Facts checked against AWS SageMaker documentation on 2 October 2026. AWS changes console labels from time to time; if your console looks different, check its current wording.
When Amazon Mechanical Turk closed on 30 September 2026, SageMaker Ground Truth labeling jobs and Amazon A2I review loops that used the MTurk workforce lost their human step. Ground Truth and A2I keep running for existing customers. What's gone is one workforce option, so the work is to replace that option without rebuilding everything around it.
What changed, and what didn't
Ground Truth and A2I support two other workforce types:
- Private workforce. People you choose, signing in to a labeling portal for your AWS account. You manage them with Amazon Cognito or your own OIDC identity provider.
- Vendor workforce. A labeling vendor you subscribe to through AWS Marketplace.
Your task templates, input and output buckets and the rest of your pipeline stay the same. The workforce setting is what changes.
Step 1: find everything that used MTurk
Search for the public workforce ARN, which ends in workteam/public-crowd/default. Check labeling job definitions and the scripts that create them, A2I human review workflows (flow definitions), and infrastructure code such as CloudFormation, CDK or Terraform. Also search for the public workforce task price setting, which only applied to MTurk.
For each match, write down the owner, weekly volume, task type, and whether the data includes personal information. You need all four to choose a workforce.
Step 2: choose a new workforce
There's no single right answer. It depends on volume, data sensitivity, and who you want accountable for quality.
- Your own staff as a private workforce. Best when the task needs deep product knowledge, or the data must stay with a small group. Full control, but you take on hiring, training, scheduling and review.
- A managed team as a private workforce. A vendor's people sign in to your portal as members of your private workforce. You keep control of access; the vendor handles training, a team lead and QA. Ask how they measure quality, whether the same people stay on your work, and what their security audits cover.
- A vendor workforce from AWS Marketplace. Useful if you want to buy through your AWS account. Check which task types each vendor supports and how pricing works.
- Automate more, review less. If a model is now good enough for most items, send only low-confidence items to people. You still need a workforce for the rest, so this combines with the options above.
- Move the task out of AWS labeling tools. If Ground Truth was only there to reach MTurk, a team can work directly in your own tools: a spreadsheet, an admin panel or a labeling tool you already run. You give up the AWS job orchestration in exchange.
The rest of this guide covers the private workforce route.
Before you start
Gather these first. Most stalled set-ups are missing one of them.
- The AWS Region your jobs and workflows run in
- An IAM role with access to your output bucket
- An S3 output bucket in the same Region
- Your existing worker task template
- Reviewer work email addresses
- Your sign-in choice: Amazon Cognito or your own OIDC provider
- IP ranges, if your security policy limits where reviewers can sign in
- A decision on personal data: which fields reviewers may see, and what is masked
Step 3: set up the private workforce
AWS allows one private workforce per account per Region, shared between Ground Truth and A2I. Inside it you create work teams; each labeling job or review workflow sends tasks to one team.
- Choose how people sign in. Creating the workforce in the console uses Amazon Cognito, which is usually simpler for an outside group: invitations go by email, and you can disable people from the console. To use your own OIDC provider, AWS says you must create the workforce with the
CreateWorkforceAPI. We couldn't confirm that an existing workforce can switch between the two, so treat it as a one-time choice. - Create a work team. In the SageMaker console, open Labeling workforces and choose Private, then create a private team. If it's your first, this also creates the workforce. Make a separate, named team for each group of reviewers, so you can route work and remove access per team.
- Invite people. Invite by work email; the documented limit is 50 addresses per invitation list. Email addresses are case sensitive, so ask people to sign in with exactly the address you entered.
- Restrict access if you need to. AWS documents a way to limit access to tasks by IP address. Test sign-in from inside and outside the allowed range. Disabling a person stops them receiving work without removing them, which is useful between projects.
Step 4: point your jobs and workflows at the new team
Ground Truth: create new labeling jobs with the private work team ARN, and leave out the public workforce task price.
A2I: the SageMaker API lists create, describe, list and delete operations for flow definitions, but no update operation. To move a workflow from MTurk to a private team, create a new one:
- In the Augmented AI section of the console, open Human review workflows and create a workflow.
- Name it: lowercase letters, numbers and hyphens, up to 63 characters, unique in the Region.
- Set the S3 output location in the same Region, and choose the IAM role.
- Choose the task type. For the Amazon Textract and Amazon Rekognition built-in types, set the conditions that start a review, such as a confidence threshold, and an optional random sample. Custom task types have no activation conditions: your code decides when to call
StartHumanLoop. - Choose the same worker task template.
- For workers, choose Private and select your new work team.
Then update your application to use the new flow definition ARN. Check any job that was still running on the MTurk workforce at the closure, and decide whether to rerun its unfinished items.
Step 5: plan how reviewers find new work
AWS documents Amazon SNS notifications for Ground Truth, but not for A2I, so reviewers won't get an email when A2I tasks arrive. Agree a schedule for checking the portal, and watch loop status on your side: A2I sends events to Amazon EventBridge when a human loop is Completed, Failed or Stopped.
Step 6: test before you move everything
- Build a gold set: 50 to 200 items with known-correct answers, checked by someone who knows the task.
- Send them through the new job or workflow, mixed with ordinary items.
- Read the results (from the S3 output, or with
DescribeHumanLoopfor A2I) and compare them with the known answers. - Where they differ, decide whether the instructions or the answer key is wrong, and fix it.
- Move the rest of the volume.
Keep the gold set after the move. It's the simplest way to see whether quality drifts later, whoever does the work.
Worked example
A team reviews invoice fields that Textract reads with low confidence. Their workflow, invoice-review-v3, pointed at the MTurk ARN ending in public-crowd/default.
- They create a work team,
ap-reviewers, in us-east-1, and invite six reviewers. - They create
invoice-review-v4with the same task template, output bucket and confidence condition, and the private team. - They change one configuration value from the v3 ARN to the v4 ARN, and update their reporting job to read both S3 prefixes.
- They send 120 invoices with known answers. Reviewers and the key disagree on 9: five come from a date-format rule the instructions never stated, so they add it; four were errors in the key.
Pitfalls
- Missing the output path change. Downstream jobs stop receiving results without any error.
- Old content classifier logic. Code written for MTurk may set
FreeOfPersonallyIdentifiableInformation. Review it, and decide on personal data explicitly for the new team. - One team for everything. Separate teams per task keep access reviews and offboarding simple.
- No answer key. Without known answers, you can't tell a careful team from a fast one.
- Assuming email alerts. A2I doesn't send SNS notifications to reviewers.
Checklist
- Approve or reject completed HITs (by 30 October 2026) and download history (by 28 January 2027)
- List every job, workflow and infrastructure file that references
public-crowd/default - Choose a workforce for each task
- Choose Cognito or OIDC, then create the private workforce and named work teams
- Invite reviewers and confirm each can sign in; set IP limits if needed
- Create new Ground Truth jobs and new A2I workflows with the private team
- Update the flow definition ARN in code, and anything that reads the old S3 prefix
- Run a known-answer batch, fix the instructions, then move full volume