The decision that shapes everything

You almost certainly cannot embed commercial CAM in a product you sell. That single licensing fact — not a technical one — is what pushes manufacturing SaaS startups toward building their own engine. Establish it early, before your architecture assumes otherwise.

Instant-quoting platforms, automated job shops, and design-for-manufacturing tools all eventually need machining intelligence available programmatically. Here is how to scope that without building a CAM system.

What your product actually needs

Founders often say "we need CAM" when they need one specific capability from it. Separating these changes the cost by an order of magnitude:

CapabilityNeedsRelative effort
Manufacturability checkFeature recognition, rule checkingLow
Cycle-time estimationFeature recognition, removal modelLow–medium
Material and cost estimationStock calculation, pricing rulesLow
Setup planningReachability analysisMedium
Production-ready toolpathsFull offsetting, verification, postsHigh

Most manufacturing SaaS never needs the bottom row. An instant-quoting platform needs to know how long a part takes and what it costs — not to emit G-code a machine will run. That distinction is the difference between a tractable project and a multi-year one.

Feature recognition is the core capability

Nearly everything above depends on one thing: looking at an imported solid and identifying what has to be machined. Holes, pockets, slots, faces, bosses, threads.

  • Rule-based recognition works well on prismatic parts with standard features and is the sensible starting point.
  • It degrades on freeform surfacing, unusual construction, and organic shapes.
  • Recognition confidence matters. Your system should know when it is unsure and route those parts to a human rather than quoting confidently on a misread.
  • It runs on imported geometry, so healing and tolerance handling come first — bad input produces bad recognition.

Cycle-time estimation without full toolpaths

You can estimate machining time credibly without generating the actual path:

  1. Compute material to remove — stock volume minus part volume, broken down per feature.
  2. Assign a strategy per feature — roughing, finishing, drilling — from rules rather than real toolpaths.
  3. Apply removal rates from your material and tooling data.
  4. Add setup, tool change and non-cutting time from empirical figures.
  5. Calibrate against actual jobs — this step is what turns a plausible model into a trustworthy one.

The calibration step is where credibility comes from. A model tuned against a hundred real jobs on your machines beats a theoretically rigorous one that has never met a shop floor.

Architecture that keeps options open

Separate geometry, recognition, and business logic strictly. Geometry handling (import, healing, queries) should sit behind an interface, so a future kernel change is a port rather than a rewrite. Recognition should output a neutral feature description that quoting, planning and any future toolpath generation all consume independently.

  • Geometry layer — import, heal, query. Open CASCADE is the usual pragmatic choice given the licensing constraint.
  • Recognition layer — geometry to features, with confidence scores.
  • Rules layer — your machining knowledge, ideally configurable rather than compiled in.
  • Application layer — quoting, planning, whatever your product does.

That last separation matters commercially: your machining rules are the asset. Keeping them in configuration rather than in code means shop-floor experts can refine them without a developer.

The licensing reality

Worth stating plainly because it catches founders late:

  • Commercial CAM licences generally prohibit embedding in software you distribute or sell as a service.
  • CAD kernel royalties can be incompatible with per-user SaaS pricing — model this before choosing.
  • Open CASCADE is LGPL, which is usable for SaaS but carries obligations worth reviewing with a lawyer.
  • Post-processors from commercial systems are typically not redistributable.

A staged build

  1. Import and heal geometry reliably. Unglamorous, and everything depends on it.
  2. Recognise the features your target parts actually contain — not a general recogniser.
  3. Estimate cost and time, calibrated against real jobs.
  4. Add manufacturability feedback, which customers value highly and which is comparatively cheap.
  5. Only then consider toolpath output, and only if your business genuinely requires it.

Narrow the part space

The most effective scoping decision is limiting what your platform accepts. A system handling 3-axis prismatic parts in aluminium and steel is achievable. One accepting arbitrary geometry in any material is not — and the narrow version can still address a large market while being honest about what it declines.

Building a manufacturing platform that needs machining intelligence? Tell us what your product must decide about a part. See our CAM service, toolpath fundamentals, and custom vs commercial CAM.

Frequently asked questions

Generally no. Commercial CAM licences are written for human operators at workstations, not for programmatic use inside a product you sell. Some vendors offer OEM or API arrangements, but terms and cost vary enormously — establish this before building an architecture that depends on it.
Consistently accurate matters more than precisely accurate. A quote that is reliably within a known margin lets you price with a defensible buffer. One that is usually right and occasionally wildly wrong is worse than no automation, because you cannot tell which case you are looking at.
Quoting, almost always. Estimating cycle time and material needs far less geometric machinery than generating production-ready G-code, it validates demand quickly, and much of the geometry work carries forward if you later add real toolpath output.