What is LQA Agent?

LQA Agent is an AI-powered tool that automatically evaluates translations for quality errors using the same MQM-based framework as a typical human LQA process. It reviews translations and records errors by category and severity, generating quality scores (using the MQM schema) without requiring a human linguist to evaluate each string.

As LLM translation volumes grow, maintaining quality oversight becomes harder to scale with human reviewers alone. LQA Agent makes it practical to monitor quality at the volume teams are actually working at, surfacing errors, flagging what needs human attention, and tracking quality trends over time at 90% of the accuracy of a human reviewer.

LQA Agent is designed to support human LQA, not replace it. It tends to score more strictly than a human evaluator and is best used as a scalable first pass that helps you prioritize and direct human evaluation efforts. LQA Agent identifies and records errors but does not modify or correct any translations.

Who should use LQA Agent?

LQA Agent is a good fit if you:

• Are scaling LLM translation volume and need quality monitoring to keep up
• Want to start quality scoring without investing in a full human LQA program
• Have an existing LQA program and want AI support to expand coverage

How it works

LQA Agent uses AI to evaluate translation output and assign standardized quality scores. When content enters a workflow step where LQA Agent is enabled, it reviews each translation against the MQM schema assigned to that step and records any errors it detects.


 

Language support: LQA Agent currently requires English source content. Based on testing, the following are recommended target locales where LQA Agent's scores most closely match human reviewers: Arabic (International), Chinese (Simplified), Chinese (Traditional), Czech, French (Canada), German, Greek, Indonesian, Italian, Korean, Polish, Portuguese (Brazil), Romanian, Spanish (Mexico), and Spanish (Spain).

Error detection

LQA Agent can detect the following error types:

• Mistranslation
• Inconsistency
• Unidiomatic
• Whitespace
• Grammar
• Omission
• Punctuation
• Spelling
• Register
• Markup/technical
• Addition
• Language variety
• Local formatting
• Duplication
• Culture-specific reference

LQA Agent does not detect glossary errors, style guide errors, preferential errors, repeated errors, or errors that fall under the Other category (e.g., cultural reference, kudos).

Schema compatibility

For best results, use the Smartling LQA Agent MQM Schema template. This template aligns with LQA Agent's detection capabilities and is customized to match the error types that LQA Agent can identify and record. If your schema includes error types that LQA Agent cannot detect, those findings will be logged as translation issues instead of LQA errors (see Note on error types and categories below).

You can adjust the severity weights in your schema to suit your quality standards. For the Acceptable Penalty Points, the recommended default is 100 (MQM passing threshold of 90). See Setting up LQA Agent for full configuration steps.

Note on error types and categories

LQA Agent does not support additional custom error categories or error types. If standard error types have been removed from your schema, the Agent will create a translation issue instead of an LQA error for those findings. Translation issues can be viewed in the CAT Tool and, if needed, converted to LQA errors by a human reviewer. The issue comment will be pre-populated in the note field of the LQA error record, but the reviewer must select the appropriate category, error type, and severity.

For example, if you remove a standard error type like "Whitespace" from the schema, LQA Agent may still detect whitespace errors and log them as translation issues rather than LQA errors.

Recommended setup approaches

LQA Agent is most effective when used with LQA Suite, which provides a dedicated LQA project with automated sampling capabilities. There are two recommended implementations:

1. LQA Agent as an AI-powered quality monitor

Use LQA Agent as your sole evaluation method. This is the fastest and most economical approach. It produces a directional MQM score that can be used to interpret general translation quality and identify trends over time.

Best for: Teams that need a baseline quality signal across high volumes and do not yet have a human LQA program.

Recommended configuration steps:
When creating the schema, set the Acceptable Penalty Points to 100 (this means the MQM passing threshold is 90). This is the recommended default for LQA Agent.

Workflow example:

Optional enhancements:
Create an automatic sampling rule to create and submit translations on a regular cadence.

2. LQA Agent + targeted human validation

LQA Agent reviews the entire sample first. Locales that pass and translations with no errors can be approved and moved out of the human validation queue. Only translations in failed locales where LQA errors are present get routed to a human evaluator for review and editing.

Best for: Teams that want broad quality coverage with human review focused on problem areas.

Recommended configuration steps: Same as above, with an additional workflow step added after the LQA Agent step. Human evaluators are assigned to this second step (not the LQA Agent step). When evaluators open the CAT Tool on this step, they can filter strings by LQA Status and select LQA Errors Present to review only the strings where LQA Agent recorded errors. From there, evaluators can modify the errors as needed and make edits to translations. 




Workflow example: 

Optional enhancements:
Create an automatic sampling rule to create and submit translations on a regular cadence.

LQA Memory

LQA Memory is a background feature that improves the accuracy of LQA Agent results over time. It stores validated LQA results and references them when evaluating new content, reducing false-positive errors and correction work for human evaluators. LQA Memory is enabled automatically for all accounts using LQA Agent and requires no configuration.

How it works

When LQA Agent evaluates a translation, it first checks LQA Memory for a record of the same string in the same project and target locale:

• If an exact match is found, LQA Agent applies the stored result directly.
• If no exact match is found, but a partial match exists, it is included as context for LQA Agent.
• If no relevant records are found, LQA Agent evaluates the string using its standard process.

LQA Memory starts empty and builds automatically as human evaluators complete reviews. Early evaluations will not benefit from memory; accuracy improvements become more noticeable as your validated error history grows.

Setting up LQA Agent

Prerequisites

• Enterprise tier account 
• Anthropic Claude enabled in LLM Provider Settings (if Claude is disabled in your account, LQA Agent will not be available) 
• LQA Suite (highly recommended) 
• A published LQA schema — the Smartling LQA Agent MQM Schema template is recommended 
• English source content

Steps

1. Create and publish a schema. Go to Account Settings > Linguistic Quality Assurance and create a new schema using the Smartling LQA Agent MQM Schema template. Set the Acceptable Penalty Points (APP) to 100 (MQM passing threshold of 90). This is the recommended value.
   
   Pass/fail messaging in the Smartling dashboard is determined by this threshold set in your schema, so adjusting the value will directly affect how scores are reported. For example, with the recommended APP of 100 (passing threshold of 90), a score of 95 would be a pass, while a score of 85 would be a fail. If you already have a compatible schema, you can reuse it. Make sure the schema is published — LQA Agent can only use published schemas. For details, see LQA: Setup Guide.
   
   
2. Create an LQA project (recommended). If you have LQA Suite, create an LQA project using the schema from step 1. As part of project creation, a new workflow is automatically created, or you can use an existing LQA workflow. This is where LQA Agent will run. If you do not have LQA Suite, you can enable LQA Agent on any post-translation workflow step in any type of project.
    
3. Enable LQA Agent. Navigate to the workflow where you would like to enable LQA Agent. For help finding and editing workflows, see Access and Manage Workflows. We recommend adding a dedicated Quality Evaluation step specifically for LQA Agent. Name it something clear like "LQA Agent" so it is easy to distinguish from other steps in the workflow. Once you've identified the step, click on it and select Manage Step.
   
   
   
   Under Quality Assurance, select Use this step for Linguistic Quality Assurance and choose your MQM-compatible schema from the dropdown. Then enable the Enable LQA Agent toggle. 
   
   
   When content reaches this workflow step, LQA Agent will automatically review the translations, record any errors, and submit the strings to the next step in the workflow. 
4. Create an additional workflow step for human validation (optional). The LQA Agent-enabled workflow step should not be assigned to human evaluators. If you want to include human validation in your process, assign evaluators to a separate workflow step after the LQA Agent step.
   
   Make sure this step is also configured to Use this step for Linguistic Quality Assurance. Without this enabled, linguists cannot access the LQA dialog on that step, which means they won't be able to view error history. 
    
5. Set up sampling (recommended). It is not recommended to run LQA Agent across all translated content. LQA and MQM are best used on a representative sample. Smartling recommends sampling 2–8% of your translations monthly, depending on overall volume. If you are using LQA Suite, you can configure automated sampling to submit translations on a regular cadence.

LQA-type steps are specific to Smartling Language Services (SLS) and are used for conducting LQA on translations completed by SLS. LQA Agent can be enabled on LQA-type steps, but this configuration is handled by SLS and customers cannot enable it on these steps themselves. If you are not using SLS, or are using a third-party translation vendor, create a Quality Evaluation-type step for LQA Agent instead.

SLS workflows such as AIT and AIHT are covered by the SLS translation satisfaction guarantee. LQA Agent is designed to be used for LLM translation workflows outside AIT and AIHT. If you have questions about using LQA Agent in an SLS workflow, please contact your SLS Project Manager or Customer Success Manager to discuss your use case.

Reporting

Errors logged by LQA Agent appear in the same LQA reports as human-recorded errors: 

• MQM results in Job Details — view MQM quality scores directly in the job details view. The tab appears for any job that uses an MQM-enabled workflow, allowing you to quickly see pass/fail results for LQA assessments completed by LQA Agent or human reviewers.
• LQA Dashboard — a visual overview of MQM scores by locale, project, or job over time. Use it to track quality trends and identify which locales or projects may need attention.
• LQA Errors & Arbitration Report — a string-level view of all recorded errors, including error category, severity, and arbitration comments. Use it to drill into specific errors and review individual evaluations.
• LQA Error Density Report — an overview of error counts and density by project, language, and job.

MQM pass/fail thresholds do not automatically adjust based on whether errors were recorded by LQA Agent or a human reviewer. Keep this in mind when interpreting scores from LQA Agent-only evaluations without human review.

Best practices

• Use the Smartling LQA Agent MQM Schema template with Acceptable Penalty Points set to 100. This ensures your schema aligns with LQA Agent's detection capabilities and establishes an appropriate passing threshold.
• Sample rather than evaluate everything. We recommend sampling 2–8% of monthly translation volume. Use automated sampling for consistent, hands-off quality monitoring.
• Combine with human validation for the most accurate results. Targeted validation of flagged content can reduce human reviewer scope by up to 90% while significantly improving score accuracy.
• Keep LQA Agent and human evaluators on separate workflow steps. Do not assign human evaluators to the LQA Agent step. When content enters the LQA Agent step, it will be automatically evaluated and then submitted to the next step in the workflow. Add a second LQA-enabled step for humans to review the errors created by LQA Agent.
• Review translation issues regularly. If LQA Agent creates translation issues (because errors couldn't be mapped to your schema), review and convert them to LQA errors as needed.
• Modifying standard error types in the schema. If you remove or rename standard error types, LQA Agent may not be able to map its findings to your schema and will log them as translation issues instead of LQA errors.

Related articles

• LQA: Overview — learn how LQA works in Smartling, including key concepts like schemas, severity levels, MQM scoring, and how LQA differs from CAT Tool Quality Checks.
• LQA: Setup Guide — step-by-step instructions for creating a schema and enabling LQA on a workflow step.
• LQA: MQM Schema Templates — compare the available MQM schema templates, including error categories, severity weights, and recommended use cases for each.
• LQA Suite: Overview — LQA Suite lets you conduct LQA in a dedicated project, separate from your production translation project. It includes features like automated sampling and translation round-trip to support ongoing quality monitoring at scale.