AI and Machine LearningAzure Machine Learningverified
Prompt flow
Prompt flow is a way to build and test an AI application as a flow instead of a single prompt. A flow can connect prompts, Python logic, tools, inputs, outputs, variants, and evaluation steps. In Microsoft Foundry classic and Azure Machine Learning, it helped teams prototype, debug, compare, deploy, and monitor LLM-based applications. In 2026, it is also a migration topic because Microsoft announced that feature development ended and full retirement is planned for April 20, 2027.
Prompt flow is a classic Microsoft Foundry and Azure Machine Learning tool for building LLM application flows with prompts, tools, variants, evaluation, deployment, and monitoring. Microsoft Learn notes feature development ended on April 20, 2026, with retirement planned for April 20, 2027.
In Azure architecture, prompt flow historically sits in the AI development platform layer, connected to workspaces, projects, model deployments, compute, storage, identity, secrets, and monitoring. A flow represents executable AI logic made from nodes, prompts, tools, and data inputs. It can be tested, evaluated, and deployed as an endpoint. For current planning, teams must distinguish legacy Prompt Flow workloads from newer Microsoft Foundry and Agent Framework patterns, because support timelines and project compatibility affect architecture decisions.
Why it matters
Prompt flow matters because many teams used it to move from prompt experiments to executable, testable AI workflows. It provided a visual graph, variants, debugging, evaluation, collaboration, and deployment paths that made prompt engineering less ad hoc. In 2026, its importance also comes from lifecycle risk: Microsoft Learn says feature development ended on April 20, 2026, and retirement is planned for April 20, 2027. Operators need to know whether an existing AI workload depends on Prompt Flow so they can maintain it safely, avoid surprise read-only behavior, and plan migration before production support becomes constrained. That changes a glossary term into a production planning concern.
⌁
Where you see it
Signals, screens, and Azure surfaces where this term usually becomes operational.
Signal 01
Foundry classic or Azure Machine Learning studio shows prompt flow assets as visual graphs with nodes, inputs, outputs, variants, runs, evaluations, deployments, and retirement notices.
Signal 02
Azure Machine Learning CLI output shows workspaces, computes, endpoints, jobs, identities, storage, and connections that often surround Prompt Flow assets and deployed flow endpoint inventory.
Signal 03
Run history, endpoint metrics, and Application Insights traces show node failures, model latency, token use, tool-call errors, dependency breaks, and behavior that must be migrated safely.
✦
When this becomes relevant
Specific situations where this term helps solve real Azure design, operations, migration, security, reliability, cost, or governance problems.
Inventory existing Prompt Flow workloads before the April 2027 retirement deadline affects production AI features.
Map flow nodes, tools, connections, and endpoints so hidden business logic is not lost during migration.
Compare prompt variants and evaluation results for an existing classic flow before choosing a replacement pattern.
Troubleshoot a deployed flow endpoint where node failures, expired connections, or model changes break users.
Plan migration from hub-based Prompt Flow assets to Microsoft Agent Framework or direct application orchestration.
◆
Real-world case studies
Different enterprise-style examples that show the term being used to hit measurable objectives.
Case study 01
Engineering consultancy inventories legacy flows
Scenario, objectives, solution, measured impact, and takeaway.
📌Scenario
An engineering consultancy had built dozens of client prototypes with Prompt Flow in Azure Machine Learning. After the retirement notice, no one knew which flows were still active.
🎯Business/Technical Objectives
Identify production and dormant flows before migration planning.
Map each flow to endpoints, model deployments, storage, and owners.
Avoid deleting client-facing endpoints by mistake.
Create a migration priority list before contract renewals.
✅Solution Using Prompt flow
The platform team treated prompt flow as an application inventory problem. They used Azure CLI to list workspaces, computes, endpoints, and deployments, then matched those resources to Prompt Flow assets found in studio. Active endpoint traffic, recent run history, and Application Insights traces identified which flows still mattered. For each active flow, architects documented prompt nodes, Python tools, input and output contracts, data connections, identities, and evaluation results. Low-traffic prototypes were tagged for retirement, while client-facing flows were prioritized for migration to Microsoft Agent Framework or application-owned orchestration. No destructive cleanup happened until owners approved the dependency map.
📈Results & Business Impact
Forty-three flows were inventoried across nine client subscriptions.
Seventeen inactive prototype endpoints were retired, saving about $6,400 per month.
Eight production flows received migration plans with named owners and timelines.
No active client workflow was deleted during cleanup.
💡Key Takeaway for Glossary Readers
Prompt flow migration starts with dependency discovery, not with rewriting prompts.
Case study 02
Pharmaceutical lab preserves experiment workflow
Scenario, objectives, solution, measured impact, and takeaway.
📌Scenario
A pharmaceutical research lab used Prompt Flow to orchestrate literature summaries, compound notes, and reviewer comments. Scientists needed continuity while engineers planned a replacement.
🎯Business/Technical Objectives
Keep the existing flow stable through ongoing research cycles.
Document tool and data dependencies before migration.
Preserve reviewer-approved output format.
Reduce failure rate from expired connections and compute issues.
✅Solution Using Prompt flow
The AI operations team reviewed the prompt flow graph and found several Python nodes that transformed research metadata before model calls. They documented each node, connection, prompt variant, and output field. Managed identity access to storage and search was verified, and stale secrets were moved back under approved Key Vault handling. Azure CLI checks recorded workspace, compute, endpoint, and diagnostic settings for the runbook. The team froze nonessential prompt changes, added monitoring for flow-run failures, and built a migration backlog that separated prompt logic from laboratory data transformations. Scientists received a freeze window for changes so baseline behavior stayed stable during comparison.
📈Results & Business Impact
Flow-run failure rate fell from 11% to 3% after connection cleanup.
Scientists kept the approved summary format through two quarterly research reviews.
Migration design time dropped by 30% because dependencies were already documented.
No research data was copied into unmanaged migration scripts.
💡Key Takeaway for Glossary Readers
Prompt flow can hide valuable orchestration logic that must be protected before replacement.
Case study 03
Sports media team retires unused fan bots
Scenario, objectives, solution, measured impact, and takeaway.
📌Scenario
A sports media company had several Prompt Flow fan-engagement bots created for seasonal events. The operations team wanted to retire stale flows without disrupting active tournament coverage.
🎯Business/Technical Objectives
Find which flow endpoints still received fan traffic.
Remove idle resources before a new tournament budget cycle.
Preserve one active workflow for migration testing.
Improve observability before the remaining bot moved platforms.
✅Solution Using Prompt flow
Engineers used Azure CLI to list Azure Machine Learning workspaces, endpoints, deployments, and diagnostic settings. They correlated that inventory with traffic logs and Prompt Flow run history. Three event-specific bots had no traffic for six months and were marked for retirement after content owners confirmed no upcoming reuse. One active tournament bot was kept and documented for migration. The team exported its prompt flow graph, prompt variants, tool calls, output schema, and monitoring requirements. Before replacement work began, they added alerts for endpoint errors and latency spikes so the live bot could be watched during the transition.
📈Results & Business Impact
Three dormant bots were retired, reducing monthly platform spend by 18%.
The active tournament bot received a complete dependency and migration document.
Endpoint error alerts were added before the next live event.
Cleanup finished without any fan-facing outage.
💡Key Takeaway for Glossary Readers
Prompt flow retirement is safer when traffic evidence separates abandoned experiments from active applications.
Why use Azure CLI for this?
As an Azure engineer with ten years of delivery scars, I use Azure CLI for prompt flow because the portal shows the flow, but incidents usually involve surrounding resources. I need to inspect workspaces, computes, endpoints, deployments, identities, storage accounts, and diagnostic settings quickly. CLI is also useful for migration inventory because many teams have forgotten which flows still exist or which endpoints receive traffic. Since Prompt Flow has a retirement timeline, scripted discovery is safer than clicking through projects manually. It creates evidence for owners, dependencies, and cleanup decisions. I also use it to build a migration backlog backed by evidence rather than memory.
CLI use cases
List Azure Machine Learning workspaces and computes that may host legacy Prompt Flow assets.
Inspect endpoints and deployments connected to flows before planning migration or cleanup.
Check managed identity, storage, and Key Vault dependencies that a flow may require at runtime.
Export resource configuration for migration assessments and retirement-readiness reviews.
Compare monitoring settings across environments before troubleshooting deployed flow latency or failures.
Before you run CLI
Confirm tenant, subscription, resource group, workspace or Foundry project, region, and permissions before inventory work.
Install or update required Azure CLI extensions, especially Azure Machine Learning commands used for workspace inspection.
Treat delete, update, or endpoint commands as risky because flows may support active production users.
Check cost impact before starting computes, running batch jobs, or triggering large evaluations during discovery.
Use JSON output and tag exports with migration owner, flow name, endpoint name, and review date.
What output tells you
Workspace and endpoint IDs show where legacy Prompt Flow assets live and which teams may own them.
Compute output reveals whether a flow depends on active clusters, serverless runtime, or idle resources wasting spend.
Identity and connection details indicate whether migration must preserve managed identity, secrets, or private access.
Deployment status and traffic fields show whether a flow endpoint is still serving users or safe to retire.
Diagnostic settings show whether operators can trace node failures, latency, and model calls during migration testing.
Mapped Azure CLI commands
Azure Machine Learning operations
direct
az ml workspace list --resource-group <resource-group>
az ml workspacediscoverAI and Machine Learning
az ml workspace show --name <workspace-name> --resource-group <resource-group>
az ml workspacediscoverAI and Machine Learning
az ml workspace create --name <workspace-name> --resource-group <resource-group> --location <region>
az ml workspaceprovisionAI and Machine Learning
az ml compute list --workspace-name <workspace-name> --resource-group <resource-group>
az ml computediscoverAI and Machine Learning
az ml workspace delete --name <workspace-name> --resource-group <resource-group>
az ml workspaceremoveAI and Machine Learning
Cognitive operations
direct
az cognitiveservices account list --resource-group <resource-group>
az cognitiveservices accountdiscoverAI and Machine Learning
az cognitiveservices account show --name <account> --resource-group <resource-group>
az cognitiveservices accountdiscoverAI and Machine Learning
az cognitiveservices account deploymentprovisionAI and Machine Learning
Ml operations
direct
az ml workspace show --name <workspace> --resource-group <resource-group>
az ml workspacediscoverAI and Machine Learning
az ml workspace create --name <workspace> --resource-group <resource-group> --location <region>
az ml workspaceprovisionAI and Machine Learning
az ml compute list --workspace-name <workspace> --resource-group <resource-group>
az ml computediscoverAI and Machine Learning
az ml model list --workspace-name <workspace> --resource-group <resource-group>
az ml modeldiscoverAI and Machine Learning
az ml online-endpoint list --workspace-name <workspace> --resource-group <resource-group>
az ml online-endpointdiscoverAI and Machine Learning
Architecture context
As an Azure architect, I treat prompt flow as a legacy orchestration asset that deserves inventory, not casual deletion. Existing flows may encode prompt templates, Python tools, data connections, evaluations, endpoint behavior, and team approval history. Before migration, I map each flow to its model deployments, workspace or project type, storage, Key Vault secrets, managed identity, private networking, compute, monitoring, and downstream consumers. I also identify whether the flow should move to Microsoft Agent Framework, a direct application service, or another Foundry pattern. The architectural risk is not only retirement; it is losing hidden business logic during migration. This prevents migration from becoming a blind rewrite.
Security
Security impact is direct for existing Prompt Flow workloads because flows can connect to model deployments, tools, data sources, credentials, and endpoints. A flow should use managed identity or approved secret storage rather than hard-coded keys. Operators must review who can edit, run, deploy, or view traces because prompts and outputs may contain sensitive data. Network exposure matters if a deployed flow endpoint is public or calls private resources. Retirement planning also has security implications: rushed migrations often loosen permissions, skip redaction, or copy secrets into new code. Treat every flow as executable application logic with data-access boundaries. Access reviews should happen before any export or rebuild.
Cost
Cost impact comes from model calls, evaluation runs, compute, storage, deployed endpoints, logging, and migration effort. Prompt flow itself is often part of a broader workspace or project cost picture. Idle endpoints, oversized compute, repeated batch runs, and verbose prompts can waste money. Migration planning also has cost because teams must rebuild, test, and validate flows before retirement. FinOps reviews should identify active flows, endpoint traffic, unused deployments, storage growth, evaluation frequency, and model-token spend. A neglected legacy flow can keep consuming resources after the business process has moved elsewhere. Tagging and ownership help separate essential flows from abandoned experiments.
Reliability
Reliability impact is direct for workloads still depending on Prompt Flow endpoints or batch runs. A flow can fail because a node changes, a tool dependency breaks, compute is unavailable, a connection expires, or the model deployment changes. The announced retirement timeline increases operational risk because teams must plan migration before the service enters read-only mode. Reliable handling includes inventorying flows, documenting owners, capturing dependencies, exporting configuration, testing equivalent behavior in the target platform, and maintaining rollback during migration. Operators should monitor endpoint health, flow-run failures, latency, token use, and evaluation scores until the workload is fully retired or replaced.
Performance
Performance impact depends on how the flow is designed. A flow with multiple LLM calls, Python nodes, external tools, retrieval steps, and large prompts can add latency compared with a simpler application path. Batch runs and evaluations can also take longer when datasets are large or compute is constrained. Operators should measure node-level timing, model latency, tool-call duration, token volume, endpoint cold starts, and queueing. Migration is a chance to remove unnecessary nodes, tighten prompts, and choose a better runtime. Performance tuning should not ignore quality, because a faster flow that skips grounding or validation may fail the business task.
Operations
Operators manage prompt flow by listing workspaces or projects, reviewing flow assets, checking endpoints, tracing failed runs, validating connections, and coordinating migration. Day-to-day work includes investigating node failures, comparing variants, confirming deployment status, monitoring endpoint metrics, and checking access to storage, Key Vault, compute, and model deployments. Azure CLI is most useful for the surrounding Azure Machine Learning, Cognitive Services, and monitoring resources. In 2026, runbooks should also include retirement milestones, migration owners, compatibility checks, evidence exports, and a plan for flows that will remain in place until replacement. That makes retirement planning visible instead of tribal knowledge stored in one team.
Common mistakes
Ignoring the Prompt Flow retirement timeline until a production migration becomes rushed and risky.
Deleting a flow because it looks unused without checking endpoint traffic and downstream consumers.
Copying flow secrets into new code instead of preserving managed identity or approved secret storage.
Migrating the visible prompt while forgetting Python nodes, tool connections, evaluations, and output contracts.
Assuming classic Foundry portal assets work the same way in newer Foundry project patterns.