Automating Excellence: Integrating TAUS Epic QE into Trados via Custom.MT

Ready to see QE in Trados in action? [Watch the full demonstration on YouTube]

For years, the standard localization workflow has remained static: run text through a translation memory, hit it with machine translation, and then send every single word to a human for review.  But as AI and MT engines have improved, this “review everything” model has become an expensive relic.

In a recent deep-dive webinar, experts from TAUS and Custom.MT demonstrated how Automated Quality Estimation (QE) is finally breaking this bottleneck by allowing teams to focus human excellence only where it truly matters.

The Shift from Niche to Necessity
We are seeing a fundamental shift in how global content is managed. Gartner predicts that by 2027, every leading translation service will use Quality Estimation as a standard part of their workflow. It is no longer a “niche experiment” but a baseline for efficiency. The reason is simple: data shows that 50% or more of MT output is often already good enough to publish immediately. Continuing to pay for human review on that “good enough” content results in millions in wasted spend. 

How TAUS Epic and Custom.MT Work Together
The webinar showcased the Epic API, described as an “AI quality companion”. Epic provides two critical functions:
● Quality Estimation (QE): It assigns a score to every translated segment, identifying what is publication-ready and what needs help.

● Automatic Post-Editing (APE): For segments that fall just below the quality bar, the system can use Large Language Models (LLMs) to automatically fix fluency or terminology errors before a human even sees them.

Custom.MT acts as the bridge, integrating these capabilities directly into the Trados environment. Through the Custom.MT console, users can orchestrate workflows that combine various engines—like Claude, OpenAI, or DeepL—with TAUS’s QE scoring.

The Power of the Threshold
The most critical concept discussed was the “threshold”. Since QE is a machine-calculated estimation of risk, users must decide what score (e.g., 85 out of 100) constitutes “good”. When a segment exceeds this threshold in Trados, the plugin can automatically confirm and lock the segment. This allows linguists to skip these sections entirely, focusing their energy on the “poor” segments, which appear with a bright red warning bar in the Trados editor.

Managing Risk and ROI
While saving up to 80% on translation costs sounds like marketing offer, the webinar broke down the math. On an annual spend of $2 million, moving from a traditional human-review-heavy workflow to an automated QE+APE model can reduce costs to approximately $90,000 by limiting human intervention to only the most critical or low-quality 20% of content.

However, the speakers were clear that this is a “risk exercise”. For mission-critical content like legal or medical translations, a higher threshold and human supervision remain vital. The goal isn’t to remove humans, but to automate the “unnecessary waste” in the review phase.

Getting Started: Calibration is Key
To implement this successfully, the experts recommended a four-step approach:
1. Select the right content: Choose a data set that represents your typical projects.
2. Calibrate your system: Find the “sweet spot” where you catch the most mistakes without losing the benefits of automation.
3. Use RAG for context: Upload your own translation memories and glossaries to the Epic API to ensure the AI understands your specific brand voice.
4. Analyze the results: Use the Custom.MT QE report in Trados to see a full breakdown of “good” vs. “poor” segments before starting the post-editing phase.

Automated QE is no longer a future prospect—it’s a functional tool available now for LSPs and enterprises looking to scale. As the industry moves toward 2027, those who master these AI-driven workflows today will be the ones leading the market tomorrow.