How UiPath 10x'd Their Dataset and Improved Model Accuracy to Near 99% with Encord

Introducing Customer: UiPath

UiPath started as an RPA (Robotic Process Automation) company, enabling people to build UI-based automations using a low-code environment. As the company has expanded, so too have the modalities it works with, moving well beyond images of UI screens or scanned documents to encompass videos of automations, text models for tool-use and customer support agents, and voice models for real-time agents.

With that breadth comes a significant annotation and data management challenge: keeping training data high quality, well-tracked, and scalable across a diverse and growing set of model types.

Problem: Annotation at Scale Across Multiple Modalities 

Annotating data sounds straightforward until you're doing it with multiple people working on the same dataset simultaneously. That is when coordination overhead, inconsistent workflows, and limited visibility into pipeline progress can quickly become a bottleneck.

Before Encord, UiPath had tried annotating with a different vendor, but found it painful, particularly when dealing with edge cases, where turnaround times were frustratingly slow. The team needed a platform that could match their pace of iteration, provide genuine pipeline transparency, and handle the complexity of multi-reviewer workflows across diverse data types.

For UiPath's table extraction model, a submodel of their broader UI understanding system, data quality was especially critical. The model structure is complex, requiring both images and labels to be precisely correct. Any annotation platform would need to support in-place editing and multiple review flows to keep their internal quality bar high.

Solution: A Flexible, Transparent Platform Built for Iteration

UiPath migrated to Encord after their Director, Applied Science, picked up the SDK, experimented with it, and found it to be both powerful and easy to move in and out of as needed.

According to their Senior Manager, Enterprise BI Analytics, a key factor in the decision was flexibility and infrastructure. The team needed a data partner that genuinely understood their use case. One that could move quickly without sending them in the wrong direction. When Encord proactively surfaced inefficiencies in the annotation process and helped address them, it confirmed they had chosen the right partner.

Encord’s native support for multiple simultaneous annotators, with fast sync between them, proved to be a significant operational upgrade. The team was able to use Encord to improve data they already had, and to iterate on their labeling process to find the right balance between speed and quality. Pipeline visibility made bottlenecks transparent and addressable in real time.

For complex annotation tasks like the table extraction model where both the image and its labels need to be exactly right, Encord's review flow gave UiPath the structured internal review process they needed.

Results: Near-99% Accuracy and a 10x Dataset Expansion

The impact of moving to Encord was most visible on UiPath's table extraction model. Before, the model was performing at around 96–97% mAP on their internal dataset, which is already strong. After using Encord to more than 10x the size of the training dataset, the team achieved more than a 4x reduction in error rate, pushing accuracy above 98% and close to 99%.

Critically, the improvement wasn't just in raw numbers: the expanded dataset also brought the model's distribution much closer to real production data, making the gains durable rather than artefacts of benchmark overfitting. Customers noticed the difference.

Beyond the table extraction model, Encord enabled UiPath to scale a handful of models across their growing portfolio, freeing the team from the overhead of starting each new project from scratch, and giving them the headspace to focus on model development rather than data infrastructure.