Encord

Learning Objectives

• Explain what an AI data platform does and why annotation, curation, and search are treated as one product rather than three
• Identify the modalities Encord handles beyond images and text, and why those matter for robotics
• Evaluate when a managed data platform beats open-source labelling tools or an in-house pipeline

What Is Encord?

Encord is a data infrastructure platform for teams that train models on multimodal data. It was founded in 2020 by Eric Landau and Ulrik Stig Hansen, who run the company as co-chief executives, and it is headquartered in London with US operations. The company went through Y Combinator and has raised $110 million in total, most recently a $60 million Series C led by Wellington Management in February 2026.

The pitch is that preparing training data is one workflow, not several. Most teams stitch together a labelling tool, a storage bucket, some notebooks for quality checks, and a spreadsheet tracking which annotator did what. Encord's argument is that those stages share the same underlying question — which data actually matters? — and that answering it well requires the platform to understand the data semantically rather than just store it.

The Product Lineup

Encord splits into three surfaces that share one data layer:

• Annotate — collaborative labelling across images, video, audio, 3D point clouds, and sensor data, with review workflows and per-annotator quality tracking. Specialist modalities including LiDAR, DICOM medical imaging, geospatial, and ECG are available as add-ons.
• Active — data curation and quality evaluation. It scores a dataset for label errors, class imbalance, and near-duplicate frames so teams can cut what adds nothing and buy more of what does.
• Index — search and management across very large datasets. This is the surface that makes petabyte-scale collections navigable rather than merely stored.

Layered over these are data agents, which put a model in the loop to pre-label or triage before a human reviews the result.

Where the Machine Learning Actually Sits

This matters for judging whether Encord is an AI product or a database with AI marketing. The learned components are load-bearing in two specific places.

Embedding-based retrieval. Index and Active work by embedding the underlying data — frames, clips, point clouds — into a vector space, then using distance in that space to surface near-duplicates, outliers, and semantically similar examples. Without the embeddings, searching five petabytes of unlabelled video for "the frames where the forklift is partially occluded" is not a feature you can offer.

Model-assisted annotation. Auto-segmentation and data agents produce a first-pass label that a human corrects, which is a substantially cheaper motion than labelling from scratch. The quality of that first pass is a direct function of the model doing it.

The Physical AI Bet

Encord's clearest strategic position is in physical AI — robots, drones, and autonomous vehicles. The company says revenue from physical-AI customers grew tenfold over the twelve months to February 2026, and that data hosted on its platform went from one petabyte to more than five in the same period. Named customers include Woven by Toyota, drone maker Skydio, geospatial-intelligence firm Vantor, and AXA Financial. Around 300 AI teams use the platform.

The reason robotics is a good market for a data platform is that robotics data is awkward in ways text is not. A robot generates synchronised streams from several cameras, depth sensors, and joint encoders, all of which have to stay aligned in time. Labelling one frame in isolation is close to useless. Platforms that only ever handled images do not extend to this cleanly.

Encord has also experimented at the edge of the category — it has run trials with the German neuroscience startup Zander Labs using head-mounted sensors that read mental states such as error, intent, and surprise, alongside first-person video and forearm muscle sensors, to enrich robotic manipulation datasets. That work is a trial rather than a shipping feature, and it is worth reading as a signal about how badly the field wants training signal, not as a product you can buy.

Pricing

Only the Starter tier carries a published price. Team and Enterprise are quoted per organisation, and specialist modalities such as LiDAR, DICOM, geospatial, and ECG are priced as add-ons. Budget for a sales conversation rather than a credit card if you need anything past the five-user ceiling.

Getting Started

1. Sign up for Starter at encord.com — free for up to five users, enough to label a real pilot dataset rather than a toy one
2. Upload a representative slice, not your whole corpus — the point of the trial is to see whether the curation scoring finds problems you did not already know about
3. Run Active against it and read the label-error and duplicate reports before you annotate anything else
4. Try a data agent on a subset to measure how much of the first-pass labelling you can actually keep
5. Price out the modality add-ons early if you need LiDAR, DICOM, geospatial, or ECG — they are not in the base tiers

Limitations & Considerations

• Pricing opacity above Starter: no per-seat or per-annotation rate is published for Team or Enterprise, which makes cost modelling hard until you are in a sales process.
• Five-user free ceiling: generous for evaluation, tight for any real annotation team, so the free tier is a trial rather than a small-team home.
• A crowded category: Scale AI, Labelbox, Roboflow, and open-source options such as CVAT and Label Studio all overlap here. Encord's differentiation is breadth of modality and the curation layer, not labelling alone.
• Auto-labelling quality varies by domain: model-assisted annotation saves the most time on common objects in ordinary conditions and the least on rare classes and unusual sensors — which is often exactly where the hard data is.
• Vendor lock-in on curation: annotations export cleanly, but the embedding indexes and quality scores that make the platform useful are not portable, so switching costs accrue in the layer you came for.

Best Use Cases

Key Takeaways

• Encord is an AI data platform covering annotation, curation, and search across images, video, audio, 3D point clouds, and sensor data
• Founded in 2020 by Eric Landau and Ulrik Stig Hansen, London-based, and backed by $110 million in total funding after a $60 million Series C led by Wellington Management in February 2026
• Embeddings and auto-labelling models are load-bearing rather than decorative — they are what make petabyte-scale search and model-assisted annotation possible
• Its strongest position is physical AI, where multi-sensor data is genuinely hard to handle and image-only tools do not extend
• The free Starter tier caps at five users, and everything above it is custom-quoted, so plan for a sales conversation