# AI product ideas: 7-Point DFM Rubric vs 65% Scrap — Build or Kill

Charlotte Higgins · November 4, 2026

> Score AI product ideas against 7 DFM points before CAD spend. Use a 65% scrap baseline, replace it with your own tooling trial data, and kill failing concepts early.

| Takeaway | Detail |
| --- | --- |
| Score every AI-generated concept against all 7 DFM points before any CAD or prototype spend. | The 7 points: minimum feature size, draft, wall thickness, undercuts, tolerance stack, material availability, tool access. |
| Treat 65% scrap as a baseline you measure from your own first-pass tooling trials, not a cited industry rate. | The 65% figure is a starting baseline; replace it with your own measured first-pass tooling trial scrap rate. |
| Kill any concept that fails a DFM point before generating a prototype. | Reader rule: all 7 points must be scored before any prototype is generated. |
| AI prototyping compresses idea-to-demo time, widening the prototype-to-product gap. | AI coding tools let teams validate a concept and build a demo faster than ever, per GeekyAnts. |

This guide applies a 7-point DFM rubric to AI-generated product concepts before any CAD or prototype spend.

It shows how scoring minimum feature size, draft, wall thickness, undercuts, tolerance stack, material availability, and tool access kills unbuildable ideas early, and how to measure your own scrap baseline from first-pass tooling trials.

![AI product ideas](https://static.mm-ais.com/article-images-ai/ai-product-ideas-7-point-dfm-rubric-vs-6-ai-98d0a4b6.jpg)

## Why AI Concepts Die at the Tool

Generative ideation tools, from LLM concept generators to text-to-CAD pipelines, are trained on visual and textual plausibility, not manufacturability. Because of that training objective, the geometry they output is optimized to look correct on screen, not to release from a mold or be cut by an end mill. This is the mechanism at play: these systems produce geometry that fails at first tooling contact, routinely emitting wall thicknesses, undercuts, and fillets that no common tooling can produce. The failure does not show up in the AI demo itself, it only surfaces at DFM review, which is exactly why so many of these concepts die the moment they reach a toolmaker.

As pointed out by GeekyAnts, the "prototype-to-product gap" is already a known hurdle for software. For hardware, that gap doubles. An AI-generated demo can validate a concept's intended function, but nothing in the generation loop checks whether that part can actually be formed or machined. A shape that can be rendered with no issue can still be geometrically locked in place, with no path for ejection or tool access, despite looking deceptively simple.

The danger is timing. Since plausibility is the reward signal, the concepts that appear the most polished and organic are often the ones most likely to hide tooling conflicts. These conflicts are not edge cases of creativity, they are direct byproducts of a system that has never been asked, "can this come out of steel?" until a human asks it. By the time that question is posed to a toolmaker, the concept has already been treated as viable by the team that generated it.

For this reason, the first check for any AI-generated hardware concept must be a tooling-based check, not a visual one. Do not ask if the part looks buildable, ask if its geometry could physically interact with a cutter or a mold cavity without collision or entrapment. If that single question cannot be answered with a concrete reason, the concept should not move to any form of physical commitment. The kill needs to happen at the idea-to-tool boundary, because that is the exact boundary these tools were never designed to respect.

![Why AI Concepts Die at the Tool — AI product ideas](https://static.mm-ais.com/article-images-ai/ai-product-ideas-7-point-dfm-rubric-vs-6-ai-5bf06e87.jpg)

## The 7-Point DFM Rubric Defined

This section alone defines the seven rubric points and their pass/fail thresholds as a single scoring instrument to be applied to every AI-generated product concept before any CAD or prototype spend. Point 1 — Minimum feature size: every internal feature must clear your process floor (0.020-inch end-mill radius for CNC, per the site's established constraint). If any pocket, hole, or slot cannot physically fit that radius, it fails outright. Point 2 — Draft: all mold-direction walls need 1–2 degrees unless the concept explicitly justifies a side-action. A straight wall with no declared draft is not a refinement, it is a fail.

Point 3 — Uniform wall: for molded parts, the max/min wall ratio ≤ 1.5. Any abrupt transition that pushes past this ratio must be corrected or the concept fails, as uneven wall sections create an unworkable path for the intended molding process. Point 4 — Undercuts: zero enclosed undercuts without a declared side-action or collapsible core. If the part cannot be ejected along the primary mold direction without trapping, the concept must either declare one of those actions or be rejected at this gate. Point 5 — Tolerance stack: critical dimensions must chain through a stated datum scheme. If the features that drive fit have no verifiable chain between them, the concept fails this point, since the error cannot be contained by the process alone.

Point 6 — Material availability: the proposed resin or stock must be obtainable for the intended process without forcing an unplanned material change to satisfy the geometry already passed by points one through five. The check is access, not preference. If no grade can meet those requirements for that process, the concept fails. Point 7 — Tool access: the required tooling elements, whether a primary mold, side-action, or specific cutter set, must be compatible with the shop path you intend to use for first-pass tooling. If the concept demands any element outside of that path with no declared alternative, it fails this point.

Each of these seven must be marked pass or fail on the same scoring sheet for the concept. No point may be skipped, and a single fail is enough to kill the concept at this gate. By forcing all seven to be answered at the same time, this instrument prevents trading one violation for another and makes the kill decision before any CAD or prototype spend.

![The 7-Point DFM Rubric Defined — AI product ideas](https://static.mm-ais.com/article-images-pixabay/ai-product-ideas-7-point-dfm-rubric-vs-6-0acf3616.jpg)

## Evidence: Where Scrap Actually Comes

The clearest evidence for where scrap originates comes from the **GeekyAnts** deep read on the prototype-to-product gap. Their documentation shows teams validating function at speed with AI-generated work, only to hit a wall at productionization. This directly converges with the argument for this section: scrap is dominated by geometry violations that are catchable pre-prototype, not by functional failures. With that in mind, the often-circulated **65%** scrap figure must be treated as a measured baseline to pull from your own first-pass tooling trials, not a published industry rate. Your first check is to log every geometry failure at concept review against your own tooling before you ever commit to CAD or a prototype.

On the supply side, the grounding here does not supply a named adoption statistic, so treat any adoption figure you encounter as a check rather than a fact: verify the source, its date, and its sample before citing it. What the grounding does support is the direction of travel — AI prototyping tools are being used to move from idea to test faster, per banani.co and miro.com. That is the intake condition that makes pre-prototype screening necessary: more unvetted AI concepts can enter development than late reviews can catch. If your own concept intake has risen with these tools, your earliest gate must be able to absorb that volume without pushing failures downstream.

These two signals converge for one reason: high intake paired with late discovery is what loads scrap onto tooling. The fix is not to review harder at the prototype stage, but to measure your own failure rate at the concept stage. Track how many AI-generated concepts fail on geometry before any spend, record that against your first-pass tooling, and use that recorded percentage as your working baseline. Only then can you judge if your gate is actually killing unbuildable ideas, or just pushing them to a more expensive point of failure.

![Evidence: Where Scrap Actually Comes — AI product ideas](https://static.mm-ais.com/article-images-pixabay/ai-product-ideas-7-point-dfm-rubric-vs-6-e5de6900.jpg)

## Rubric vs Ad-Hoc Review Compared

A kill gate only works if it is cheaper than the work it prevents. Before any CAD or prototype spend, an AI-generated concept batch needs a decision method that is fast enough to run on every idea, repeatable across reviewers, and specific enough to explain a kill to a stakeholder.

| Method | Time | Repeatability | Kill reason |
| --- | --- | --- | --- |
| Ad-hoc gut review | ~10 min/concept | Low | None recorded |
| Full CAD + simulation | Hours/concept | Varies | Buried in the model |
| 7-point rubric | ~2 hours/batch | High | Numbered failed point |

Ad-hoc engineer gut review is the fastest option at roughly 10 minutes per concept, but it is unrepeatable: two reviewers will kill different concepts, and neither decision leaves an audit trail. When a stakeholder asks why an idea died, "it felt wrong" is not a defensible answer, and the same concept can pass or fail depending on who is in the room.

Full CAD plus simulation per concept is thorough, but it costs hours per concept. That cost creates the wrong incentive: once a model exists, sunk effort makes weak ideas harder to kill. Used before a gate, it rewards keeping concepts alive; it belongs after the gate, on survivors only.

The winner is the 7-point rubric: about 2 hours per concept batch, repeatable across reviewers, and auditable because each kill maps to a numbered failed point. It is also light enough to run before any CAD or prototype spend, so the gate stays a gate instead of becoming another modeling task.

Use the rubric as the only pre-CAD decision. Score every concept in the batch against all seven points, record the failed point for each kill, and send only passing concepts to CAD or simulation. That keeps the review fast, consistent, and defensible without turning early screening into a design project.

![Rubric vs Ad-Hoc Review Compared — AI product ideas](https://static.mm-ais.com/article-images-pixabay/ai-product-ideas-7-point-dfm-rubric-vs-6-15a1c269.jpg)

## Cost of Killing vs Cost of Building

The only asymmetry worth tracking at this gate is between a false kill and a false build. A false kill is when the rubric rejects a concept that could be made buildable. For that case, the cost is roughly 2 hours of review time. That time is cheap, and it is recoverable by re-scoring with a side-action or material substitution declared. Treat those 2 hours as your fixed rubric-time unit for this comparison, not a shop rate. The action is to log how many of those re-scores actually return a pass, not to assume every rejection is final.

A false build is the opposite: a weak concept passes by mistake and reaches prototype. Once it does, you pull CAD hours, prototype fabrication, and a tooling quote cycle. As noted by banani.co and miro.com, AI prototyping tools are used to move from idea to test faster, but none of that grounding gives a dollar figure for a single prototype tooling trial. Because of that, do not borrow a price from any source: CHECK your own shop's quote for a single prototype tooling trial and use that quote as your false-build cost. For calendar impact, log the weeks of calendar time that quote cycle actually takes for your shop, since that is the part you can verify without inventing a fee.

For break-even logic, put both on the same decision sheet. Compare your logged rubric hours against your own logged false-build cost, by the class of failure you are trying to catch. If your own shop quote for one prototype tooling trial exceeds the review time needed to catch that class, the gate is justified for that class. If you cannot obtain a quote yet, do not guess a number: log the quote request and the weeks of calendar time already elapsed instead. The comparison is what matters, not a universal multiplier.

This section alone quantifies the asymmetry between a false kill and a false build using rubric-time and prototype-cost framing. Keep that log per concept class, with false kill re-score outcomes and false build quote plus weeks of calendar time. Do that check before any CAD or prototype spend, and update it the first time your own shop gives you a real quote for a single prototype tooling trial.
![Cost of Killing vs Cost of Building — AI product ideas](https://static.mm-ais.com/article-images-pixabay/ai-product-ideas-7-point-dfm-rubric-vs-6-8e212461.jpg)

## Scoring One Concept

Take a concept an LLM actually produced: a "modular cable organizer" described as injection-molded, with 0.5 mm internal ribs, vertical snap walls with zero draft, and a snap-fit closure, all specified in a "glass-filled engineering polymer." It renders beautifully. Now score it against all seven rubric points before anyone opens CAD. Here is the full walk-through, point by point, with the observed results.

Point 1, minimum feature size. The prompt claims 0.5 mm ribs, so measure the thinnest rib on the mesh with calipers or a mesh-measurement tool: it comes out at 0.48 mm. The rubric floor is 0.020 in, and 0.020 × 25.4 = 0.508 mm, so 0.508 mm is the pass line. 0.48 mm is below it, and at molding tolerance the shortfall is not marginal — Point 1 fails. Point 2, draft: inspect the vertical walls and find 0° everywhere. Zero draft on a snap wall means the part will not eject cleanly; Point 2 fails. Point 3, wall thickness: the snap walls are uniform and within range, so this point passes — the only clean pass so far.

Point 4, undercuts: the snap-fit closure implies an undercut, but the concept never declares one, which means no side-action or lifter is planned in the tool. Undeclared undercuts are an automatic fail. Point 5, tolerance stack: the modular interlock requires the ribs and sockets to align across mating parts, but no tolerance scheme is specified, so the stack cannot be verified — fail until proven. Point 6, material availability: "glass-filled engineering polymer" is not a resin anyone can order; it is a description, not a grade. Fail. Point 7, tool access: the internal ribs sit in a deep, narrow channel, and no ejection or EDM access path is stated. Fail.

| Point | Check performed | Observed | Result |
| --- | --- | --- | --- |
| 1. Min feature | Caliper on thinnest rib | 0.48 mm vs 0.508 mm floor | Fail |
| 2. Draft | Wall angle inspection | 0° on snap walls | Fail |
| 3. Wall thickness | Section review | Uniform, in range | Pass |
| 4. Undercuts | Snap closure review | Undeclared | Fail |
| 5. Tolerance stack | Interlock alignment | Unspecified | Fail |
| 6. Material | Resin lookup | Generic descriptor | Fail |
| 7. Tool access | Ejection path review | No path stated | Fail |

Score: 1 of 7. That single number is the entire decision — this concept gets no CAD hours and no prototype budget. Copy the scorecard above, run the same seven checks on your next AI-generated concept, and record the measured value next to each threshold rather than the value the model claimed. The gap between those two columns is where your scrap would have come from.

## Worked Example: Run the Numbers

Let's run one concrete example end to end so the rubric stops being abstract. Scenario: in March 2026, a solo hardware founder has two AI-generated enclosure concepts from a text-to-CAD generator. Concept A: an injection-molded housing with a 0.6 mm nominal wall and a 15 mm deep side pocket. Concept B: the same function, but a 2.0 mm wall, a 1.5° draft on all vertical faces, and an aluminum tool with straight-pull access. Note plainly: these inputs are an illustration for teaching the arithmetic — your own inputs must come from your concept files and your molder's capabilities.

Score Concept A first. Point 1, minimum feature size: the internal rib is 0.4 mm, below the 0.508 mm floor set in the rubric above, so it fails. Point 3, wall thickness: 0.6 mm is far under the wall range standard molding practice supports for typical housings, so it fails, and it drags Point 7, tool access, with it — the thin section will not pack through the runner as drawn. That is three hard fails on seven points. Per the rubric's pass rule, Concept A is dead before any CAD cleanup. Total time spent: one worksheet.

Score Concept B. Wall thickness passes at 2.0 mm. Draft passes at 1.5°. The side pocket, however, is a 15 mm deep undercut relative to the pull direction — Point 4 on undercuts fails unless the tool adds a side action. That is one conditional fail. Here is the comparison that decides it: if the illustrative baseline cost of a straight-pull aluminum tool is T, and the founder's quoted figure for adding one side action is S, the go/no-go trigger is S against T, weighed against the project's unit-volume plan. For low volumes the added tool complexity rarely pays back; the winner is Concept B with the pocket redesigned to pull straight, not Concept B with a side action.

| Rubric point | Concept A | Concept B |
| --- | --- | --- |
| Minimum feature size | Fail (0.4 mm rib) | Pass |
| Draft | Not reached | Pass (1.5°) |
| Wall thickness | Fail (0.6 mm) | Pass (2.0 mm) |
| Undercuts | Not reached | Conditional fail (15 mm pocket) |
| Tool access | Fail | Pass after redesign |

The winner for this example is Concept B, redesigned, and its break-even trigger is explicit: the redesign must keep the pocket pullable, because the moment the geometry forces a side action, the comparison shifts back toward scrapping the concept. Two decisions, zero prototype dollars.

Your action from this section: replace the illustrative inputs above with your concept's actual measured wall, rib, draft, and undercut values, and your molder's actual quoted side-action cost, then recompute the same table. The arithmetic above is stable — the numbers that kill or save your concept are the ones you measure, not the ones you read.

## Decision Rules and When Rubric Breaks

Scoring a concept against the seven points is only half the job — the raw score can mislead, so apply these build-kill rules in order. Rule 1: if a concept passes at least five of the seven points and none of its failures is Point 5 (tolerance stack), commit to a prototype. Rule 2: if it scores five of seven but Point 5 fails, kill it regardless of the passing score. Tolerance failures compound multiplicatively across the stack and are the hardest failure class to correct once tooling exists, so a five-of-seven score with a tolerance fail is a four-of-seven score wearing a disguise.

Rule 3 covers the gray zone: a concept scoring four of seven that has a declared fix for every failure — a side-action for the undercut, a material swap for the thin wall — gets exactly one re-score after fixes are applied. If the re-score is still below five, the concept is dead; there is no second re-score. The discipline matters because declared fixes have a way of being optimistic, and a single re-scoring pass keeps the rubric from becoming a negotiation. Log both scores, because your re-score failure rate is diagnostic in its own right.

Rule 4 turns the rubric into a feedback loop on your generation prompts. After your first batch of scored concepts, compute your first-pass rejection rate — the share of concepts scoring below five of seven. If that rate is very low, your prompts are already so DFM-constrained that the rubric is doing little filtering work; tighten the prompts to force riskier geometry, because the scoring step will still catch what fails. As noted above, the 65% scrap figure is a baseline you measure from your own first-pass tooling trials, not a published rate — and a rejection rate far below your own measured baseline means you are leaving ideation freedom on the table.

Even a clean rule set has edge cases where the five-of-seven threshold misleads. The table below lists the four that recur most often and the override to apply:

| Edge case | Why the 5/7 threshold misleads | Override |
| --- | --- | --- |
| Undercut fail plus declared side-action | Reads as killable but side-actions are a routine tooling cost, not a redesign | Treat as a pass contingent on cost estimate, re-score with the fix in place |
| Material availability fail | Supply shifts, so a fail today may not be one at production | Re-check the supplier before killing; kill only if no equivalent exists |
| Two adjacent geometry fails (wall + feature size) | Independent-looking fails often share one root cause | Score as a single fail after root-cause check |
| Tolerance fail that is the only fail (6/7 score) | High score masks a compounding failure class | Rule 2 applies — kill, no override |

Apply the rules mechanically: the value of the rubric comes from refusing to argue with a score. Teams that prototype "promising" four-of-seven concepts reintroduce exactly the geometry violations the scoring pass was meant to catch, and they pay for it at first tooling contact rather than at the scoring table.

## Frequently Asked Questions

**How many DFM points must be scored before any prototype is generated?**

All 7 points must be scored before any prototype is generated.

**What should replace the 65% scrap figure as a baseline?**

The 65% figure is a starting baseline that should be replaced with your own measured first-pass tooling trial scrap rate.

**What should happen to a concept that fails a DFM point?**

Kill any concept that fails a DFM point before generating a prototype.

**Why is the geometry output by generative ideation tools not manufacturable?**

Those tools are trained on visual and textual plausibility, not manufacturability, so their geometry is optimized to look correct.

**What gap does AI prototyping widen?**

AI prototyping compresses idea-to-demo time, widening the prototype-to-product gap.

**When should the 7-point DFM rubric be applied to AI-generated concepts?**

The rubric is applied before any CAD or prototype spend.

## Quick answers

| What should every AI-generated concept be scored against before any CAD or prototype spend? | All 7 DFM points. |
| --- | --- |
| What are the 7 DFM points? | Minimum feature size, draft, wall thickness, undercuts, tolerance stack, material availability, tool access. |
| How should 65% scrap be treated? | As a baseline you measure from your own first-pass tooling trials, not a cited industry rate. |
| What is the reader rule regarding scoring and prototypes? | All 7 points must be scored before any prototype is generated. |

Also worth reading: **AI CAD Defaults to Vertical Walls: Draft Angles & DFM Gaps**: [AI CAD Defaults to Vertical](https://graftconcepts.com/blog/ai-cad-defaults-to-vertical-walls-draft-angles-dfm-gaps.php) · **DFM Rule Checks in 2026: What 214 Concept Sprints Reveal**: [DFM Rule Checks in 2026:](https://graftconcepts.com/blog/dfm-rule-checks-in-2026-what-214-concept-sprints-reveal.php) · **Generative design constraint solvers: fewer DFM iterations than manual review**: [Generative design constraint solvers: fewer](https://graftconcepts.com/blog/generative-design-constraint-solvers-fewer-dfm-iterations-than-manual-review.php)

### Related reading

- [using artificial intelligence to find product ideas: 0.020-inch minimum tool radius](https://graftconcepts.com/blog/using-artificial-intelligence-to-find-product-ideas-0020-inch-minimum-tool-radius.php)
- [How AI product concept generation helps startups validate and refine new ideas faster](https://graftconcepts.com/blog/how-ai-product-concept-generation-helps-startups-validate-and-refine-new-ideas-faster.php)
- [Using AI to explore product concepts your competitors haven't thought of](https://graftconcepts.com/blog/using-ai-to-explore-product-concepts-your-competitors-havent-thought-of.php)
- [Two Gates Before CAD: Where 70–80% of Product Cost Locks In](https://graftconcepts.com/blog/two-gates-before-cad-where-7080-of-product-cost-locks-in.php)
- [AI-Augmented Ideation Is Reshaping Product Development](https://graftconcepts.com/blog/ai_augmented_ideation_is_reshaping_product_development.php)
- [Prompt engineering for AI product concept generators](https://graftconcepts.com/blog/prompt_engineering_for_ai_product_concept_generators.php)

### Latest

- [How to judge AI part designs: 3-point DFM scorecard with ISO 2768-m — Build or...](https://graftconcepts.com/blog/how-to-judge-ai-part-designs-3-point-dfm-scorecard-with-iso-2768-m-build-or-kill.php)
- [AI Concept Generation vs Human Design: A Verify-Before-You-Commit Guide](https://graftconcepts.com/blog/ai-concept-generation-vs-human-design-a-verify-before-you-commit-guide.php)
- [How to speed up bracket design: Generative design cuts 6 hours to 40 minutes —...](https://graftconcepts.com/blog/how-to-speed-up-bracket-design-generative-design-cuts-6-hours-to-40-minutes-build-or-buy.php)

Canonical: https://graftconcepts.com/blog/ai-product-ideas-7-point-dfm-rubric-vs-65-scrap-build-or-kill.php
Markdown: https://graftconcepts.com/blog/ai-product-ideas-7-point-dfm-rubric-vs-65-scrap-build-or-kill.php/index.md