What access does someone need to view an Azure ML job?

An Entra ID identity in the tenant plus an RBAC role on the workspace — typically Reader or AzureML Data Scientist. For anyone outside the tenant, that means a guest account first.

How do I share results with someone outside the tenant?

Pull the metrics and artifacts from the job with the SDK or CLI, render them into a page, and publish it. No guest account, no role assignment.

Does this work for AutoML runs?

Yes, and it is a good fit — an AutoML run produces a leaderboard most stakeholders should see summarized rather than explored.

Can a pipeline publish its own summary?

Yes. A final step that renders and PATCHes an existing report gives every scheduled run one stable, current URL.

What about responsible-AI dashboards?

Export the figures you need. The interactive dashboard is a Studio surface; the fairness and error-analysis charts travel fine as static images with their numbers stated in text.

Share Azure ML job results

The training job finished. Azure ML Studio has the metrics, the run history, the artifacts, the responsible-AI dashboard.

To see it, someone needs to be in your Entra ID tenant and hold an RBAC role on the workspace. For a colleague that is a ticket. For a partner or a client it is a guest account, which is a conversation with IT.

The access ladder is steep

Azure's identity model is thorough, and for a workspace holding training data that is correct. It is also disproportionate to the actual request, which is usually: what was the AUC, and is it better than what we have?

Between that question and the answer sit a guest invitation, a role assignment, an MFA enrolment and a Studio UI the person has never used. The predictable outcome is a screenshot in Teams and a number nobody can verify.

Publish the job's results

from azure.ai.ml import MLClient
from azure.identity import DefaultAzureCredential

ml = MLClient(DefaultAzureCredential(), sub_id, rg, workspace)
job = ml.jobs.get(job_name)
# read logged metrics and download the plots you want to include

Assemble a page that leads with the decision:

What we recommend, in one sentence.
Candidate vs incumbent on the metric that decides it.
The regression — whichever slice or metric got worse.
Appendix — job name, compute target, dataset version, environment.

An AutoML run benefits especially from this treatment: the leaderboard's top row is the answer, and the other forty rows are for you, not for the reader.

Responsible-AI figures travel as images

Fairness dashboards and error analysis are Studio-interactive. Export the charts you need as images and state the numbers in text alongside them — a stakeholder needs "false-negative rate is 8 points higher for the under-25 segment", not a widget they cannot filter.

Pairs with sharing a Fairlearn or SHAP plot for the same reason: the explanation has to survive leaving the tool.

The conversation has to survive too

A published page keeps the discussion anchored to the number being discussed — see commenting on HTML. In Teams the same exchange scrolls away, and six months later nobody can reconstruct why the model shipped with that regression.

Worth knowing

Restrict the report. Metrics often reveal training-data composition — see security.
Publish from the pipeline's last step, updating one report id, so the URL in the ticket is never stale.
Include the job name. Anyone with Studio access should be able to get from the page to the run in one step.
5 MB per report body.

Try it

Free — unlimited reports, commenters and revisions.

Read the API reference →