Building What’s Next: Agent Workflows in Speckle

What's next for Speckle

We're excited to share some sneak peeks at what's coming next in Speckle.

Over the last year, we've helped you put your BIM model data to work like never before: dashboards across your portfolio, data validation checks that run automatically on every model version, and Speckle AI to ask questions across every project.

What's next? Let our agents run those processes for you, so your team can focus on the work that matters.

Remember, agents are only as good as the data they can reach. That's why Speckle’s agents are built on the same open, governed data layer you already use, and work across every authoring tool and CDE in your stack. They do the repetitive work so your experts make the calls.

Read on to learn how you can avoid moving from data siloes to agent siloes!

Building in the open

We're building our agents with the teams who will use them, on live projects. This page is where we'll share each workflow as it takes shape: what it does, how it works, and what it's delivered so far.

Bookmark it. We'll add a new workflow every few weeks between now and launch.

Workflow 01: Drawings that carry their data

Our first workflow makes the drawing set readable by AI. When Speckle extracts drawings from a Revit model, each PDF now ships with the properties of every element drawn on it, embedded in the file.

The problem

Drawings are still how most project information moves between companies: to clients, contractors, consultants and authorities. But a PDF is a dead end. The model behind it knows every wall type, material and fire rating. The PDF only knows lines and text.

So anyone downstream who needs that information has to work it back out: reading the linework, re-keying it by hand, or running OCR and trusting the confidence scores. Drop a drawing into an AI tool today and it does the same thing, guessing from pixels.

How it works

1. Extract drawings from your model. Speckle generates the sheets straight from your Revit model.
2. Embed the element data. Each PDF carries the properties of every element drawn on it as parquet attachments, a slice of the model's own tables.
3. Add a hidden text layer. Any LLM that only reads the page still knows what's drawn where.
4. Ship it as normal. The PDF looks identical to the one you'd print today. Nothing changes for people who just need to look at it.

What it looks like in practice

Drag one of these PDFs into Claude and ask: "What's the fire rating of the exterior walls on this plan?" It answers from the model's data, not from squinting at linework. No OCR, no bounding boxes, no confidence scores.

The drawing stays a drawing. It just stops being a dead end.

Coming next

We'll add each of these here as we share them:

• Workflow 02: Auto-classification for cost, filling in missing or incorrect codes so estimators can price from the model with confidence.
• Workflow 03: Clash detection and prioritization, surfacing the few clashes that need a person
• Workflow 04: Model QA/QC, checking every model version against your standards
• Workflow 05: Quantities and estimates that update the moment the model changes

Build it with us

We're working with a small group of Build Partners to shape Speckle Agents on live projects before launch. Partners get agents tailored to their workflows, hands-on delivery support, and a measured ROI report for their business case.