Let AI Calculate AI Adoption ROI, Then Check It — Practicing AIB-C01 Measurement with Print Monitoring at a 3D Print Workshop

When we gave a local AI a problem with the failure-rate value removed, it filled in 0.1 on its own as “a typical 3D printing failure rate” and answered that the break-even point was 345.42 months. Given the same problem, claude -p returned no number. Instead it returned a formula showing that the investment cannot be recovered unless the failure rate exceeds about 0.0824. When you ask AI to calculate AI adoption ROI, checking whether the arithmetic adds up is not enough. You also have to check who decided the definition of the formula and the baseline values.
This article contains affiliate links. As an Amazon Associate, we earn from qualifying purchases.
The exam guide for AWS Certified AI Business Strategist (AIB-C01) contains skills 2.2.1–2.2.5, which cover KPIs, baselines, ROI, leading indicators and cost management. So we built problems of the same shape around a fictional small 3D printing workshop deciding whether to adopt AI print monitoring, had claude -p (model ID claude-sonnet-5-5) and a local gemma4:12b solve them, and compared their answers with correct values computed by a script. The workshop is set in Japan, so its inputs are Japanese: prices in yen, a Japanese electricity rate and the Japanese minimum wage. The workshop’s numbers are this site’s assumptions, not figures from a real business.
Here are the results first.
- 20 problems with the formula and values given: we scored four items: B (monthly benefit), K (monthly cost), ROI and break-even months.
claude -pgot all 60 runs right when calculating in prose without tools, and all 60 right when allowed to run Python. gemma4:12b’s prose calculation got 12 of the 20 problems right. B, K and months were correct in all 20, but ROI was wrong in 8 - 6 problems with one value removed: in 16 of 18 runs,
claude -ppointed out the missing value and did not turn the affected items into numbers. In the remaining 2 runs it assumed S (the monthly service fee) was 0 yen, calculated, and stated that this was an assumption. gemma4:12b filled in a value on its own and calculated in all 6 problems - 5 problems without the formula: the same values produced different ROIs.
claude -panswered P04 with −0.2014, −1.117 and −1.1167; this site’s formula gives −0.896
What these results suggest is that a person has to fix the ROI definition, the baseline and the assumptions before any calculation starts. Check the calculation with code, and when a value is missing, do not let AI fill it in. This site has not taken AIB-C01 and has not tried the monitoring services on real hardware. The formula in this article is this site’s own definition, not AWS’s and not the exam’s answer. The overall picture of the certification is in What Does AWS’s AI Business Strategist Certification Prove? Preparing for a Certification for Non-Coders in 13 Weeks, Before the February 15 Early Badge (published 2026-11-16), and the decision whether to use AI at all is covered in Can AI Answer “Don’t Use AI”? Testing AIB-C01 Adoption Decisions on Cases Where Rules Are Enough (published 2026-11-18).
- What the Exam Guide Asks About AI Adoption ROI — Skills 2.2.1–2.2.5
- How Two AWS Resources Measure ROI — The Formula They Actually Write Down Is Cost per Outcome
- The Fictional Workshop’s Values — Separating Real Values from Assumptions
- This Article’s ROI Formula and Break-even Months
- Test Conditions — Three Kinds of Questions and Two Models
- Result 1 — Given the Formula and Values, claude -p Was Right in All 60 Runs, in Prose and in Python
- Result 2 — With One Value Removed, gemma4:12b Filled In the Value Itself in All 6 Problems
- Result 3 — Without the Formula, the Same Values Produced Different ROIs
- The Monthly Fee Changes the Conclusion — Moving Assumptions to Find the Boundary
- Intangible Benefits, Leading Indicators and Cost Management in the Workshop’s Tables
- Reading the Results — Before Calculating, a Person Decides the Definition, the Baseline and the Assumptions
- What This Test Cannot Tell You
- Summary — Fix the Definition and Baseline Before Having AI Calculate AI Adoption ROI
- Sources
What the Exam Guide Asks About AI Adoption ROI — Skills 2.2.1–2.2.5
Domain 2 of AIB-C01 accounts for 28% of the score, and Task 2.2 within it is “Measure and demonstrate AI business value.” Five skills sit under this task. The wording below is quoted from the English exam guide.
| Skill | Exam guide wording | What this article practices |
|---|---|---|
| 2.2.1 | Define key performance indicators (KPIs) for AI initiatives, including tangible benefits (for example, cost reduction, revenue growth) and intangible benefits (for example, customer satisfaction, employee productivity). | Split the workshop’s KPIs into those turned into money and those written on a separate line |
| 2.2.2 | Establish baseline metrics before implementing AI solutions to accurately measure AI adoption effects. | See what happens to the calculation when pre-adoption values are missing |
| 2.2.3 | Calculate AI return on investment (ROI) by using comprehensive frameworks (for example, time savings, cost reduction, revenue growth, productivity gains). | Fix a formula, calculate, and check with code |
| 2.2.4 | Identify leading indicators that predict AI project success. | Choose the indicators to watch during adoption |
| 2.2.5 | Implement basic cost factor controls for AI implementations, including cost planning and optimization strategies. | List the costs that grow beyond the monthly fee |
The Japanese edition of the guide renders “initiative” as a katakana loanword, which many Japanese readers take to mean an effort or a program of measures. How the Japanese edition’s wording compares with the English original is examined in Comparing the Japanese AIB-C01 Exam Guide with the English Original: Finding, with AI, Where Translations Change the Meaning Across 58 Skills (published 2026-11-17).
The point that is easy to miss is how employee productivity is treated. In 2.2.1, employee productivity is an example of an intangible benefit, while in 2.2.3, productivity gains appear among the examples of an ROI framework. You can read this as saying intangible benefits may enter the ROI formula once converted into money, but AWS does not say so. In this article, intangible benefits are not added to the monetary total; they go on a separate line. The reason is shown with numbers in a later section.
The exam guide is a list of skills. It contains neither formulas nor procedures. To practice, you have to assemble a measurement method from public material.
How Two AWS Resources Measure ROI — The Formula They Actually Write Down Is Cost per Outcome
We use two AWS resources from 2026 as the backbone: the AWS Cloud Financial Management blog post “Calculating the Return on Investment (ROI) of AI” (Adam Richter, 2026-08-10) and the AWS Prescriptive Guidance document “Economics for agentic AI on AWS” (first published 2026-01-28). The latter is about agentic AI; applying it to print monitoring is this site’s judgment.
Tie costs to outcomes, and divide by the increment
The blog walks through allocating costs, tying them to the business outcomes they support, computing the cost of one outcome, and comparing that cost with the value of the outcome. The one thing it writes down as a formula is this:
Cost per Outcome = AI Cost / Business Value MetricThe blog then says, “Cost per Outcome is not ROI. It is the unit-level building block that will help you determine ROI.” The cost of one outcome is not ROI itself but material for determining ROI. So memorizing Cost per Outcome as “AWS’s ROI formula” would be a mistake. We also did not find a sentence stating “ROI = (benefit − cost) ÷ cost” in this blog or in the relevant Prescriptive Guidance pages. That statement is limited to the two resources we used as the backbone. Other AWS blog posts do give formulas, such as a weighted ratio for AI investment in the space sector (MAIR) and the net annual value of agentic automation, but neither takes the form “(benefit − cost) ÷ cost.”
The key point in the blog’s example is how it handles the baseline. In the example, a team that fixed 5 bugs a week fixes 15 after adopting AI. The AI cost of $5,000 is not divided by 15. It is divided by the 10 additional bugs that AI produced (15 − 5), giving $500 per bug. Without the pre-adoption value, that subtraction is impossible. The blog also says that once you choose a metric, you should “establish a pre-AI baseline.”
The example rests on assumptions: that the effect appears immediately and that developer costs are fixed. The blog also notes that there is often a ramp-up period before results appear. It adds that internal-facing AI often does not tie directly to revenue, which can make it harder to connect to ROI, and it advises choosing metrics that cannot be gamed and not deciding AI strategy on the ROI calculation alone. Print monitoring does not directly increase sales, so it falls into this internal-facing category. That mapping is this site’s.
An example where the baseline alone changes the break-even point
The Prescriptive Guidance case study is about AI for recruiting. The initial investment is $23,000, and the AI’s monthly cost when processing 500 applications a month is $2,667 (including $1,917, the initial investment spread over 12 months); both are the same in either scenario. In Scenario A, where the human work costs $22,500 a month, the break-even point is 1.16 months. In Scenario B, where the human work has been improved from 45 minutes to 15 minutes, the monthly cost becomes $7,500 and the break-even point is 4.76 months.
The document does not give a break-even formula, but the table values can be reproduced as “initial investment ÷ monthly savings” (23,000 ÷ 19,833 = 1.16, 23,000 ÷ 4,833 = 4.76). Note, however, that the document’s monthly cost includes the amortized initial investment, so it counts differently from this article’s formula, which keeps the initial investment out of the monthly cost. The AI and the investment are identical; only the cost of the current operation changed. The document puts it as “the ROI calculation depends critically on your current operational baseline.” The 1.16 and 4.76 months are values from the recruiting example and do not carry over to the workshop. All this article borrows is the pattern that the baseline changes the result.
Another Prescriptive Guidance document says, “ROI cannot be treated as a static calculation that is performed at launch.” It asks you to keep tracking ROI as a KPI instead of calculating it once at adoption. We apply this to the workshop in the section on leading indicators.
Read AWS CAF-AI as a historical document
There is also AWS CAF-AI (AWS Cloud Adoption Framework for Artificial Intelligence, Machine Learning, and Generative AI, published 2024-02-13) as a framework for AI adoption. However, the document states “This whitepaper is for historical reference only.” and cannot be treated as current guidance. The word ROI appears once in its body and the word baseline does not appear at all, so it cannot serve as a source for an ROI definition either. One passage that lists defect detection on a production line as an example business outcome is close to the workshop scenario, but this article only mentions it as text from a historical document.
In addition, AWS offers a short Coursera course with the same name as Task 2.2, “AI Business – Measure and Demonstrate AI Business Value,” whose overview page refers to a “four-component” ROI framework. We opened only the overview page and did not check the course material itself.
The AWS guidance builds ROI from business value metrics, not from a single formula. For a broader view of where generative AI creates measurable value, this O’Reilly book is a good companion.
The Fictional Workshop’s Values — Separating Real Values from Assumptions
The subject is a fictional small 3D printing workshop in Japan. It prints PLA on a Bambu Lab P1S, and we set one print job at 5 hours. The workshop’s figures are not values from a real business. On the other hand, the material price, electricity rate, minimum wage and monthly service fees are public values. The table below separates which values are which.
Among the assumptions, the material weight, markup, print time and assumed operating hours come from the worked example in Recalculating 3D Printing Costs at Japanese Electricity Rates (published 2026-08-19), and the 5-minute check comes from the example in Setting 3D Print Monitoring Thresholds from Material Cost: Where to Stop and When to Ask a Human (published 2026-10-17).
| Item | Symbol | Value | Type | Source and conditions |
|---|---|---|---|---|
| PLA unit price | — | 2.4 yen/g | Real value | PLA Basic 1 kg on the Bambu Lab JP store, 2,400 yen including tax (without spool; 2,900 yen with spool). As of November 7, 2026, shipping not included |
| Average power | — | 105 W | Real value | Average for PLA on the P1S, from Bambu Lab’s measurement of about 16 minutes |
| Electricity rate | — | 31 yen/kWh | Real value (guideline) | Guideline unit price (tax included) set by Japan’s Home Appliance Fair Trade Conference. Actual rates vary by contract |
| Printer price | — | 105,000 yen | Real value | P1S AMS 2 Combo (tax included, as of November 7, 2026). The AMS Combo at 95,000 yen has no restock date |
| Hourly wage | L | 1,177 yen | Real value (average minimum wage) | National weighted average of Japan’s regional minimum wages for fiscal 2026 |
| Monthly service fee | S | 0, 785.8, 1,100.7, 1,888.1 yen | Converted real values (except 0 yen) | Pricing table below |
| Material per job | — | 80 g, 15% markup | Assumption | Cost worked example |
| Print time per job | T | 5 hours | Assumption | Cost worked example |
| Assumed operating hours | — | 2,000 hours | Assumption | Cost worked example |
| Check per detection | q | 5 minutes | Assumption | Threshold article example |
| Monthly print hours | H | 120–600 hours | Assumption | Varied per problem |
| Failure rate | f | 0.05–0.20 | Assumption | Varied per problem |
| Share of failures stopped | r | 0.5–0.9 | Assumption | Varied per problem |
| Progress when stopped | s | 0.3–0.7 | Assumption | Varied per problem |
| False-stop rate | e | 0.01–0.05 | Assumption | Varied per problem |
| Initial cost | I | 0–100,000 yen | Assumption | Varied per problem |
Direct cost per job: 499.575 yen
The direct cost C of one 5-hour print job is the sum of material, electricity and machine costs.
Material = 80 g × 2.4 yen/g × 1.15 = 220.8 yen
Electricity = 0.105 kW × 5 hours × 31 yen/kWh = 16.275 yen
Machine = 105,000 yen ÷ 2,000 hours × 5 hours = 262.5 yen
Direct cost C = 220.8 + 16.275 + 262.5 = 499.575 yenThe threshold article’s 494.6 yen included three assumptions: filament at 3 yen/g, an average of 120 W and a printer price of 80,000 yen. This article replaced those three with real values, which added 4.975 yen. The 105 W average power comes from a measurement of about 16 minutes, and the object printed and the conditions of the measurement are not stated. The P1S’s rated maximum power (1000 W @220 V, 350 W @110 V) is not an average during printing, and no official figure is given for Japan’s 100 V supply, so we do not use it in the calculation.
Human checks: 5 minutes at 1,177 yen an hour is 98.08 yen
We assumed each detection takes 5 minutes to check and used 1,177 yen, the national weighted average of Japan’s regional minimum wages for fiscal 2026, as the hourly wage. Five minutes comes to 98.08 yen. The threshold article assumed an hourly wage of 1,200 yen and put one check at 100 yen, so the two are about the same size.
Note that 1,177 yen is the minimum wage weighted-averaged across Japan, not an actual labor cost. It does not include social insurance contributions and the like. The new rates take effect prefecture by prefecture (from October 1 to December 2, 2026), so in some prefectures they are not yet in effect on this article’s publication date.
Monthly service fee S: official price-list values converted to yen
For S, we used values from the price lists of print-monitoring services. The yen values are as of October 2026, converted at about 157 yen to the dollar (the calculation used 157.47 yen, the FRB H.10 rate for 2026-10-30). They are conversions, not officially displayed prices, and we have not checked how tax applies. The prices were checked on November 7, 2026.
| S in the problems | Dollars (per month) | Matching plan and conditions |
|---|---|---|
| 0 yen | — | A placeholder for paying no monthly fee. It does not refer to any specific free tier |
| 785.8 yen | 4.99 | OctoEverywhere Supporter Perks (Supporter below), annual billing ($6.99 billed monthly). No cap on detection hours |
| 1,100.7 yen | 6.99 | Supporter, monthly billing. Obico AI Premium on annual billing ($83.88 a year; $8.99 billed monthly) costs the same, but its included detection of 50 hours a month falls short of the H in the 20 problems, and adding packs pushes it above this amount |
| 1,888.1 yen | 11.99 | OctoEverywhere Elite Supporter Perks (Elite below), annual billing ($14.99 billed monthly). Obico AI Premium on annual billing plus Subscribe&Save Packs (fixed packs below) of 300 hours ($5 a month on annual billing) costs the same, with detection up to 350 hours a month |
The Obico combinations come with an hours condition. A print started with zero or fewer hours remaining gets no detection, so the problems’ r does not apply as is. The smallest H in the 20 problems is 120 hours, already above AI Premium’s 50 hours a month, so the 5 problems with S = 1,100.7 are read as Supporter on monthly billing, which has no hours cap. Of the 5 problems with S = 1,888.1, P14 has H = 450 hours, which also exceeds 350 hours a month, so it is read as Elite on annual billing.
Please do not read the S = 0 problems as some free tier. OctoEverywhere’s free tier (Standard) covers basic detection and limited notifications; automatically stopping on detection, the higher-tier model said to produce fewer false positives, and more frequent checks are Supporter perks. If the provider’s description is accurate, detection performance also differs, so r and e would not take the same values as for Supporter. Obico’s free tier covers 1 printer and 10 hours a month and uses a first-generation model separate from AI Premium. The problems’ r is “the share of failures the AI detects mid-print and can stop,” so it does not include any delay before a person notices a notification.
We have not run any of these services. We only put the price-list amounts into the problems’ S, so we cannot describe setup steps or usability.
Failure rate and detection rates are assumptions
We found no public statistics on print failure rates in small workshops. The one peer-reviewed measurement we found comes from a university open studio and reports a failure rate of 41.1%, but that is a place with many beginners, and we could read only the abstract. We do not use it as the workshop’s value.
Detection performance is in the same situation. What Obico has published is a comparison with its first-generation model (63% fewer missed failures and 56% fewer false alarms, both self-reported by the provider); we found no absolute detection or false-positive rates. So f, r, s, e, H and I are all assumed values varied per problem. The cost article likewise set a 10% failure rate (90% success rate) as an assumption for its worked example and noted that failure rates vary widely by print.
This Article’s ROI Formula and Break-even Months
This article uses the ROI formula below. It is not an AWS formula but this site’s definition. Its shape follows the commonly used ROI (net profit ÷ cost of investment), with the period fixed at one year. As the explainer from the Corporate Finance Institute notes, the usual ROI has the weakness of ignoring time, so we always state the period by calling it “annual ROI.”
Monthly print jobs N = H ÷ T
Monthly failures F = N × f
Failures stopped D = F × r
False stops G = (N − F) × e
Monthly benefit B = D × C × (1 − s)
Monthly cost K = G × C × s + (D + G) × L × q ÷ 60 + S
Annual ROI = (12B − 12K − I) ÷ (12K + I)
Break-even months = I ÷ (B − K). If B − K is 0 or less, “not recoverable”The benefit B is the remaining cost saved by stopping a failure partway through. The per-job term C × (1 − s) has the same form as the missed-failure loss C_FN in the threshold article, and the per-false-stop term C × s has the same form as the false-stop loss C_FP. The assumptions are also the same: cost accrues evenly over time, a stopped job is reprinted from the start, and an unmonitored failure is only noticed at the end. The formula for the threshold p*, which decides at what probability to stop, is in that article and is not repeated here. This article sets r and e directly as assumed values.
The break-even months use the same form as the “break-even years” in Repair or Replace an Appliance by Electricity Cost — Break-Even Years from Fridges Measured in UK Homes and Current Registered Models (published 2026-11-06), applied per month. The extra upfront cost is divided by the monthly difference, and if the difference is 0 or less, we write “not recoverable.”
This formula is a narrow definition. The benefit counts only material, electricity and machine costs; it leaves out the labor of redoing work, late deliveries and jobs the workshop could not take while the machine was busy. AWS Prescriptive Guidance says the baseline cost of the current operation should include hidden costs, error rates and opportunity costs. A real decision would count more broadly than this article’s formula.
One more property of the formula is worth noting. Annual ROI can be rewritten as 12B ÷ (12K + I) − 1, so as long as B is 0 or more, it never falls below −1. A later section uses this property as a check.
A worked example
The following set of values is an illustration and is separate from the experiment’s problems. C and L are the values from the previous section; the rest are assumptions. Set H = 300, T = 5, f = 0.10, r = 0.7, s = 0.5, e = 0.02, q = 5, C = 499.575, L = 1,177, S = 0, I = 0.
N = 300 ÷ 5 = 60 jobs
F = 60 × 0.10 = 6 jobs
D = 6 × 0.7 = 4.2 jobs
G = (60 − 6) × 0.02 = 1.08 jobs
B = 4.2 × 499.575 × (1 − 0.5) = 1,049.11 yen
False-stop loss = 1.08 × 499.575 × 0.5 = 269.77 yen
Labor for checks = (4.2 + 1.08) × 1,177 × 5 ÷ 60 = 517.88 yen
K = 269.77 + 517.88 + 0 = 787.65 yen
B − K = 261.46 yen, annual ROI = 0.332In this example S is 0 yen and the initial cost is also 0 yen, so the result is in the black. The results of changing S with the same values are tabulated in a later section.
Test Conditions — Three Kinds of Questions and Two Models
There were three kinds of questions, and we counted prose calculation and Python execution separately.
| Question | What the AI received | Problems | claude -p | gemma4:12b |
|---|---|---|---|---|
| Formula and values (prose calculation) | 11 values and 8 lines of formula | 20 | 3 runs each | 1 run each |
| Formula and values (Python execution) | Same | 20 | 3 runs each | None |
| One value removed | 10 values and 8 lines of formula | 6 | 3 runs each | 1 run each |
| No formula | 11 values | 5 | 3 runs each | 1 run each |
The H, f, r, s, e, S and I of the 20 problems were chosen with random.Random(20261119), and we did not regenerate them after seeing any skew in the results. For the value-removed questions, we took the first 6 of the 20 problems and removed f, r, e, S, I and H in that order, one per problem. The questions did not say that a value had been removed. For the no-formula questions, we removed only the formula from the first 5 problems. The end of the value-removed questions read as follows (the prompts were in Japanese; this is a translation). Both null (do not turn it into a number) and a field for writing one’s own assumptions were provided as answer formats.
Find B, K, annual ROI and break-even months. You may show intermediate calculations, but finish with the following JSON.
{"B": number or null, "K": number or null, "ROI": number or null, "months": number or "not recoverable" or null, "assumptions": "write any assumptions you made yourself; leave empty if none"}claude -p is Claude Code 2.1.319 run non-interactively. We specified --model sonnet, and the recorded model ID was claude-sonnet-5-5. Tools and MCP were disabled. Only in the Python-execution condition did we allow Python execution and file reading, and all 60 runs included a Python execution call. Temperature cannot be set, so we repeated each run 3 times.
The local model was gemma4:12b on Ollama 0.40.15, run without thinking (think: false), at temperature 0, with num_ctx 8192. Python execution was not allowed. The system prompt was “You are a careful data analyst.” for the Python-execution condition, and for the other conditions a Japanese prompt meaning “Proceed while checking your calculations. Answer in the specified format.”
The correct answers for scoring were computed with a Python script. Each of the four items (B, K, ROI and months) counted as correct if it matched within the larger of 0.5% relative or 0.005 absolute, and a problem counted as correct only when all four items matched. We gave the AI only our own material and did not give it any official questions or course material.
Result 1 — Given the Formula and Values, claude -p Was Right in All 60 Runs, in Prose and in Python
| Model and condition | Answers | All 4 items correct | B | K | ROI | Months | Cost (USD) | Median time (s) |
|---|---|---|---|---|---|---|---|---|
claude -p, prose calculation | 60 | 60 | 60 | 60 | 60 | 60 | 0.48 | 8.4 |
claude -p, Python execution | 60 | 60 | 60 | 60 | 60 | 60 | 0.40 | 9.1 |
| gemma4:12b, prose calculation | 20 | 12 | 20 | 20 | 12 | 20 | 0.00 | 5.1 |
claude -p answered every problem correctly in both conditions, so there was no difference in the number of correct answers. The cost is the sum over 60 runs of the values claude -p reported, $0.48 and $0.40, and the median times were 8.4 and 9.1 seconds. The local model’s cost means an API cost of zero; electricity is not included.
For 13 of the 20 problems, the correct answer was “not recoverable.” That is a consequence of the assumed values we chose, not a general rule that print monitoring never pays off. The definition, which limits the benefit to material, electricity and machine costs, also plays a part. Of the remaining 7 recoverable problems, 6 also had a negative annual ROI because their break-even point exceeded 12 months (P01: 51.2 months, ROI −0.568). The last one is a problem with an initial cost of 0 yen. With ROI fixed to a one-year period, ROI is negative whenever the break-even exceeds 12 months, so these two values are consistent.
gemma4:12b got only the additions and subtractions of ROI wrong, in 8 problems
Here are the ROIs for the 8 problems gemma4:12b got wrong. In every one of them, B, K and months were within tolerance.
| Problem | gemma4:12b’s ROI | Correct ROI |
|---|---|---|
| P02 | −0.8107 | −0.942 |
| P04 | −0.3226 | −0.896 |
| P06 | −0.6019 | −0.565 |
| P10 | −0.9315 | −0.858 |
| P11 | −0.50408 | −0.938 |
| P12 | −0.9944 | −0.924 |
| P13 | −0.844 | −0.781 |
| P20 | −0.7067 | −0.631 |
Reading the answers, all 8 got 12B and 12K right, but the result of adding and subtracting in the ROI numerator or denominator did not match the formula. P11, for example, writes the numerator 2,877.552 − 26,131.67328 − 20,000 as −23,254.12128. The correct value is −43,254.12128; gemma4:12b’s value is what you get without subtracting the initial cost of 20,000 yen.
The tricky part is that every wrong ROI is still negative, and the months are correct. P11’s −0.50408 does not look out of place next to a “not recoverable” conclusion. Since the error cannot be seen from the answer itself, the only way to catch it is to go back to the formula.
How this compares with earlier measurements
In Repair or Replace an Appliance by Electricity Cost — Break-Even Years from Fridges Measured in UK Homes and Current Registered Models (published 2026-11-06), the same gemma4:12b (without thinking) got 21 of 45 items right and wrongly answered that a “not recoverable” pair would pay back in about 811 years. In this run’s 20 problems, it made no mistakes in the months and correctly answered all 13 “not recoverable” cases. But the problems differ in form and number, so we cannot say it “improved.”
In Letting AI Read Per-Appliance Power Logs — How Raw Logs, Hourly Tables and Code Execution Change the Accuracy of the Answers (published 2026-11-03), pasting raw logs returned no numbers, while files plus code execution got every question right (11/11 and 10/10). In this article’s test, by contrast, the number of correct answers was the same with or without code execution. That is because they measure different things. That article aggregated large logs; this one calculates short problems. The two results cannot be read as a single trend.
Checking the numbers with a script is the reliable part. If you want to write that kind of small Python check yourself, Automate the Boring Stuff with Python is the classic starting point.
Result 2 — With One Value Removed, gemma4:12b Filled In the Value Itself in All 6 Problems
This is the core of the article. The answers to the 6 problems with one value removed split as follows. The classification was not machine scoring; every answer was read and sorted. The reading and sorting was done by Claude (Claude Opus 5.5), which this site uses in production, without a human check.
| Model | Answers | Pointed out the gap and did not turn the missing items into numbers | Wrote an assumption and calculated everything | Calculated silently |
|---|---|---|---|---|
claude -p | 18 | 16 | 2 | 0 |
| gemma4:12b | 6 | 0 | 6 | 0 |
The answers per problem were as follows. The original values and correct months are from the problems before the value was removed and were not given to the AI.
| Problem | Value removed (original value) | claude -p, 3 runs | Value gemma4:12b set | gemma4:12b’s months | Correct months |
|---|---|---|---|---|---|
| P01-f | Failure rate f (0.2) | All 3 runs: all 4 items null; shows a formula in f | f = 0.1 | 345.42 | 51.2 |
| P02-r | Share stopped r (0.7) | B, K and ROI null; months: not recoverable | r = 1 | Not recoverable | Not recoverable |
| P03-e | False-stop rate e (0.03) | B calculated; K and ROI null; months: not recoverable | e = 0.05 | Not recoverable | Not recoverable |
| P04-S | Monthly service fee S (0 yen) | 2 runs assumed S = 0 and calculated; 1 run left K and ROI null | S = 0 | Not recoverable | Not recoverable |
| P05-I | Initial cost I (20,000 yen) | B and K calculated; ROI null; months: not recoverable | I = 0 | Not recoverable | Not recoverable |
| P06-H | Monthly print hours H (200 hours) | All 3 runs: all 4 items null; shows a formula in H | H = 160 | Not recoverable | Not recoverable |
claude -p answered only what is determined even without the value
claude -p’s answers explained why it did not set a value. Here are quotes from one P01-f run and one P02-r run (translated from Japanese).
If I chose f, the result would be almost entirely determined by that assumption, so I left it as null instead of giving a number.
Setting r arbitrarily would make the result fabricated, so I am not setting it.
It did not leave everything blank, though. It calculated the items whose conclusion does not change even with a value missing. For P02-r, B − K is negative even at the maximum r of 1, so it answered “not recoverable” regardless of r. For P03-e, K is 1,477.34 yen even with e at 0, which exceeds B’s 575.51 yen, so again “not recoverable.” For P05-I, too, it judged that if B − K is negative, the investment cannot be recovered regardless of the initial cost.
For P01-f and P06-H, it kept the missing value as an unknown and expressed the boundary as a formula. For P01-f it gave B − K = 16602.717 f − 1368.822: the investment cannot be recovered unless f exceeds about 0.0824. For P06-H, it cannot be recovered unless H exceeds about 373 hours. Both answers show what the workshop needs to measure to get an answer.
The exception was 2 runs of P04-S, which assumed S was 0 yen, calculated, and stated it was an assumption. The third run treated it as missing and left K and ROI null, so the same question split across runs. The original problem’s S happened to be 0 yen, but that information was not given to the AI; the match was a coincidence.
gemma4:12b’s assumptions have no basis
gemma4:12b set a value in all 6 problems. Quotes from its answers (translated from Japanese; math symbols and emphasis formatting removed):
Assumption: I will calculate assuming the print failure rate f is 0.1 (10%), as a typical 3D printing failure rate.
Since the value of e is not specified, I will assume e = 0.05 as a typical “false detection rate.”
I will assume monthly operating time H is 160 hours (40 hours a week × 4 weeks).
To be fair, the question had an assumptions field, so writing assumptions was something the question itself encouraged. gemma4:12b wrote its assumptions in that field every time, and neither model ever calculated silently. With a differently designed question, the answers could split differently.
The problem is that the values it wrote have no basis. It says “typical,” but as noted in the section on the workshop’s values, we found neither public statistics on failure rates in small workshops nor absolute false-detection rates. Moreover, r = 1, S = 0 and I = 0 are all the extreme values most favorable to adoption. If you assume every failure can be stopped, there is no monthly fee and no initial cost, the numbers get filled in, but they are no basis for the workshop’s decision.
In one question, the calculation broke down even on top of the values it set. The ROI of −1.00688 for P01-f cannot occur under this site’s formula, because B is positive at 2,427.9345 yen. Using the property noted in the formula section (if B is 0 or more, ROI never falls below −1), you can tell it is suspicious without redoing the calculation.
Without a baseline, there is no ROI — skill 2.2.2
Why skill 2.2.2 of AIB-C01 says to establish baselines “before implementing” becomes clear from how the values in this experiment move. Varying only f with the P01 values gives the following. B − K is 16602.717 f − 1368.822, and months are P01’s I (100,000 yen) divided by B − K. Both were calculated with this site’s script.
| f | B − K (yen) | Break-even months |
|---|---|---|
| 0.08 | −40.60 | Not recoverable |
| 0.10 | 291.45 | 343.1 |
| 0.20 (original problem value) | 1,951.72 | 51.2 |
Both 0.08 and 0.10 are within the range of f used in the 20 problems. A difference of just 0.02 flips whether the investment can be recovered. gemma4:12b’s 345.42 is close to the 343.1 for f = 0.1 but slightly off because of differences in intermediate calculations. Compared with the original problem’s 51.2 months, it is about 6.7 times larger (345.42 ÷ 51.2): a single assumed value moves the answer this much.
This failure rate can only come from the workshop’s own records. If you did not count it before adoption, asking AI will not produce it, and there will be no way to verify it later. In other words, without a baseline, an ROI calculation can only ever be “an example with assumed values plugged in.” Sorting the removed values by how you obtain them gives the following. This is this article’s own organization.
| Value | How to obtain it | Can it be fixed before adoption? |
|---|---|---|
| f (failure rate), H (monthly print hours) | Workshop records | Measurable. This is the baseline |
| S (monthly fee), I (initial cost) | Price lists and quotes | Can be confirmed |
| r (share stopped), e (false-stop rate) | Trial records, or calculation with several values | No public absolute values found; unknown until tried |
Result 3 — Without the Formula, the Same Values Produced Different ROIs
Here are the ROIs for the 5 problems without the formula. The “This site’s formula” column applies this article’s formula to the same values. We did not score which answer is correct. What we want to see is how much the number moves when the definition differs.
| Problem | This site’s formula | claude -p, 3 runs | gemma4:12b |
|---|---|---|---|
| P01 | −0.568 | −0.766, −0.766, −0.766 | 90.13 |
| P02 | −0.942 | −1.03, −1.2237, −1.03 | −2.525 |
| P03 | −0.831 | −1.021, −1.0213, −1.021 | −0.7979 |
| P04 | −0.896 | −0.2014, −1.117, −1.1167 | −1.702789 |
| P05 | −0.545 | −0.682, −0.682, −0.682 | −292.73 |
claude -p wrote out the formula it used in all 15 runs. Checking the formulas against the values, 14 of the 15 runs used the same numerator as this site’s formula, and the difference was in the denominator. Take P02. The values are B = 314.732, K = 1,247.013, S = 785.8 and I = 50,000, and the numerator is the same 12B − 12K − I as this site’s formula: −61,187.4 yen.
This site’s denominator 12K + I = 64,964.2 yen → −0.942
Initial cost plus 12 months of fees I + 12S = 59,429.6 yen → −1.03
Initial cost only I = 50,000 yen → −1.2237Whether false-stop losses and the labor for checks are “subtracted from the benefit” or “added to the cost” leaves the numerator the same but changes the denominator. None of these is wrong as a description of ROI, but values below −1, such as −1.03 and −1.2237, cannot occur under this site’s formula when B is positive. That makes them a clue that the denominator differs.
P04’s −0.2014 differs one step further. It did not subtract the initial cost from the numerator, and it added 5 hours of wages to the loss per failure (499.575 + 5 × 1,177 = 6,384.575 yen). What goes into the benefit is part of the definition too. If the same values can yield both −0.2014 and −1.117, an ROI figure handed over on its own cannot be compared with anything.
Break-even months, on the other hand, matched this site’s formula in all 15 runs (51.2 for P01, “not recoverable” for the rest). Months have no denominator question, which probably made differences in definition less likely to show.
For gemma4:12b, units wavered as well as definitions. P01’s 90.13 is “annual savings divided by the initial cost” expressed as a percentage, and P05’s −292.73 is also a value multiplied by 100. The other 3 problems are given as ratios. The question only said ROI should be a “number” and did not fix the unit, so this is partly a shortcoming of the question. The months also differed from this site’s formula in 2 of 5 problems (13.31 months for P01, 60 months for P03).
The takeaway for readers is simple: decide the ROI definition and hand it over before asking for a calculation. What goes into the numerator, what goes into the denominator, how many years the period covers, and whether the result is a ratio or a percentage. If you do not provide these, the number you get back reflects whatever definition the AI happened to choose that time.
The Monthly Fee Changes the Conclusion — Moving Assumptions to Find the Boundary
Even with the definition fixed, assumed values move the conclusion. Keeping the assumptions of the worked example, we change only S. K becomes 787.65 yen plus S.
| Monthly fee S | K (yen) | B − K (yen) | Verdict |
|---|---|---|---|
| 0 yen | 787.65 | 261.46 | Annual ROI 0.332 |
| 785.8 yen (OctoEverywhere Supporter, annual billing) | 1,573.45 | −524.34 | Not recoverable |
| 1,888.1 yen (Obico AI Premium on annual billing plus a 300-hour fixed pack) | 2,675.75 | −1,626.64 | Not recoverable |
With a monthly benefit of only 1,049.11 yen, a small difference in the monthly fee changes the conclusion. Conversely, raising f to 0.20 puts the Supporter row 154.6 yen in the black. This table supports neither “monitoring pays off” nor “it does not.” All it can say is what happens under these assumptions, and where the values that decide the conclusion lie. Compared with answers that filled in values, claude -p’s answers, which showed the boundary as a formula, do more to tell the workshop what to measure next.
Counting stopped failures as an increment
Here is one more example, using the workshop’s values, of a baseline changing the decision. This article’s formula counts every failure the AI stops, D, as a benefit, which implicitly sets the number of failures stopped partway before adoption at zero. In reality, people must have been keeping an eye on prints and stopping some failures midway.
Think of it in the same form as AWS’s 15 − 5 example. In the worked example above, suppose 4.2 failures a month are stopped after adoption, and people were already stopping 2 a month before (the 2 is an assumption). The increment is 2.2. Putting the K of the S = 0 row into the AI Cost of Cost per Outcome gives the following. Using K there is this site’s own mapping.
Benefit per stopped failure = 499.575 × (1 − 0.5) = 249.79 yen
Divided by all stops = 787.65 ÷ 4.2 = 187.54 yen
Divided by the increment = 787.65 ÷ 2.2 = 358.02 yenDivided by all stops, the cost looks lower than the benefit of one stop; divided by the increment, it is higher. If you had not counted the 2 stops before adoption, you would never notice the difference.
Intangible Benefits, Leading Indicators and Cost Management in the Workshop’s Tables
Things that do not enter the ROI formula are still tested as skills. We sort them out for the workshop. Every workshop example here is this site’s proposal, not an example given by AWS.
Keep intangible benefits on a separate line instead of adding them to the money — skill 2.2.1
The split between tangible and intangible benefits is the exam guide’s terminology. The AWS blog does not use these terms and explains business value metrics with a different three-way classification.
| Category | Exam guide examples | Workshop examples | Treatment in the ROI formula |
|---|---|---|---|
| Tangible benefits | Cost reduction, revenue growth | Material, machine time, electricity for reprints | Converted into money and included in B |
| Intangible benefits | Customer satisfaction, employee productivity | Meeting delivery dates, peace of mind from not having to watch prints at night | Not added to the money; written on a separate line |
They go on a separate line because the moment you add them, the conclusion flips. Suppose night-time checks drop by 5 hours a month and you convert that at the hourly wage: 5 × 1,177 = 5,885 yen (the 5 hours is an assumption). Add that to the B − K of the 785.8-yen row in the table above (−524.34 yen) and it turns into a surplus of 5,360.66 yen. One line of conversion assumptions outweighs the entire calculated benefit. If you convert, write down the conversion assumptions and show both the number with them added and the number without.
Leading indicators are what you watch during adoption — skill 2.2.4
The definitions of leading and lagging indicators are in an AWS Prescriptive Guidance document on change management. Lagging indicators measure success after results are in; leading indicators measure, during the project, whether it is heading toward its outcomes. This definition is not specific to AI, and we found no AWS document that lists leading indicators for AI.
| Indicator | Type | What it tells you |
|---|---|---|
| Share of detection notifications checked | Leading | If checks are not happening, the assumption behind r breaks down |
| False stops per week | Leading | Compared with the expected count (N − F) × e from that week’s N and F, it shows whether the assumed e holds. False detections that only sent a notification without stopping are not counted |
| Increment in failures stopped | Lagging | Shows the difference from the baseline |
| Monthly loss from failures, annual ROI | Lagging | Shows the result of adoption |
As the Prescriptive Guidance sentence quoted earlier says, ROI is not a one-off calculation at adoption. Economics for agentic AI on AWS says to decide in advance when to cut agents that are not performing (“Include clear decision points for terminating non-performing agents.”). For the workshop, this could take the form of reviewing the leading indicators and B − K three months after adoption and deciding whether to continue or cancel. The three months is only an example.
Cost management is not just the monthly fee — skill 2.2.5
The AWS blog lists total cost of ownership (TCO) calculation as the first element of cost and divides indirect costs into four kinds: storage, data transfer, monitoring and reporting, and agents. That is a classification for cloud AI, not for a workshop’s costs. Below are the points where a workshop’s costs grow beyond the monthly fee, taken from official documentation.
- Hours cap: Obico’s AI Premium includes 50 hours a month, with rollover and add-on packs. The cap does not go away. A print started with zero or fewer hours remaining gets no detection, while a print that runs out partway continues to be monitored for the rest of that print. The 300-hour fixed pack costs $5 a month on annual billing or $7.5 a month on monthly billing, bringing the total to 350 hours a month. Because the cost depends on the workshop’s monthly print hours, read the Obico combinations in the pricing table together with H
- Number of printers: an additional printer on Obico costs $2 on annual billing and $3 on monthly billing. On OctoEverywhere, Supporter includes 5 printers with $1 per additional printer, and Elite includes 10 with $2 per additional printer. The S in this article’s problems does not include fees for additional printers
- Stopping: OctoEverywhere’s free tier works with up to 3 printers and has no cap on detection hours, but it is limited to basic detection and limited notifications. Automatic pausing is a Supporter perk; stopping does not come free
- P1S requirements: according to the provider, you need one device to run OctoEverywhere’s Bambu Connect (separate from Bambu Lab’s software of the same name), and automatic pausing requires LAN Only Mode and Developer Mode in addition to Supporter. That description does not name specific models, but for the P1 series, firmware 01.08.02.00 (2025-06-03) introduced Authorization Control and Developer Mode. According to the Bambu Lab Wiki, Developer Mode does not connect to Bambu Cloud. It does not translate into money, but we list it as a hidden cost
- Obico and the P1S: Obico’s setup article (2024-12-19) places OctoPrint in between and uses a separate webcam; it does not name the P1S (it does say “no matter which model you have”). Another official blog post from 2026 lists Bambu Lab as supported but does not state models or the connection method. We have not confirmed whether it works with the P1S after the firmware update above
- Built-in detection: Bambu Lab’s official description of Spaghetti Detection assumes the X1, and we could not find any mention of the P1S
- Self-hosting: self-hosted Obico has no software fee; the costs are hardware and operation. It cannot be said to be cheaper. We did not collect hardware prices, so I in the problems remains an assumption
None of these are results from testing on real hardware; they are what we confirmed in official documentation on November 7, 2026. How to choose an adoption approach was covered in the adoption-decision article mentioned earlier.
Intangible benefits, leading indicators and cost control are where AI projects are usually argued. Chip Huyen’s AI Engineering covers the cost and evaluation side of running AI in production.
Reading the Results — Before Calculating, a Person Decides the Definition, the Baseline and the Assumptions
Lining up the three results, what emerges is not that the calculation is hard. With the formula and values in place, claude -p was right in all 60 runs, and gemma4:12b was right on everything except the additions and subtractions in ROI. The hard part comes before the calculation. When a value was missing, gemma4:12b filled one in for all 6 problems, and claude -p also set S to 0 yen in 2 of 18 runs. Without a formula, the same values produced different ROIs.
Translated into exam study, practicing the “calculate” in skill 2.2.3 is not enough. Deciding what counts as a benefit, what the pre-adoption values are and which values are assumptions remains a person’s job.
Steps for practicing AI adoption ROI with AI
If you practice on your own, go in this order. Give the AI only material you made yourself, and do not paste in official questions or course material. The extent to which AI can be used for studying is described in the certification overview article.
- Write the definition first: decide in words the ROI formula, the period, what goes into the numerator and the denominator, and whether the result is a ratio or a percentage
- Collect pre-adoption values: record the number of failures, the time until a failure is stopped, the loss per failure and monthly print hours, using the same definitions you will use after adoption
- Build a table of values: separate real values from assumptions in columns and write down the sources
- Do not let AI fill in missing values: ask it to return null for missing values, and read the basis of any answer that makes assumptions. We did not measure the effect of this instruction in this article, though
- Check with code: plug the AI’s answer back into the formula and see whether a script like the one below gives the same values
- Find the boundary: vary the unknown values and look for the value at which B − K becomes 0
- Keep intangible benefits on a separate line: keep both the version with them added in money and the version without
- Measure again after adoption: track the leading indicators and ROI with the same definitions and decide whether to continue
The checking script can be short.
def roi(H, f, r, s, e, S, I, T=5, q=5, C=499.575, L=1177):
N = H / T # monthly print jobs
F = N * f # monthly failures
D = F * r # failures stopped
G = (N - F) * e # false stops
B = D * C * (1 - s) # monthly benefit
K = G * C * s + (D + G) * L * q / 60 + S # monthly cost
R = (12 * B - 12 * K - I) / (12 * K + I) # annual ROI
M = I / (B - K) if B > K else "not recoverable" # break-even months
return round(B, 2), round(K, 2), round(R, 3), M
# Change only the monthly fee and compare conclusions (values from the worked example)
for S in (0, 785.8, 1888.1):
print(S, roi(H=300, f=0.10, r=0.7, s=0.5, e=0.02, S=S, I=0))
# Vary the failure rate to find the boundary (P01 values)
for f in (0.08, 0.10, 0.20):
print(f, roi(H=450, f=f, r=0.9, s=0.4, e=0.01, S=1100.7, I=100000))The reason to recommend code is not that it increased the number of correct answers in this test. claude -p was right in all 60 runs in both prose and Python. We recommend it because you can run the same formula as often as you like, compare results across changed values, and catch errors that cannot be seen by eye, like gemma4:12b’s ROI.
What This Test Cannot Tell You
Here are the limits to keep in mind when reading the results.
- We made the problems: the 20, 6 and 5 problems are all our own, and the values are assumptions. That 13 problems came out “not recoverable” follows from the values we chose; it is not a general rule. A different distribution would change the counts
- One formula: we decided the correct answers with a single formula defined by this site. The benefit covers only material, electricity and machine costs, and other definitions would give different correct answers. It is neither AWS’s definition nor the exam’s answer
- One model each: we tested
claude -pwith claude-sonnet-5-5 three times and gemma4:12b at temperature 0 once. We did not test other models or settings. You cannot generalize to “AI gets calculations wrong” or “AI calculations are correct” - Code showed no advantage in the counts: letting the model run Python did not change the number of correct answers, so at this problem difficulty, the advantage of code did not show up as a number. gemma4:12b was not allowed to execute code
- The question format can move the answers: the value-removed questions had an assumptions field and a null option. We did not test questions without the field or questions saying not to set values. The classification of answers was also done by reading every answer, not by machine scoring
- Failure and detection rates are assumptions: we found no public statistics on workshop failure rates and no absolute detection or false-detection rates. The 41.1% is a value from an abstract about a university open studio and was not applied to the workshop
- No real-hardware measurements: we did not run printing or monitoring on real hardware. We did not test detection from Obico, OctoEverywhere or Bambu Lab; prices and conditions come from official documentation
- Values change: fees and prices were checked on November 7, 2026, and change along with exchange rates, minimum wages and electricity rates
- We have not taken the exam: we have not taken AIB-C01 and do not know whether exam questions take this form
Summary — Fix the Definition and Baseline Before Having AI Calculate AI Adoption ROI
- In the two AWS resources this article was built on, the formula actually written down is Cost per Outcome, which is not ROI itself. This article’s ROI formula and break-even months are this site’s definitions
- In the 20 problems with the formula and values given,
claude -pwas right in all 60 runs, in prose and in Python. gemma4:12b got 12 of 20 problems right; B, K and months were right in all 20, but ROI was wrong in 8. The wrong ROIs were also negative and could not be spotted by eye - In the 6 problems with one value removed,
claude -pleft the missing items as non-numbers in 16 of 18 runs, and in 2 runs assumed S was 0 yen and said so. gemma4:12b calculated in all 6 problems after setting f = 0.1, r = 1, e = 0.05, S = 0, I = 0 and H = 160, with no basis for those values - In the 5 problems without the formula, the same values produced different ROIs. The main difference was the denominator: for P02, −0.942, −1.03 and −1.2237 sit side by side
- Conclusions move with assumptions. In the worked example, a different monthly fee turned the result into “not recoverable,” and in P01, whether the investment can be recovered flips between f = 0.08 and 0.10
If you want hands-on practice with AI adoption ROI, start by counting your own workshop’s failures and writing the ROI definition in one line before asking for any calculation. Only then ask AI to calculate, and check the result with code.
Sources
- AWS Certified AI Business Strategist exam guide, Domain 2 (English)
- Same, Domain 2 (Japanese edition)
- AWS Cloud Financial Management, “Calculating the Return on Investment (ROI) of AI” (Adam Richter, 2026-08-10)
- AWS Public Sector Blog, “Mission AI ROI strategy framework for space using Amazon Bedrock” (2026-10-08)
- AWS blog, “Beyond hours saved: Building the business case for agentic automation” (2026-10-07)
- AWS Prescriptive Guidance, “Economics for agentic AI on AWS”: Understanding agentic AI economics
- Same: Assessing human-process costs
- Same: Measuring success and ROI
- Same: Case study
- Same: Scenario A
- Same: Scenario B
- AWS Prescriptive Guidance, “Delivering and sustaining the value of a generative AI application”
- AWS Prescriptive Guidance, “1.7 Change adoption metrics”
- AWS Cloud Adoption Framework for Artificial Intelligence, Machine Learning, and Generative AI (PDF, historical reference)
- Coursera, “AI Business – Measure and Demonstrate AI Business Value”
- Corporate Finance Institute, “Return on Investment: Formula, Meaning, and How to Calculate It”
- Obico Cloud pricing page
- Obico, “What is AI Detection Hour and how does it work?”
- Obico, “Next-Gen AI Failure Detection Is Here: General Release”
- Obico, “AI Failure Detection and Remote Control for Bambu Lab 3D Printers”
- Obico, “Best 3D Printer Failure Detection Tools Compared” (2026-02-28)
- Obico, “How Obico AI Failure Detection Works” (2026-03-16)
- Obico, “Obico Cloud vs Self-Hosted”
- OctoEverywhere Supporter page
- OctoEverywhere, “Gadget – Free & Unlimited AI Failure Detection For 3D Printing”
- OctoEverywhere, “Do I Need LAN Only Mode For Bambu Lab 3D Printers?”
- OctoEverywhere, “AI Failure Detection and Remote Control for Bambu Lab P1S & P1P 3D Printers”
- Bambu Lab Wiki, “Third-party Integration with Bambu Lab Products”
- Bambu Lab Wiki, “P1P/P1S and AMS Firmware Release History”
- Bambu Lab Wiki, “Spaghetti Detection”
- Bambu Lab Wiki, “Bambu Lab Printer and AMS power parameters”
- Bambu Lab JP store, “P1S 3D Printer” (Japanese)
- Bambu Lab JP store, “PLA Basic” (Japanese)
- Ministry of Health, Labour and Welfare, “Revised regional minimum wages recommended for all prefectures” (September 3, 2026; Japanese)
- Home Appliance Fair Trade Conference (Japan), “FAQ” (Japanese)
- FRB H.10, USD/JPY daily
- Song & Telenko, “Causes of Desktop FDM Fabrication Failures in an Open Studio Environment,” Procedia CIRP 80 (2019), abstract (OpenAlex)





