Direct answer: tokens measure compute consumption; they do not measure engineering acceptance. For an equipment maker, AI becomes economically accountable when a frozen design obligation is delivered as a native artifact, verified through generator-independent post-execution evidence, and accepted by a qualified engineer—or explicitly refused when the evidence is insufficient. ROI is proven only when total cost per accepted result improves against a frozen baseline.
Claim boundary: token pricing is a rational way to meter computation. This article does not argue that every AI project loses money, that every AI provider is profitable or that MST guarantees customer ROI. It defines a stricter way to measure industrial AI.
Both ledgers matter. They are not interchangeable.
Tokens, seats, reserved capacity, inference time and platform usage.
Answers: What did computation cost?
Native artifact, review time, correction cost, rework, refusal quality and verified throughput.
Answers: Did the workflow produce usable value?
The provider’s meter measures computation. The factory’s ledger measures accepted engineering work. A credible ROI claim must reconcile both.
Adoption, supplier revenue and customer value are different claims
General-purpose AI services are naturally metered by inputs, outputs, capacity or seats. OpenAI’s token documentation explains that API usage is priced by input, output and cached tokens, while Scale Tier allocates token throughput. Those sources establish the computation meter. They do not establish supplier profitability, customer ROI or the value of a specific workflow.
The economic asymmetry is narrower and more important: a supplier can recognize consumption revenue before the customer has demonstrated an accepted operating result. The customer carries integration, review, correction, workflow redesign and organizational adoption costs that are not visible in a token invoice.
The public evidence supports a value gap—not a universal loss claim.
Organizational adoption and AI-company revenue grew, while measured productivity effects remained task-dependent. Narrow, observable work is easier to evaluate than ambiguous deep-reasoning work.
About six percent of respondents met McKinsey’s high-performer definition: at least five percent EBIT impact attributed to AI plus reported significant value. The survey results do not show that the remaining organizations lost money because of AI.
Nearly three quarters of respondents said their most advanced initiative was meeting or exceeding ROI expectations, while most experiments were not expected to scale quickly. Different samples and definitions produce different answers.
The conclusion is not that AI has no value. It is that adoption, usage, supplier revenue, productivity and attributable P&L impact are separate measurements. A credible industrial-AI program declares which one it is measuring before deployment.
The unit of industrial-AI value is an accepted engineering obligation
An accepted engineering obligation is a bounded requirement whose inputs, native result, verification method, failure behavior and human authority are declared before execution. It is not a prompt, a session or a generated-file count.
- Frozen inputsIdentified P&ID, BOM, approved part revisions, Rulepack, geometry and spatial boundary.
- Declared obligationThe exact connection, placement, routing, assembly or verification task.
- Native executionOperation of the real engineering toolchain rather than a rendered imitation.
- Editable artifactA result in the customer’s working engineering format, using approved existing models.
- Post-execution read-backRealized facts extracted after execution instead of accepted from the generator’s own report.
- Bounded verdict
AUTO_PASS,PASS_WITH_ENGINEER_REVIEW,FAIL_CLOSED,UNSUPPORTEDorCANNOT_VERIFY. - Human release authorityA qualified engineer retains responsibility for formal design and production release.
Many equipment-design obligations can be converted into typed, tolerance-aware checks. Others remain engineering judgments and must stay under qualified human authority. A successful SolidWorks rebuild or a set of mates is necessary evidence for some obligations; it is not proof of process correctness, manufacturability, EHS, safety or regulatory compliance.
A real deployment is evidence of field use—not automatic proof of ROI.
MST has deployed this bounded engineering workflow with a semiconductor process-equipment manufacturer in China. The customer remains anonymous; no name, logo, endorsement or confidential engineering data is published. Deployment establishes field use; it does not by itself prove ROI, enterprise-scale rollout or customer endorsement.
Quantified value requires a customer-approved baseline, observation window, sample size, acceptance definition and the full cost of review, correction and exception handling.
From P&ID evidence to a native SolidWorks result
- P&ID PDF
Preserve page, region, tag, line and ambiguity evidence.
- Directed topology
Compile source observations into conditional engineering obligations.
- PDM + B-Rep
Resolve customer-authorized part identity, revision and physical interface truth.
- 3D physical layout
Solve discrete parts, continuous XYZ pose, layer changes, keep-outs and orthogonal routing.
- Assembly IR
Compile components, transforms, mates, routes and proof obligations.
- Native SolidWorks
Create a native
.sldasmthat remains editable and uses approved existing models. - Reverse proof
Reopen the artifact and map realized facts back to each in-scope obligation.
- Human release
Present evidence, exceptions and verdicts to the qualified engineer.
MST’s current public scope is configured modular semiconductor process-equipment assembly under controlled libraries, rules and boundaries. It is not arbitrary whole-machine autonomy, universal PDF/CAD/PDM support or production release without engineer approval.
Evidence and refusal are part of the delivered result
A factory cannot accept “the answer looked plausible.” The reviewer needs approved part identity and revision, BOM conservation, target and realized mates, rebuild state, interface facts, spatial constraints and P&ID-to-physical-path coverage. When required evidence is missing, a fail-closed refusal prevents an unsupported artifact from being represented as complete.
Generator-independent post-execution read-back means that realized SolidWorks facts are extracted after execution rather than accepted from the generator’s own report. It does not mean third-party certification. NIST’s AI Risk Management Framework supports testing, documentation of limitations, validity/reliability evaluation and safe failure beyond knowledge limits; NIST does not certify MST or prescribe this architecture.
Measure total cost per accepted result
A customer should freeze the baseline, result definition and valuation rules before the pilot. Engineering hours cannot be counted simultaneously as labor savings, added capacity and lead-time value unless a preregistered rule prevents double counting.
| Measure | Frozen baseline | Pilot result | Customer-set gate | Evidence source |
|---|---|---|---|---|
| Accepted obligations per engineer-week | Manual workflow | Observed, not estimated | Declared before run | Task and approval ledger |
| Median / P90 cycle time | Same task family | Full elapsed time | No post-hoc threshold | Timestamped run record |
| Review + correction minutes | Same definition of done | Include exceptions | Customer-approved | Reviewer record |
| First-rebuild and downstream rework | Comparable modules | Report both | Quality guardrail | Native artifact + change record |
| False pass / false refusal / correct refusal | Frozen fixture | Include sample size | Risk-based limit | Golden verdict set |
| P&ID-to-physical coverage | Declared obligations | Pass, review or refusal | No silent omission | Evidence graph |
| Cost per accepted result | Fully loaded manual cost | Fully loaded AI cost | Improves vs baseline | Finance-approved model |
Annual economic benefitverified labor savings + verified rework or scrap avoidance + non-overlapping capacity or lead-time contribution + expected-loss reduction
Annual AI total costannualized integration + platform or license + inference + human review + correction + exception handling + maintenance
Net engineering valueannual economic benefit − annual AI total cost
ROI(annual economic benefit − annual AI total cost) / annual AI total cost
Cost per accepted resultannual AI total cost / qualified accepted results
Five readers, five decisions
PR / ANALYSTSClaim native artifacts, evidence and bounded refusal.Inspect the category definition →
STRATEGIC INVESTORSDiligence the compiler, verifier and deployment economics.Review the investment wedge →
ENGINEERSTest the obligation from source edge to realized face.Inspect the proof chain →
ENGINEERING LEADERSTrack throughput, review, rework and cost per result.Open the measurement guide →
Commercial acceptance should follow the engineering contract
An industrial-AI pilot can combine platform fees, implementation fees, compute and milestone payments. The commercial model does not need to be purely outcome-priced. It does need staged acceptance gates: formalized intent, qualified part/interface coverage, solved physical layout, native artifact, post-execution evidence and human acceptance.
Stopping rules matter. If authoritative part data is unavailable, the solver cannot satisfy the frozen fixture, exception-handling cost overwhelms the benefit or the accepted-result cost fails to improve, the scope should stop or narrow. More token consumption is not a remedy for a broken engineering contract.
The investment thesis lives above the model layer
Some inference capacity is becoming easier to obtain, but models still differ in capability, latency, security and migration cost. MST’s potential durable layer is elsewhere: customer-specific data contracts, PDM identity and revision handling, B-Rep interface truth, deterministic compilers, domain Rulepacks, native-tool adapters, acceptance fixtures, evidence ledgers and reviewed failure history.
Those assets become a defensible business only if repeated deployments reduce configuration and exception-handling cost while preserving customer isolation. Investors should examine deployment time, reusable-kernel versus customer-specific work, variable cost per accepted result, compute share of revenue, review burden, gross margin, renewal, expansion and customer concentration. A technical barrier is a potential moat; repeatable economics prove that it has become one.
Inspect the result contract from three directions.
PRODUCT CONTRACTP&ID to native SolidWorksSee the bounded workflow →
FAILURE CONTRACTVariation, refusal and benchmark gatesInspect the evaluation method →
DEPLOYMENT CONTRACTCustomer-local data, IP and executionReview the boundary →
INVESTMENT WEDGEEvidence levels and diligence questionsOpen the investor page →
QUALIFIED REVIEWRequest the technical diligence packageContact the engineering team →
What this article does not prove
It does not prove that every AI customer loses money, that every model provider wins, that result-based pricing always produces better economics, or that MST has already demonstrated universal ROI. It also does not turn native CAD execution into proof of safety, manufacturability or regulatory compliance.
The claim is more precise: consumption-led AI becomes economically fragile when usage is mistaken for value. Industrial AI becomes accountable when accepted results, total cost, evidence and refusal are measured together.
References
- OpenAI — What are tokens and how to count them? Accessed 2026-08-08.
- OpenAI — Scale Tier for API customers. Accessed 2026-08-08.
- Stanford HAI — 2026 AI Index, Economy. Accessed 2026-08-08.
- McKinsey — The state of AI: Global Survey 2025. Accessed 2026-08-08.
- Deloitte — State of Generative AI in the Enterprise, Q4 survey. Accessed 2026-08-08.
- NIST AI RMF Core — Govern, Map, Measure and Manage. Accessed 2026-08-08.
Limit: this article defines an economic and engineering acceptance thesis. It does not claim that all AI vendors are profitable, that all generic AI projects lose money, or that MST guarantees universal customer ROI. MST performance and customer value must be demonstrated on a frozen scope with customer-reviewed evidence.
MST Engineering AI. “AI Should Ship an Engineering Result, Not a Token Bill.” MST Engineering AI. Updated 2026-08-08. https://mst-us.ai/ai-should-ship-engineering-result-not-token-bill/
CUSTOMERS · PARTNERS · INVESTORS
Talk to the engineering team.
Bring a non-confidential P&ID question, pilot boundary or technical diligence request. Protected engineering material is exchanged only after identity, NDA and scope review.
Operated by MOORE SOLUTION TECHNOLOGY PTE. LTD. · Singapore UEN 202315239E