How to judge AI part designs: 3-point DFM scorecard with ISO 2768-m — Build or Kill

This guide delivers a 3-point DFM scorecard with ISO 2768-m to judge AI part designs.

Use the fixed gate order to make a build or kill decision without reviewing later gates.

Why AI CAD Fails at the Tolerance Level

The only tolerance-level failure unique to AI-generated parts is inheritance: generative/AI CAD tools optimize for mass and load paths but emit unconstrained dimensions that default to whatever tolerance note the drawing carries — so the failure is invisible until quoting. Outputs from tools like Fusion 360 generative design and text-to-CAD tools do not emit GD&T callouts, so every feature inherits the drawing's general tolerance block. For that reason, ISO 2768-m must be declared explicitly, not assumed.

This becomes a scale problem. An AI part with hundreds of organic surfaces can carry hundreds of dimensions that inherit the drawing's general tolerance block. Each one silently becomes a commitment the shop must quote against under that block. The model can appear dimensionally complete in CAD while carrying no stated tolerance intent at all. Uncalled dimensions still require verification effort even if they are not individually labeled, so confirm the general tolerance note exists and is the intended class before scoring anything else.

The check method for this gate is one of declaration, not decoration. First, locate the general tolerance note on the drawing and confirm it exists. Second, confirm it is the intended class for this part, not a leftover note from another drawing. Do not scan the feature tree for callouts to substitute for this note, as their absence is the expected output of these tools. If the note is missing entirely, the part cannot be scored for this gate and must be treated as failing it.

This is strictly a gate of intent. The issue is not the organic form itself, but the lack of a stated tolerance scope. With platforms like MechSimulator offering interactive virtual simulators for precision measurement instruments, the act of measuring is treated as its own task, not an assumed byproduct of the model. If the general tolerance block is not explicit at the moment of review, the decision is to kill at this gate before moving to any later evaluation.

Gate 1: The Geometry Kill Test

Before any talk of tolerances, the only geometry gate to run first is the ordered geometry checklist — tool radius, depth-to-diameter, wall thickness, and undercuts — that must pass before any tolerance discussion. This order is not a suggestion. If the shape itself cannot be cut, probed, or held by the intended process, every minute spent past this point is wasted. Run these four in sequence, and stop at the first failure.

Start with the tool-radius check for machining. Confirm every internal corner radius is ≥ the shop's smallest end mill. A 0.020 in (0.5 mm) minimum tool radius is a practical floor for prototype work. Pull the radius value for each internal corner straight from the STEP file, not from a rendered view. If any AI-generated fillet falls below that floor, flag it as a kill unless the part moves to EDM or additive. A sharp internal corner is not a "tight" feature, it is an uncuttable one.

Next, run the depth-to-diameter check. For every hole or pocket, divide its depth by its tool diameter directly from the model. Any hole or pocket deeper than 3× its tool diameter needs special ops. AI lattices and generative channels routinely violate this, so measure, don't eyeball, in the STEP file. The question is not "can a tool reach it," but "can a tool reach it while clearing chips and keeping rigidity" for that ratio. If the answer is no for the intended setup, fail it here.

Third, the wall-thickness check: for the intended process, find the thinnest continuous wall and judge it against clamping and cutting, not just the load path. A wall can look efficient for mass reduction and still be too flimsy to hold without distortion under the forces of the cut. If fixturing cannot support that minimum wall for the chosen process, it fails this gate with no need to argue strength alone.

Finally, scan for undercuts. For each undercut, ask if the intended setup can reach it with a standard tool. If reaching it forces a setup change, a custom form tool, or a process swap just to access that one feature, treat it as a geometry kill. Run all four in this exact order, and the first one that cannot pass is your kill decision for Gate 1.

Gate 2: Scoring Against ISO 2768-m

Only run this gate after passing the geometry gate. For ISO 2768-m, treat the medium-class linear tolerances as a strict pass/fail audit: ±0.1 mm (6–30 mm), ±0.2 mm (30–120 mm), ±0.3 mm (120–315 mm), ±0.5 mm (315–1000 mm) for linear dimensions, plus the chamfer/radius rule. Pull every linear dimension from the AI drawing and place it onto that ladder by its nominal size. Mark each one as within or outside of m-class, with no rounding up to justify a pass.

Do not audit every dimension equally. Isolate only the critical dimensions that directly affect fit or function. For those, compare the required deviation to the m-class value for its exact size band. A function that needs tighter than ±0.1 mm on a 6–30 mm feature is already outside of m-class, and must be flagged the moment it is found. Apply the same scrutiny to any chamfer or radius that falls under the chamfer/radius rule, as those can be just as restrictive as a linear dimension.

The kill threshold for this gate is more than 3 critical dimensions falling outside m-class achievable tolerances. Once that count is reached, the design needs explicit tighter callouts. If the AI output cannot produce those explicit tighter callouts in a usable form, regenerate. Do not attempt to "split" violations across non-critical dimensions to dilute the count. The number is of critical dimensions only.

Keep a running tally as you go, and stop at the first time the count exceeds 3. This is not a grading exercise to see how close it can get, but a pass/fail count against a fixed ladder. If it fails here, do not move to the next gate. The decision is made at this count alone.

Gate 3: Rescue Cost — Fix or Regenerate

At this point, the only question to answer is the rescue-cost rule: any single fix costing more than one regeneration cycle (regeneration is near-zero marginal cost) means kill, with the two-sided cost of being wrong. The cost of wrong by killing is discarding a design that was fixable and wasting the review time already spent. The cost of wrong by building is committing to a failed machine setup or a scrapped batch. These are not equal weights, and they must be weighed against each other for this gate alone.

Do not count how many edits would be needed to "make it work". Only look for one fix action. If the path to buildability is a chain of changes instead of a single fix, that chain is already pushing past one regeneration cycle by nature, and the decision should be kill. The question is not effort of editing, but the cost of that action compared directly to regenerating once.

Because regeneration marginal cost approaches zero, bias toward kill. Before accepting any fix, verify the exact process against your shop's quoted rate for that process, not a general assumption. The scale tips the moment that single fix exceeds one regeneration cycle, and at that moment, no later consideration should enter the decision. Kill, and regenerate.

The Scorecard: Worked Example and Rules

The only new item to use for this step is a copy-usable 3-point scorecard template with pass/fail thresholds, plus the full worked audit of one hypothetical AI bracket showing a kill in under 15 minutes. The rule for it is strict: run in order of geometry → tolerance class → rescue cost, and the first gate it fails is the kill decision. Do not spend review time on the gates after it.

To see that in practice, take an AI-generated 6061 aluminum sensor bracket with these inputs: 85 mm overall, 12 mm thick, four M5 holes, one 0.3 mm internal fillet, and one 40 mm deep × 8 mm diameter pocket (D/d = 5). Checkpoint 1 is the fillet against a 0.5 mm tool-radius floor: 0.3 mm < 0.5 mm, so FAIL. Checkpoint 2 is the pocket depth at 5× diameter against a 3× limit: 5× > 3×, so FAIL. Checkpoint 3 is the hole pattern at m-class ±0.1 mm: PASS. With the order enforced, the verdict is KILL at Gate 1 in ~10 minutes, with the direct action to regenerate with a minimum-feature constraint added. The later fails are noted only to show they exist, not to justify continuing review.

For your own sheet, use this scorecard structure and apply it to the job at hand:

For Gate 1, confirm the floor values against the actual cutter set for that bracket, not a general guess. For Gate 3, do not assign a currency figure unless you have a written quote to compare against a written regenerate plan. The value of this scorecard is not completeness across all three, but the forced stop at the first fail to keep the review under 15 minutes.

Frequently Asked Questions

What is the maximum time allowed to make a build or kill decision?

Kill or build an AI-generated part design in under 15 minutes.

In what order must the three gates be scored?

The gates must be scored in the order of geometry → tolerance class → rescue cost, with the first gate it fails being the kill decision.

At which gate must ISO 2768-m compliance be verified?

Gate 2 requires ISO 2768-m tolerance compliance before evaluating rescue cost.

Why must ISO 2768-m be declared explicitly rather than assumed?

Outputs from tools like Fusion 360 generative design and text-to-CAD tools do not emit GD&T callouts, so every feature inherits the drawing's general tolerance block.

What makes tolerance inheritance a scale problem for AI parts?

An AI part with hundreds of organic surfaces can carry hundreds of dimensions that inherit the drawing's general tolerance block, with each one silently becoming a commitment.

How fast can an AI model design 3D printed bone scaffolds?

AI Model Designs 3D Printed Bone Scaffolds in 12 Seconds (3D Printing Industry).

Quick answers

Also worth reading: DFM Rule Checks in 2026: What 214 Concept Sprints Reveal: DFM Rule Checks in 2026: · Measuring ROI on AI Product Concepts: Measuring ROI on AI Product · AI CAD Defaults to Vertical Walls: Draft Angles & DFM Gaps: AI CAD Defaults to Vertical