AWS CloudFormation declare and provision infrastructure from code and give teams a place to define and review those scaling settings.

How much history do predictive scaling features need?

About a week on Azure and a day on AWS, at minimum. Azure predictive autoscale needs at least seven days of CPU history and works best with a rolling 15-day window, while EC2 Auto Scaling predictive scaling starts after 24 hours of metric data and is more accurate with two full weeks. For monthly or yearly patterns such as an annual sale, Microsoft recommends schedule-based or metric-based autoscale on Azure.

Can AI resolve incidents automatically on Azure or AWS?

Not by default. Azure's Observability Agent correlates alerts and investigates them, but it doesn't restart resources, change configuration or resolve issues on its own. CloudWatch investigations can suggest a Systems Manager Automation runbook, which the team decides whether to run. Automatic fixes are best limited to known, low-risk runbooks that have already worked under human approval.

 

Getting AI into your existing Azure or AWS DevOps toolchain

Pick one use case where your pipeline already has past data and a repeatable decision. Predictive scaling for cyclical traffic is the usual first choice, followed by machine learning alert thresholds for noisy alerts. Let the forecast-only mode run for two weeks while you compare predicted and actual load, and only after that let the model start changing capacity on its own. Standard scaling rules and human review stay throughout.

 

AWS teams also have a less pleasant job, which is settling their code review replacement before the January and April 2027 Amazon Q Developer dates. Arbisoft's DevOps team works on infrastructure as code with automated provisioning, plus performance monitoring with predictive alerting. If you'd like help scoping that first use case, Arbisoft's DevOps services team can work through it with your own traffic data.

...Loading Related Blogs

Explore More

From Introduction to Proposal in Days

Discovery Call
Our sales team reviews your message and asks for a discovery call to gather more information.
Expert Input
Our veterans go through your requirements to provide their take, backed by decades of experience.
Proposal
We provide a proposal specific to what you're building, for you to review at your own pace.

Trusted by top platforms for our transformative solutions and exceptional results:

How Can We Help You Build?

We recommend using your work email.
We'll send a mutual NDA before the discovery call if requested. Zero obligation.