Dassault Systèmes announced over a dozen AI features for SolidWorks at 3DEXPERIENCE World 2026. The marketing materials show engineers describing geometry in plain English, generating drawings on demand, and having AI explain why a feature failed. Most of that is real — but the gap between what runs on your desktop today and what requires a 3DEXPERIENCE subscription is larger than the announcements suggest.

This is a precise breakdown: which features ship, which platform they require, what they can and cannot do, and where API macros still outperform them.

Two completely separate AI tiers

SolidWorks AI features divide into two groups with no overlap in deployment model:

Tier 1 — Desktop-local, no cloud, released: Built into standard SolidWorks desktop (2024+), runs entirely on-machine, no internet connection, no additional licensing. These use local pattern-matching models, not large language models. They are production-stable.

Tier 2 — Cloud companions, 3DEXPERIENCE required, Early Access: Two AI agents — AURA (Knowledge Companion) and LEO (Engineering Companion) — that run on Dassault cloud infrastructure. They require: active 3DEXPERIENCE platform subscription, cloud connectivity during use, and token-based credits that are consumed per interaction. These use generative AI (LLMs + computer vision). They are in Early Access as of mid-2026, meaning rapid iteration and inconsistent results on complex geometry.

The practical consequence: approximately 85% of SolidWorks seat holders are on perpetual desktop licenses with no 3DEXPERIENCE connection. Tier 1 is their only option.

Tier 1: What runs on desktop today

Command Predictor

When enabled (Tools → Options → System Options → General → Enable Command Predictor), a small panel appears in the Command Manager showing 3-5 suggested next commands based on your recent workflow. The model is trained on aggregate usage patterns across SolidWorks users.

What it does well: Reduces menu navigation for common sequences. When you finish a base flange it correctly suggests Sketched Bend, Edge Flange, or Miter Flange depending on which you’ve used most. When you finish a sketch it surfaces Pad, Revolve, or Pocket based on context.

What it doesn’t do: Adapt to an unusual workflow, predict cross-document operations, or help when you’re off the beaten path. It predicts the most common next command for the current state — if your workflow is non-standard, predictions degrade quickly. It also cannot explain why you’d use one command over another.

Where it fits: Useful during learning (reduces trips to Help), neutral for experienced users who already know their sequences.

Fastener Recognition and Pattern Assistance

Analyzes assembly geometry to identify fastener instances (bolts, screws, pins) and assist with pattern-based mating. When you insert a fastener and start the Mate command, it proposes the concentric + coincident combination automatically.

What it does well: Eliminates the 2-click mate sequence for standard fasteners. Recognizes ANSI/ISO standard screw geometry patterns from the SolidWorks Toolbox library and third-party STEP imports that have the same geometry profile.

What it doesn’t do: Recognize non-standard fasteners with unusual head geometry, or handle inserts and rivets reliably. Still requires manual mate selection for anything outside the trained geometry library.

Selection Accelerators

During fillet and chamfer operations, when you select one edge, the tool predicts related edges that typically belong in the same fillet chain (tangent edges, loops, concentric edge patterns). A small “add related” suggestion appears.

Practical value: Measurable time savings on uniform fillets across sheet metal bends and machined pockets. Less useful for complex surfaced geometry where edge tangency is ambiguous.

Repair Sketch Relations and Repair Missing Mate References

When a sketch has broken or conflicting relations (the familiar red/yellow warnings), the Repair Sketch Relations tool analyses the under/over-defined state and proposes restorations. Repair Missing Mate References does the same for assembly mates when referenced geometry has been deleted or renamed.

What it does well: Identifies the root cause of “dangling” relation warnings that users typically fix by deleting and recreating relations manually.

What it doesn’t do: Fix relations that are geometrically impossible to satisfy simultaneously, or handle designs where the reference geometry was intentionally removed. In those cases it correctly identifies the conflict but cannot propose a valid resolution.

Other local features

Fastener Simplification: Recognizes over-detailed fastener bodies (fully-threaded solid thread geometry) and offers to replace them with simplified cosmetic thread representations — meaningful for large assemblies where render thread geometry tanks performance.

Pattern Driven Component Pattern: When a part contains a feature pattern, the assembly can automatically place matching components at the same pitch and count.

Driven Silhouette Part Defeature: Creates linked, simplified part representations for large assembly performance. Useful for reducing VRAM pressure when sharing assemblies with visualization-only recipients.

Tier 2: AURA and LEO (3DEXPERIENCE required)

AURA — Knowledge Companion

AURA is a chat-based assistant embedded in the SolidWorks UI when 3DEXPERIENCE cloud is connected. You type requests in natural language; AURA interprets them against the active document and executes or explains.

Assembly Performance Analysis: Given an open assembly, AURA identifies the highest-cost components by polygon count, thread geometry, and SpeedPak eligibility. It outputs a ranked list with specific actions (suppress cosmetic threads on X fasteners, create SpeedPak for Y sub-assembly). This is the most production-ready cloud feature — the recommendations are deterministic and grounded in geometry metrics.

What’s Wrong Analysis: When a feature rebuilds with an error, AURA can explain the failure in plain language and propose corrective action. For standard errors (missing reference, incompatible constraints) this works reliably. For complex failures in surfaced geometry it surfaces the official Help link rather than a specific fix.

Sketch Analysis: Identifies under/over-defined sketch conditions and proposes which relations to add or delete. More useful than the standard “over defined” warning alone, which doesn’t identify which relation is in conflict. Works best on sketch-heavy parametric models; struggles with imported geometry that has constraint artifacts from the source CAD.

Natural-Language Selection: “Select all holes smaller than 5mm diameter” executes a selection filter via natural language. Geometry-type selections (all fillets, all flat faces, all cosmetic threads) work consistently. Dimensional comparisons work for simple criteria. Boolean logic across multiple criteria (select all holes smaller than 5mm that are on faces with a specific material appearance) is unreliable.

Material Appearance Manager: Finds and reassigns appearances via text description. “Change all parts with aluminum appearance to the brushed steel appearance” works. More specific requests (change only parts in the top-level assembly, not sub-assemblies) require careful phrasing and verification.

LEO — Engineering Companion

LEO focuses on model creation and transformation tasks that previously required either deep API knowledge or manual reconstruction from scratch.

BREP to Parametric: This is the headline feature. When you import a STEP or IGES file, LEO can attempt to reconstruct the parametric feature tree — converting dead solid geometry into a sketch-driven history. It works reliably on prismatic geometry (extruded profiles, revolved bodies, simple machined parts). Sheet metal with bends, complex surfaced bodies, and imported assemblies with many components return partial results or fail.

The underlying approach is recognizable-feature decomposition: LEO detects faces that match known feature patterns (cylindrical boss = Extrude from circular sketch, pocket = Extrude cut, fillet = Fillet feature) and reconstructs a feature chain. What it cannot do: recover the design intent of a complex parametric model, infer configurable dimensions, or produce a feature tree that rebuilds efficiently on subsequent edits.

AI Drawing Creation: Describe what views you want and LEO creates a drawing. “Create a front, right, and isometric view on A3 paper with a title block” executes correctly. Dimension placement follows an automated algorithm and typically requires significant human correction — dimensions are geometrically present but not positioned for readability, and GD&T is not generated. This is a time-saver for boilerplate drawing setup, not a full drawing automation solution.

Assembly Instructions: Generates step-by-step manufacturing assembly instructions with callout images. Produces an exportable document. Useful for first-draft work instructions; still requires review for sequence accuracy on complex assemblies.

Macro Generation: LEO can generate VBA macros from a text description of the desired workflow. “Write a macro that iterates all sheet metal parts in the assembly, exports each flat pattern as DXF to a folder named after the part number.” The output is syntactically correct VBA more often than ChatGPT for SolidWorks-specific tasks, because LEO is grounded in the SolidWorks API documentation. However, it makes the same class of errors documented for AI macro generation generally — incorrect units, missing null checks, wrong interface versions — and requires the same review process.

The practical difference from using ChatGPT with the SolidWorks API docs: LEO outputs compile-ready code without requiring you to paste the API reference into a chat session. It also handles some common patterns (opening documents in silent mode, SaveAs4 arguments) more reliably. For complex workflows involving events, add-in registration, or property manager pages, it still produces incomplete stubs that require manual completion.

What Copilot cannot replace: where macros and add-ins still win

After going through every announced Copilot feature, these categories remain firmly in API-and-macro territory:

Batch operations: AURA/LEO operate on the currently active document. Iterating hundreds of assembly components, all DXF exports from a vault query, bulk custom property writes — none of this is achievable through the chat interface. These require a loop in code.

PDM integration: No Copilot feature reads from or writes to the vault programmatically. Workflow transitions, vault state queries, batch PDM operations all require EdmLib or the PDM API. See the SolidWorks workflow automation guide for the decision framework.

Repeatable, auditable processes: Copilot features are interactive and session-specific. A VBA macro or C# add-in runs identically every time and produces a log. For ISO/IATF processes where auditability is required, scripted automation remains the standard.

Complex drawing automation: AI Drawing Creation handles view layout and basic annotation setup. Parametric title blocks linked to model custom properties, BOM-driven balloon placement, revision table management, and sheet metal flat pattern views all require the drawing API. The macro-driven drawing creation guide remains the authoritative approach for these.

Custom selection logic: Natural-Language Selection handles simple filters. Multi-pass, logic-heavy selections (all parts matching a supplier code in custom properties, all components with mass above threshold in a specific sub-assembly) require IAdvancedSelectionCriteria or a property-traversal loop.

The platform split in practice

If your organization runs perpetual SolidWorks licenses with no 3DEXPERIENCE subscription:

  • Tier 1 features are already available — turn on Command Predictor, enable Fastener Recognition.
  • AURA and LEO are unavailable. There is no workaround short of acquiring a 3DEXPERIENCE subscription.
  • ChatGPT with the SolidWorks API documentation pasted in is the practical equivalent for macro generation — with the same quality caveats.

If your organization is on a 3DEXPERIENCE subscription (SolidWorks Connected or SOLIDWORKS Cloud Offer):

  • All Tier 2 features are accessible once Cloud Services are activated in the SolidWorks options.
  • AURA and LEO are in Early Access — expect iterative quality improvements over 2026, and treat outputs as starting points requiring engineering review.
  • The most mature cloud feature is Assembly Performance Analysis. The most experimental is BREP to Parametric on surfaced geometry.

When evaluating whether Copilot replaces your current VBA automation: the answer depends on batch scale and repeatability. Single-document interactive tasks where guidance or first-draft generation is valuable are where Copilot earns its keep. The macro vs add-in upgrade decision framework still applies to everything that runs at scale or requires system integration.

What to expect over the next 12 months

Dassault’s trajectory from the 2026 release cadence:

  • BREP-to-Parametric will expand coverage from prismatic to sheet metal geometry (explicit roadmap item).
  • Macro Generation is planned to move beyond single-macro output toward multi-file add-in stub generation.
  • The token/credit model for cloud features is expected to shift — currently each interaction consumes credits from a pool; Dassault has signaled a move toward usage-based subscription tiers to remove the per-interaction cost friction.

What is not on the roadmap: offline LLM inference on desktop (Tier 1 features will remain pattern-based), vault-integrated batch operations through Copilot, and automated GD&T generation from design requirements.

The split between what runs locally and what requires cloud is not a transitional state — it reflects a deliberate architecture where desktop SolidWorks remains the modelling environment and 3DEXPERIENCE cloud is the intelligence platform. Engineers planning automation strategies should assume this split is permanent and design accordingly.