AI Machining Quotation: From Drawing to Quote, and Acceptance Criteria

“RFQs keep coming in, but only two or three people in the company can look at a drawing and put a price on it. Our replies are late, and we lose the job.” This is a concern we hear often from presidents and sales managers at Japanese-affiliated contract machining companies in Thailand. That is when AI machining quotation comes onto the table. Here is the conclusion up front. Introducing AI into machining quotation does not mean letting AI decide the price. It means breaking the estimator’s judgment into four steps, “reading”, “similarity search”, “cost calculation” and “approval”, and deciding where the boundary lies between the steps you hand to AI and the steps people remain accountable for.

What decides success is not the accuracy of the AI itself, but whether you have settled the following three points before you place the order.

• Which drawings may be passed through AI quoting (scope)
• When the AI misreads a drawing, who stops it, and where (approval gate)
• What you measure accuracy with, and what counts as a pass (acceptance criteria)

Note that among the amounts, hours, counts and ratios in this article, all those relating to “Model Factory Q”, described later, are original estimates and assumptions made for this article. They are neither industry averages nor survey figures. Please read them as a “calculation template” into which you substitute your own measured values.

What Is Happening with AI Machining Quotation in 2026

Money and products are starting to converge on quoting

From summer into autumn 2026, there was a run of major moves around machining quotation.

On September 15, 2026, CADDi announced that it had raised JPY 17.7 billion in a Series D round, bringing its valuation to JPY 182 billion. It listed the uses of the funds as developing its proprietary AI technology, expanding its platform and strengthening its North American business, among others. Earlier, on August 6, it announced the next-generation “CADDi Manufacturing AI Data Platform”, and among its business-specific products it lined up “CADDi Quote”, the quotation platform it launched in September 2024, and “CADDi Cost Review”, a cost assessment product scheduled for release within the year. According to the company’s announcement, more than half of Nikkei 225 companies have adopted it, and it is used in 22 countries.

In the UK, it was reported on September 9, 2026 that CloudNC, known for its AI CAM product “CAM Assist”, had raised USD 20 million and plans to launch “Quote Agent”, a new product that applies AI to quoting and estimating for machining, within 2026. According to the reports, CEO Theo Saville said, “Quoting is one of the biggest hidden bottlenecks in precision machining.” In other words, quoting is seen as the “invisible bottleneck” of the machining floor. However, at the time of writing, Quote Agent has not yet been released, and neither its accuracy nor its price has been disclosed.

US quoting software company Paperless Parts announced in April 2026 that it had been granted its fifth US patent. The patent covers AI that reads a drawing at the moment it is uploaded and converts it into structured information usable for quoting and procurement. The company lists the pillars of the patent as interpretation of geometric dimensioning and tolerancing (GD&T), material callouts, surface finish and thread specifications; extraction of requirements from PDF drawings; mapping of notes to 3D geometry; and robustness to rotated or degraded drawings. Before that, in September 2025, the company announced “Requirements Review”, which automatically highlights GD&T and critical notes at the quoting stage and flags high-risk conditions based on rules, with general availability planned for October 2025 after a private beta.

Buyer-side tools and seller-side tools are different things

One distinction needs to be made here. Products called “AI quoting” actually include two types with different positions.

The first is procurement tools for buyers (the ordering side). CADDi Quote is a cloud service for the buyer’s quoting work, featuring bulk sending of RFQs, automatic generation of comparison tables, and suggestions of similar-part candidates based on past quotation and procurement data. The instant quotes offered by manufacturing platforms such as Xometry and Protolabs in the US are likewise mechanisms in which the buyer enters a drawing or 3D data and gets back a price and lead time. In March 2026, Xometry announced that it had quadrupled the training data for its machining lead-time prediction model and would use a model that adjusts the price for each quote based on geometric features, the customer’s purchase history and other factors. In February 2026, Protolabs announced “ProDesk”, which combines AI quoting with manufacturability analysis, and according to the company’s own explanation, it expects about 50,000 customers to use it in 2026. MISUMI’s meviy began offering, in March 2026, a function that automatically links dimensional tolerance information from iCAD’s 3D CAD “iCAD SX”. It covers milled (prismatic) parts and sheet metal, and the company says it reduced quotation setup effort by about 25% compared with the conventional manual re-entry. This is a figure announced by MISUMI, and the details of the comparison conditions have not been disclosed.

The second is systems for sellers (contract manufacturers) to produce their own quotes. Paperless Parts and CloudNC’s Quote Agent fall into this group. Most readers of this article are likely in this position.

The two serve different purposes, but they are not unrelated. As procurement tools spread on the customer side, contract manufacturers come under pressure to return quotes “faster and in a consistent format”. Once your quote sits in a customer’s comparison table, companies that reply slowly or submit quotes with missing items are easily dropped from consideration. AI quoting on the contract manufacturer’s side is also a way to prepare for this pressure.

Management feels the effect, but it has not yet reached the quoting floor

In an online survey CADDi conducted from August 4 to 10, 2026 among 484 executives of manufacturing companies (announced September 10), 57.8% of executives said they felt a management-level impact from investing in and using AI. On the other hand, the company says only 17.2% of companies saw the effect of AI extend to specialist tasks such as design, quoting and procurement, or to processes on the shop floor. This is a survey by a company that itself sells AI quoting products, and its respondents are Japanese executives. Still, the picture of “senior management feels results, but AI has not reached specialist work such as quoting” probably matches what Japanese-affiliated factories in Thailand are experiencing too.

There is an investment flow on the Thai side as well. According to the Thailand Board of Investment (BOI), in the first half of 2026 the “Smart and Sustainable Industry” measure, which covers machinery upgrades, adoption of digital technology, and automation and robot integration, received 132 applications worth about 17.2 billion baht (these are application-based figures, not actual investment amounts). As for current demand, according to Trading Economics, reporting an announcement by the Federation of Thai Industries (FTI), automobile production in August 2026 was 124,646 units, up 10.93% year on year, but vehicle exports in August fell 2.04% year on year, and the FTI forecasts full-year 2026 production to decline by 3.33%. The harder it is to read the peaks and troughs of orders, the more valuable a system becomes that lets a small team of estimators handle many requests and reliably win the jobs that can be won.

What AI Machining Quotation Means: Splitting a Quote into Four Steps

If you break down the work an experienced estimator does all at once in their head, it falls roughly into the following four steps.

1. Reading: Pick out from the drawing the material, quantity, dimensional tolerances, geometric tolerances, surface roughness, and notes such as heat treatment and plating
2. Similarity search: Find whether you have made a similar part before, and how much it actually cost at that time
3. Cost calculation: Build up material cost, setup and machining time, outsourced processes, inspection, and packing and transport
4. Approval: Decide the final price and the promised lead time, looking at profit, delivery and risk

Discussions of AI quoting easily become confused because people lump these four steps together and ask, “Can AI automate quoting?” Looked at step by step, the parts where AI is strong and the parts people should remain accountable for separate clearly. Reading and similarity search are steps where AI is easy to put to use (on the premise that people check the reading results). Cost calculation is suitable or unsuitable depending on the element. Approval is people’s work.

AI quoting is also easily confused with surrounding systems. Quoting for configurable products, where the price is determined by selecting options, is handled by a different system called CPQ (for details, see our article on CPQ adoption in manufacturing). Generating machining toolpaths with AI is the domain of AI CAM, which has a different purpose from quoting. That said, a link in which machining times calculated by AI CAM are used in the quote’s cost calculation is possible.

Step 1: Reading the Drawing (2D Drawings, 3D Data, GD&T and Critical Notes)

Reading 2D drawings (PDF)

On the floors of contract machining companies in Thailand, many drawings still arrive as 2D PDFs. What needs to be read is the title block (drawing number, revision, material, quantity), dimensions and dimensional tolerances, geometric tolerances, surface roughness, and notes such as heat treatment, plating, grinding and deburring. The symbols and interpretation of geometric tolerances are defined in ISO 1101 for ISO-based drawings and in ASME Y14.5 for drawings from US customers. Japanese customers use JIS or ISO-based systems, and US customers use ASME-based systems, so drawing systems can be mixed within a single factory. This needs to be kept in mind at the stage of defining the reading specification.

3D data (STEP AP242) and PMI

When you receive 3D data, ISO 10303-242 (STEP AP242), whose fourth edition was published in August 2025, includes in its scope dimensional and geometric tolerance data and product manufacturing information (PMI) from the design and manufacturing planning stages. With PMI-bearing data, you can read geometric features and tolerances without going through a drawing. In practice, however, many STEP files are exported with geometry only and no PMI. In that case, tolerance information does not reach the AI, and you need to read the 2D drawing separately. The fact that meviy released a function that automatically links tolerance information from iCAD SX can be seen as reflecting the view that the reliable way to pass tolerances is from the 3D model side.

Limits of reading drawings with general-purpose generative AI

You can also have general-purpose generative AI such as ChatGPT, Claude or Gemini read a drawing PDF. However, each vendor’s official documentation states clear constraints.

Anthropic’s official documentation states, as limitations of image understanding, that errors can occur with low-quality images, rotated images, and small images under 200 pixels, and that coordinate and position outputs are approximate, and it asks that the model not be used without human oversight for tasks that require perfect precision. For PDFs, it gives a maximum request size of 32MB and a maximum of 600 pages per request (100 pages for requests with a context of less than 1 million tokens), says password-protected or encrypted PDFs cannot be handled, and says each page uses roughly 1,500 to 3,000 tokens. Images have a resolution limit, and large images are scaled down before processing, so you should assume that fine tolerance symbols on an A3 drawing may become hard to read.

Google’s Gemini API accepts PDFs of up to 50MB or 1,000 pages, treats each page as equivalent to 258 tokens, and scales large pages down to a maximum of 3072×3072. OpenAI’s API limits PDF input to under 50MB per file and 50MB in total per request, and provides a parameter to adjust the level of detail of image processing.

In other words, if you read drawings with general-purpose AI, you need to work on the premise that the drawing’s resolution, page layout and any rotation are normalized in preprocessing, and that people always check the reading results.

The purpose of reading is to catch “dangerous requirements” before calculating a price

The most important thing in reading is, in fact, not producing a price. Missing a tolerance or a note is not a matter of the quoted amount being slightly off; it leads directly to post-order losses and defects. For example, if you miss a heat treatment instruction, the entire outsourcing cost is left out; if you miss a tight geometric tolerance, both the machining method and the inspection method change.

It is symbolic of this point that Paperless Parts’ Requirements Review puts “highlighting GD&T and critical notes” and “flagging high-risk conditions”, rather than price calculation, at the forefront, and is designed on the premise of human review, with assignment of reviewers and audit history. The acceptance criteria for reading AI should put “not missing critical notes” ahead of price error.

Step 2: Similar-Drawing Search and Actual Costs

“How much did we quote for a similar part before?” is not enough

The first thing an estimator does on seeing a new drawing is try to remember “whether we have made something similar before”. AI-based similar-drawing search separates this task from human memory. Mechanisms that retrieve past drawings by closeness of geometry, dimensions, material and notes are covered in detail in our article on drawing search AI.

What matters in quoting is what is linked to the past jobs retrieved. If only past quoted “amounts” are linked, you cannot tell whether those amounts were profitable. They include amounts for jobs won with discounts, jobs that made a loss and jobs that were lost. What determines accuracy is whether past jobs are linked to actual costs, that is, the setup and machining times actually spent, outsourcing costs, and rework due to defects.

Data preparation is the first big hurdle

At Japanese-affiliated factories in Thailand, information is often scattered: quotes in Excel, drawings on a file server, actual labor hours in the production management system, and outsourcing costs in the accounting system. To make similarity search useful for quoting, you need to link these together using the drawing number and revision as keys. If drawing revisions are misaligned, you end up quoting a new revision with the actual costs of an old revision, which causes errors. For how to think about drawing revision control, our article on design change management AI is also a useful reference.

This preparation is unglamorous, but it is an important step that accounts for a quarter of the initial investment in the model estimate described later.

Step 3: Cost Calculation (Elements Easy and Hard to Hand to AI)

Cost calculation is a build-up of material cost, setup time, machining time, outsourced processes, inspection, and packing and transport. Among these, the elements that are easy to hand to AI and those that are hard to hand over separate clearly.

Some academic research shows that, under limited conditions, machine learning can estimate manufacturing costs with a certain degree of accuracy. A study posted to arXiv in August 2025 extracted about 200 features from 13,684 2D drawings (24 product groups) of automotive suspension and steering parts and predicted manufacturing costs with gradient-boosted trees, reporting a mean absolute percentage error of almost 10% across product groups. However, this is a preprint that has not yet been peer reviewed, and the result was obtained on a specific dataset (automotive suspension and steering parts). This figure cannot be read as “the general accuracy of AI quoting”. The accurate understanding is that, while there is research showing this level of estimation is possible under limited conditions, you will not know the accuracy unless you measure it on your own parts.

At present, we have found no public or independent benchmark for the accuracy of AI quoting. For any accuracy figure a vendor presents, always check which group of parts, which period and which criteria it was measured on.

Step 4: The Approval Gate (Writing Down When People Must Stop the Quote)

People approve the final price and the promised lead time. What matters here is not just saying “people will look at it” but writing down under which conditions an AI quote must not go through as is, and must be stopped and checked by a person. If the conditions are vague, checks become a formality, especially in busy periods.

Here are examples of conditions for stopping at the approval gate. None of these are industry standards; they are examples of criteria you set yourself.

On the approval screen, display not only the AI’s quoted amount but also the similar past jobs it was based on, highlights of the critical notes it read, and the reason it was stopped, side by side. Whether the screen lets the approver confirm “why this amount” within a few tens of seconds is an item that should be included in the acceptance criteria.

How Machining, Sheet Metal and Plastic Molding Differ

Even within “machining quotation”, the kind of difficulty differs by process. The following is a general overview of where care is needed.

CNC machining: machining time varies greatly with geometric features (holes, pockets, threads, grooves and so on), tolerances and surface roughness. Turned parts and prismatic parts follow different processes, and the number of setups and the need for fixtures also have an effect. Reading and machining time estimation are at the center of the quote, and connecting to machining-side data such as AI CAM increases the basis for estimates.

Sheet metal: cost is driven by the flat pattern, the number of bends, the number of holes and notches, and nesting efficiency, that is, how many pieces can be cut from one sheet of material. Because the way material is used changes with quantity, you need a system that produces quotes by quantity. When welding or painting is involved, the weight of outsourcing and downstream processes rises.

Plastic molding: the structure of the quote changes completely depending on whether a mold exists. Mold cost and per-part molding price must be considered separately, and when a new mold must be made, mold design judgment comes in, so automatic quoting from drawings becomes even more difficult.

Common to every process is not targeting all processes and all materials from the start. The realistic approach is to start with the process and material with the most jobs, and widen the scope once the acceptance criteria are met there.

How to Measure Accuracy: Building Acceptance Criteria for AI Quoting

Backtesting (comparison on past jobs)

The most reliable way to check the accuracy of AI quoting is to pass drawings of jobs you actually won and made in the past through the AI and compare the AI’s quote with the actual cost. You do not compare against past quoted amounts, because those quoted amounts may themselves have been wrong. The jobs used for comparison must be kept separate from the jobs used for training.

Look at the “distribution of errors”, not the average

Judging accuracy by a single figure for average error is dangerous. Even if the average is good, if some jobs are far off, you make a loss on those jobs. What you should look at is the distribution of errors, and in particular “how many jobs were far off on the low side (quoted too cheaply)”.

Example metrics to include in the acceptance criteria

The pass values shown here are all examples you set yourself. Because the appropriate level differs by part group and customer, you must not adopt a vendor’s “accuracy of X%” as your acceptance criterion as is. How to decide where to end a PoC and move to full deployment is also covered in our article on exit criteria for AI PoCs.

Cost and ROI: Labor Savings Alone Will Not Pay It Back

From here, we run estimates using this article’s original model factory. To repeat, all of the following figures are original assumptions and estimates made for this article, not industry averages.

Assumptions for Model Factory Q

• A Japanese-affiliated contract manufacturer of precision machining and sheet metal in Chonburi Province, Thailand, with 320 employees
• 1,800 requests for quotation (RFQs) per year (150 per month)
• Average estimator working time per quote of 2.0 hours (assumed value, including drawing review, process design, inquiries to subcontractors, calculation and approval)
• Annual quoting hours: 1,800 quotes × 2.0 hours = 3,600 hours
• Estimator labor rate of 400 baht per hour (a placeholder value for this article, covering a technical estimator’s salary plus social security and similar costs)
• Annual labor cost of quoting: 3,600 hours × 400 baht = 1,440,000 baht
• Win rate of 20%, so 1,800 quotes × 20% = 360 orders won per year
• Average order value of 300,000 baht per job and a contribution margin of 25%, so contribution profit per job is 75,000 baht
• Current average quote response lead time of 5 business days

Assumed configuration and costs

The configuration is “AI drawing reading + similar-drawing search + cost tables, with the final price approved by a person”.

Key point 1: Labor savings alone take 10 years to pay back

Suppose quoting time is cut by 50% (from 2.0 hours to 1.0 hour per quote). The hours saved are 3,600 hours × 50% = 1,800 hours, which in labor cost comes to 1,800 hours × 400 baht = 720,000 baht per year.

Subtracting the annual operating cost of 600,000 baht leaves an annual net benefit of only 120,000 baht. The simple payback period is 1,200,000 ÷ 120,000 = 10.0 years. The five-year cumulative result is 120,000 × 5 − 1,200,000 = −600,000 baht, a loss. Even if you halve the labor hours, labor savings alone do not recover the investment.

A 1-point rise in the win rate changes the picture

Next, suppose, as an effect of shorter response lead times, that the win rate rises by 1 percentage point, from 20% to 21%. The additional orders are 1,800 quotes × 1% = 18 jobs, and the increase in contribution profit is 18 jobs × 75,000 baht = 1,350,000 baht per year.

In this case, the annual net benefit is 720,000 + 1,350,000 − 600,000 = 1,470,000 baht. The simple payback period is 1,200,000 ÷ 1,470,000 = about 0.8 years (about 10 months), and the five-year cumulative result is 1,470,000 × 5 − 1,200,000 = 6,150,000 baht.

Three cautions about the calculation. First, the win-rate effect is counted as “contribution profit”. Adding the 300,000 baht order value directly to the benefit would greatly overstate it. Second, the labor savings of 720,000 baht (labor cost) and the win-rate effect of 1,350,000 baht (contribution profit) are different kinds of benefit, so they are added together; and because the number of RFQs, 1,800, does not change, winning more orders does not increase quoting work. Third, this estimate assumes there is enough production capacity for the additional orders.

The conclusion is clear. What drives the investment decision is not labor savings but the win rate. However, the win rate is also affected by price, quality, delivery, the economy and circumstances on the customer side, and proving that a change is the effect of AI quoting is not easy. That is exactly why you need to decide at the outset on a design in which you record the current win rate, response lead time and reasons for lost orders for 30 days before introduction, and compare the same metrics after introduction. Conversely, you are entitled to be skeptical of any proposal that promises an improvement in the win rate.

Key point 2: The cost of a single misread

Here is one more yardstick. Suppose that missed tolerances or notes such as heat treatment and plating lead to 4 orders a year being won below cost, with a loss of 150,000 baht per job. The loss is 4 jobs × 150,000 baht = 600,000 baht per year.

This equals about 83% of the 720,000 baht labor saving in Key point 1 (600,000 ÷ 720,000 = 0.833). Note that this loss figure is presented as a yardstick for comparison with the labor saving; it is not meant to create a separate estimate by subtracting it from the net benefit in Key point 1.

What this shows is that the moment you remove the approval gate, AI quoting can lose most of the labor cost it saved. This is the basis for including in the acceptance criteria both the stop conditions at the approval gate and a screen where people check the reading results (highlighting critical notes).

About tax incentives for SMEs in Thailand

Accounting firm commentary reports that Royal Decree No. 802, which allows a 200% tax deduction for digital investment by small and medium-sized enterprises, took effect in Thailand in February 2026. It applies to companies with paid-up capital of 5 million baht or less and annual revenue of 30 million baht or less, covers expenditure from June 1, 2025 to December 31, 2027, and is said to be capped at 300,000 baht, on the condition that software and hardware are registered with the Digital Economy Promotion Agency (depa). The interpretation needs to be confirmed, including whether the cap applies to the additional deduction or to the total, and a company the size of Model Factory Q may not meet the eligibility requirements. Please confirm individually with a qualified tax professional or the Revenue Department whether it applies to you.

12 Items to Put in the RFP for a Machining Quotation System

Many disputes in ordering AI quoting arise because the specification does not state “what counts as complete”. We recommend writing at least the following 12 items into your request for proposal (RFP).

Item 5, “Ownership of and editing rights to the cost model”, is especially easy to overlook. If the cost model becomes a vendor black box, you will need to ask the vendor every time material unit prices or machine rates change, and you will no longer be able to explain the basis of your own quotes. See also our article on how to choose an AI development company for assessing developers, and our article on system development contracts for the points to secure in the contract.

What to Check in FAT/SAT

In the factory acceptance test (FAT) and site acceptance test (SAT), check the parts you cannot see in a demo.

• Backtesting on past jobs: On past won jobs not used for training, compare estimated cost with actual cost and check whether the error distribution meets the acceptance criteria
• Zero misses on critical notes: Build your own test drawing set that deliberately includes heat treatment, plating, tight geometric tolerances, special inspection instructions and so on, and check whether all of them are picked up
• The approval gate reliably stops: Feed in drawings that meet stop conditions and check that every single one is routed to an approver
• No personal names remain: Check that the design ensures designer and approver names in the title block do not remain in reading results or training data
• Handling of drawings with different revisions: When drawings with the same drawing number but different revisions are fed in, check that they are treated as separate items and that actual costs from an old revision are not mistakenly used
• Rotated and scanned drawings: Check how the system behaves with skewed scanned drawings or low-resolution drawings, and whether it displays “unreadable” when it cannot read them

It is important that the test drawing set be built from your own actual drawings, not prepared by the vendor. Vendor-supplied samples do not reflect the quirks of your drawings or the ways notes are commonly written by your customers.

Thailand-Specific Issues: Drawing Confidentiality, PDPA, Export Control and Multiple Languages

Customer drawing confidentiality and cloud AI

Customer drawings are, in many cases, covered by non-disclosure agreements (NDAs). First, check whether your NDA with the customer restricts providing drawings to third parties or storing them in the cloud.

When using cloud generative AI APIs, the major vendors state in their terms or official documentation that they do not use commercial API customer data for training. Anthropic’s Commercial Terms of Service (effective June 17, 2025) state that it does not train models on customer content from its commercial services. OpenAI’s official documentation states that data sent to the API is not used for training unless you explicitly opt in, and also states that abuse monitoring logs are retained for up to 30 days and that Zero Data Retention is available subject to approval. Japan, Singapore and other regions can be selected as data storage locations, but Thailand is not on the list.

The point to note here is that “not used for training” and “not stored” are different things. Even if data is not used for training, logs may be stored for a certain period. Personal plans also have different terms from commercial contracts. Customer drawings should always be handled through a commercial contract concluded by the company, and the practice of employees having drawings read through their personal accounts should be prohibited.

PDPA (Personal Data Protection Act)

Thailand’s Personal Data Protection Act (PDPA) defines personal data as “any information relating to a person which enables the identification of such person, whether directly or indirectly”. The geometry and tolerances of a drawing are not normally considered to fall under this. However, designer and approver names in the title block, and the names and contact details of customer contacts in RFQ emails, may be personal data. It is safest to design the system so that personal names are removed from reading results and training data. Please confirm the specific handling individually with your legal department or a qualified professional.

Export control

It has been reported that, under a Ministry of Commerce notification based on Thailand’s trade control law relating to weapons of mass destruction, from July 30, 2026 the export and re-export of Category 0 (nuclear-related) dual-use items require a permit through the system of the Department of Foreign Trade (DFT). The direct impact on quotes for general machined parts is considered limited, but depending on the content of a drawing, care may be needed in handling technical data. The information in this article is based on secondary sources such as news reports, so please confirm individually with the Department of Foreign Trade (DFT) of the Ministry of Commerce or a qualified professional whether this applies to you.

Multiple languages

At Japanese-affiliated factories in Thailand, Thai estimators, Japanese sales staff, and the Japanese head office or Japanese customers are all involved in the same quote. Notes on drawings can also mix Japanese, English and Thai. Unless the item names in reading results, especially material designations and surface treatment names, are standardized in a terminology dictionary, the same material will be registered under different names, and both similarity search and cost calculation will break down. Appointing an owner for the terminology dictionary is also part of preparing for introduction.

A 90-Day Plan for Introducing AI Machining Quotation

Finally, here is how to proceed with introduction over 90 days.

The most important thing in the first 30 days is not to play with the AI but to record the current state. If you do not know the win rate and response lead time before introduction, no one can explain what improved after introduction. As we saw in Key point 1, what drives the investment decision is the win rate. The foundation for comparing that win rate can only be built before introduction.

In the parallel run in days 31-60, the AI’s quotes are not sent to customers but compared side by side with human quotes. This is where you start to see which kinds of drawings the AI tends to get wrong, and under which conditions quotes should be stopped at the approval gate.

Frequently Asked Questions (FAQ)

Q1. What is AI quoting? How is it different from a conventional quoting system?

A conventional quoting system was a tool in which a person entered the material and machining time and a fixed formula produced an amount. AI quoting differs in that AI takes on the preceding parts: “reading the drawing”, “finding similar past jobs” and “estimating machining time”. However, the dividing line, that people approve the final price and the delivery response, does not change.

Q2. How accurate is automated quoting from drawings?

There is no general accuracy figure. We have found no public or independent accuracy benchmark, and the figures from academic research are pre-peer-review results obtained on a dataset of a specific group of parts. The only way is a backtest in which you pass your own past won jobs through the AI and compare with actual costs, measuring the error distribution and the reading recall. You set the pass values yourself.

Q3. Does the difficulty of AI quoting differ between CNC machining and sheet metal?

The difficulty is different. In general terms, machining centers on estimating machining time from geometric features and tolerances, while for sheet metal the flat pattern, the number of bends and nesting efficiency drive cost. For plastic molding, the structure of the quote changes depending on whether a mold exists. The realistic approach is to start by narrowing down to the process and material with the most jobs.

Q4. Is it safe to send customer drawings to cloud AI?

First, check whether your NDA with the customer restricts providing drawings to third parties or storing them in the cloud. The major vendors state that they do not use commercial API data for training, but “not used for training” and “not stored” are different things, and logs may be stored for a certain period. Names in the drawing title block and similar information may be personal data, so please confirm individually with your legal department or a qualified professional, including PDPA handling. It is also safest to confirm individually with the relevant authority, the Department of Foreign Trade (DFT), or a qualified professional whether the content of a drawing may be subject to export control.

Q5. Can a small or mid-sized contract manufacturer introduce AI quoting? What does it cost?

In this article’s estimate for Model Factory Q (320 employees, 1,800 RFQs a year), we set an initial investment of 1,200,000 baht and an annual operating cost of 600,000 baht. These are assumed values and vary widely with configuration and scope. Please check, using your own numbers, the structure in which labor savings alone take 10.0 years to pay back, while a 1-point rise in the win rate pays it back in about 10 months. Thailand’s 200% tax deduction for SME digital investment has capital and revenue requirements and its interpretation also needs to be confirmed, so please confirm individually with a qualified tax professional or the Revenue Department.

Q6. How long does introduction take?

This article presents a 90-day plan: recording the current state and preparing data in the first 30 days, a limited-scope parallel run in the next 30 days, and FAT/SAT and the decision on widening the scope in the last 30 days. How well your past actual costs are organized greatly changes the workload of the first 30 days.

Summary

• Introducing AI into machining quotation does not mean letting AI decide prices; it means splitting a quote into four steps (reading, similarity search, cost calculation and approval) and defining the boundary between AI and people.
• The purpose of reading is, before price calculation, not to miss dangerous requirements such as tolerances, heat treatment and plating.
• The accuracy of similarity search depends on whether it is linked to actual costs, not to past quoted amounts.
• Write down the conditions for stopping at the approval gate, and make it possible to check the basis and critical notes on the approval screen.
• Measure accuracy on your own parts, with backtesting on past won jobs and the distribution of errors.
• In the model estimate, labor savings alone took 10.0 years to pay back, and what drove the investment decision was the win rate. On the other hand, losses from misreads could reach about 83% of the labor saving, so the approval gate cannot be removed.
• Recording the current state in the 30 days before introduction is the foundation for every comparison.

You are welcome to consult us even at the stage of working out how to record your current quoting hours and win rate, or where to draw the line on which drawings to pass through AI quoting. Including cases where you want to rerun the estimate with your own numbers, please feel free to reach out via our contact page.

References

• The Next Web (CloudNC funding and Quote Agent): https://thenextweb.com/news/cloudnc-20m-cam-assist-quote-agent
• Robotics 24/7 (IMTS 2026, CloudNC): https://www.robotics247.com/article/imts-2026-cloudnc-raises-20m-to-expand-ai-tools-for-precision-machining
• Paperless Parts (fifth patent awarded): https://www.paperlessparts.com/press/paperless-parts-awarded-fifth-patent-advancing-its-position-as-the-leading-intelligence-platform-for-the-industrial-supply-chain/
• Paperless Parts (Requirements Review announcement): https://www.paperlessparts.com/press/paperless-parts-new-ai-features-surface-critical-requirements-helping-shops-quote-faster-with-confidence/
• CADDi (next-generation platform announcement, Japanese): https://caddi.com/ja-jp/news/press/20260806
• CADDi (Series D funding, Japanese): https://caddi.com/ja-jp/news/press/series-d-funding
• CADDi (survey of 484 manufacturing executives, Japanese): https://caddi.com/ja-jp/news/press/20260910_aiimpact
• Mynavi News (official launch of CADDi Quote, Japanese): https://news.mynavi.jp/techplus/article/20240918-3027513
• CIMdata (expansion of the Xometry Instant Quoting Engine): https://www.cimdata.com/en/industry-summary-articles/item/29596-xometry-deepens-ai-native-marketplace-advantage-with-new-enterprise-lead-time-intelligence-and-personalized-pricing-models
• Netshop Tantosha Forum / PR TIMES (tolerance linkage between meviy and iCAD SX, Japanese): https://netshop.impress.co.jp/index%2Ephp/r/prtimes/items/000000332.000012100
• Digital Commerce 360 (Protolabs ProDesk): https://www.digitalcommerce360.com/2026/02/17/protolabs-launches-prodesk-ai-quoting-production-management/
• ISO 1101:2017: https://committee.iso.org/standard/66777.html
• ISO 10303-242:2025 (STEP AP242): https://committee.iso.org/standard/84300.html
• ASME Y14.5-2018: https://www.asme.org/codes-standards/find-codes-standards/y14-5-dimensioning-tolerancing/2018
• arXiv 2508.12440 (manufacturing cost prediction from 2D drawings, not peer reviewed): https://arxiv.org/abs/2508.12440
• Anthropic (Vision): https://platform.claude.com/docs/en/build-with-claude/vision
• Anthropic (PDF support): https://platform.claude.com/docs/en/build-with-claude/pdf-support
• Anthropic (Commercial Terms of Service): https://www.anthropic.com/legal/commercial-terms
• OpenAI (Your data): https://developers.openai.com/api/docs/guides/your-data
• OpenAI (PDF files): https://developers.openai.com/api/docs/guides/pdf-files
• Google Gemini API (Document processing): https://ai.google.dev/gemini-api/docs/document-processing
• Thailand Personal Data Protection Act (PDPA), English translation: https://pdpa.msu.ac.th/wp-content/uploads/2022/05/EN-PDPA-2019.pdf
• Nation Thailand (export permit system for dual-use items): https://www.nationthailand.com/news/policy/40068484
• Lex Bangkok (Thailand dual-use export controls): https://lexbangkok.com/thailand-dual-use-export-controls/
• Thailand Board of Investment (BOI), first-half 2026 announcement: https://www.boi.go.th/upload/content/PR117_2569EN.pdf
• Trading Economics (Thailand vehicle production, FTI announcement): https://tradingeconomics.com/thailand/total-vehicle-sales/news/587114
• Forvis Mazars (200% tax deduction for SME digital investment in Thailand): https://www.forvismazars.com/th/en/insights/doing-business-in-thailand/tax/how-smes-benefit-from-the-200-tax-deduction