Limitless CNC and Dirac: The 80/20 Rule of Manufacturing AI — Augment the Human, Don't Replace Them

Company Profiles

Limitless CNC is a Tel Aviv-based startup founded by David Priev that applies AI to CNC programming — specifically, the generation of NC toolpath programs from CAD geometry and machining specifications. CNC programming is currently a skilled trade: experienced programmers use CAM software to manually select cutting strategies, define toolpaths, set feeds and speeds, and verify the result. For routine parts with standard features, this work is repetitive and well-characterized. For complex, non-standard parts, it requires deep expertise. Limitless CNC targets the routine 80%.

Dirac was founded by Filip Aronstein to apply AI to manufacturing work instructions — the step-by-step documentation that operators follow to assemble, inspect, or test products. Creating work instructions is currently labor-intensive: a manufacturing engineer documents each step, often by memory and tribal knowledge, creating instructions that may be incomplete, inconsistent, or out of date. Dirac uses AI to generate work instructions from CAD models, process plans, and existing documentation, dramatically reducing the manual effort while improving completeness.

Both companies serve mid-sized manufacturing organizations — the segment that is large enough to have systematic process documentation requirements but not large enough to have dedicated teams for CNC programming optimization or work instruction management.

---

The 80/20 Framework

The framing that both Limitless CNC and Dirac operate from is worth examining before the technology: the 80/20 augmentation model.

In any manufacturing engineering workflow, a subset of tasks is routine, well-characterized, and high-volume. These are the tasks where the output is predictable if the inputs are defined correctly — standard CNC programs for common feature types, work instructions for standard assembly sequences, quality inspection steps for known acceptance criteria. This is the 80%.

The other 20% — the novel parts, the difficult setups, the edge cases, the non-standard assemblies, the quality escapes that don't fit the pattern — requires human judgment, domain expertise, and accountability. This is the work that experienced engineers and machinists are actually good at and cannot be automated without unacceptable risk.

AI deployment that targets the 80% routine work gets adoption because it makes experienced people more productive without threatening their role in the 20% that matters. AI deployment that attempts to automate the 20% gets resistance — and often fails technically as well, because the 20% is difficult by definition.

Both Priev and Aronstein are explicit about this framing. It is not a marketing hedge. It is a deployment strategy that reflects how manufacturing organizations actually adopt new technology.

---

What Limitless CNC Built

CNC programming today follows a workflow that has not fundamentally changed in 30 years:

1. Import CAD geometry into CAM software
2. Define the machine, tooling, and material
3. Select machining strategies for each feature (pocket, contour, hole, surface)
4. Generate toolpaths and verify clearance
5. Post-process to machine-specific G-code
6. Prove-out on the machine

For complex aerospace or medical parts, steps 3–4 require significant expertise and may take days. For routine prismatic parts with standard features — brackets, plates, flanges, housings — steps 3–4 are largely repetitive: the programmer applies the same strategies they have applied hundreds of times before.

Limitless CNC's AI engine handles the routine programming workflow. Given a CAD file, material, and machine specification, the system:

• Recognizes standard feature types (pockets, bores, profiles, faces) and their machining requirements
• Selects appropriate cutting strategies from a strategy library trained on experienced programmer decisions
• Sets feeds and speeds based on material, tool, and feature geometry
• Generates G-code for the selected machine, ready for prove-out

For standard parts, the programmer receives a first-pass NC program rather than starting from a blank CAM session. For complex parts, the AI generates the straightforward features and flags the difficult ones for human attention.

The productivity impact: experienced CNC programmers using Limitless CNC report 50–70% reduction in programming time for routine part types, with first-pass programs requiring minimal editing. More importantly, junior programmers can handle a much higher proportion of the quote volume independently, freeing senior programmers for complex work.

---

What Dirac Built

Work instructions have a documentation problem that is almost the inverse of the CNC programming problem: instead of a skilled person doing repetitive work, you have a skilled person documenting what they are doing in a way that someone less experienced can follow. The documentation is incomplete because the expert takes for granted the knowledge that the novice needs.

Dirac's AI generates work instructions from multiple inputs:

CAD geometry: The system interprets assembly geometry — components, interfaces, fasteners, clearances — and generates step-by-step assembly sequences that are geometrically feasible.

Existing documentation: PDFs, old work instructions, engineering notes, and process plans that exist in the PLM or document management system provide context about the process that the AI incorporates.

Expert input: Dirac's interface allows manufacturing engineers to review AI-generated draft instructions and add context, corrections, and domain-specific warnings. The AI handles the documentation framework; the expert handles the knowledge gaps.

Visual generation: Dirac generates illustrated work instructions with annotated 3D views of each assembly step, not just text descriptions. This dramatically reduces the ambiguity in step interpretation.

The output: work instructions that are more complete, more consistent, and faster to produce than manually authored instructions. Programs that previously had work instruction gaps — steps that experienced assemblers just "knew" but weren't documented — get explicit documentation.

The tribal knowledge capture value is significant: when an experienced assembler's knowledge is encoded in Dirac-generated instructions, it becomes part of the program record rather than departing with the employee.

---

Results

Limitless CNC outcomes:

• CNC programming time reduction: 50–70% for standard part types, with experienced programmers reporting full shift productivity gains on routine work
• Junior programmer capacity: customers report 2–3x increase in parts programmable by junior programmers without senior review, freeing senior capacity for complex work
• Quote support: faster program generation enables faster cycle time estimates, which reduces quote turnaround (complementary to quoting tools like up2parts)

Dirac outcomes:

• Work instruction authoring time: 60–75% reduction in first-draft authoring time for standard assembly processes
• Completeness: AI-generated instructions systematically include steps that manual authors frequently omit (torque specifications, orientation callouts, inspection checkpoints)
• Onboarding acceleration: programs with complete Dirac-generated work instructions report 20–30% faster qualification of new assemblers to production standards
• Tribal knowledge capture: manufacturing engineers report capturing process knowledge from retiring workers more completely and efficiently using Dirac's structured documentation workflow

---

Lessons Learned

1. The augmentation framing is not optional — it is the adoption strategy. Both companies have found that "AI replaces the programmer/engineer" framing kills adoption regardless of technical quality. "AI handles the routine work so you can focus on the complex work" gets traction. The framing has to be true, and both companies built accordingly.

2. Routine work is expensive even though it's easy. The 80% of routine work that AI targets is not trivial in cost terms — it consumes the majority of experienced workers' time. Shifting even half of it to AI creates substantial capacity for complex work.

3. Completeness outperforms elegance in work instructions. The highest-value improvement from Dirac-generated instructions is not style or format — it is completeness. Systematically including steps that human authors skip reduces the defect rate in work instruction-driven processes.

4. Tribal knowledge capture requires a workflow, not just a repository. Documentation systems exist in every manufacturing company. The barrier to tribal knowledge capture is not storage — it is the workflow for extracting knowledge from experts in a structured, reusable form. Dirac's structured instruction generation is that workflow.

5. Augmentation requires clear human accountability for the 20%. The 80/20 model only works if the 20% that humans remain responsible for is actually well-defined. Programs that blur the boundary between AI-handled and human-accountable tasks get worse outcomes than programs with clear delineation.

---

Implementation Advice

For manufacturing organizations evaluating CNC programming or work instruction AI:

Start with a pilot on your most routine, high-volume part families or assembly processes. These are the places where the AI is most accurate, adoption friction is lowest, and ROI is clearest. After demonstrating value there, expand to progressively less routine work.

Involve the experienced practitioners early. CNC programmers and manufacturing engineers who help configure the AI — teaching it their preferred strategies, reviewing its outputs, correcting its errors — become advocates rather than resistors. Their expertise is the AI's training signal.

Measure tribal knowledge capture explicitly. Count the number of processes that existed only in experienced workers' heads before AI-generated documentation, and track how many are now documented. This is a real business value that is invisible in standard productivity metrics.

---

About the Source

This case study is drawn from AI Across the Product Lifecycle Episode 14, a podcast conversation with David Priev (CEO, Limitless CNC) and Filip Aronstein (CEO, Dirac). See also: [[CNC Machining PLM]], [[Work Instructions in PLM]], [[AI in Manufacturing]], [[Knowledge Management in PLM]].