Can AI Say “Do Not Use AI”? Testing the AIB-C01 AI Adoption Decision on Cases Where Rules Suffice

Consider a quote that is nothing more than material weight and print time multiplied by unit prices. When I asked a local AI model, “How should we introduce AI into this task?”, it answered that the best move was to deploy “a dedicated chatbot for quotes.” That is exactly the moment the AI adoption decision in AIB-C01 asks you to handle: can you say “this task does not need AI”? If the AI you study with leans toward adoption every time, its answers are no good as an answer key.
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) includes a skill for determining when AI is not the appropriate solution. So I turned the work of a small 3D printing shop into 19 cases, split them into ones that rules can handle, ones that suit AI, and ones a person should decide, and had claude -p (Claude Sonnet 5.5) and a local gemma4:12b answer them. The shop has 8 printers and 3 staff. Both numbers are assumptions made for this article, not figures from a real business.
Here is the result up front. Under the three conditions that asked for a three-way choice of Rules, AI or Human, both models answered every case correctly. A difference appeared only when the question changed shape. Asked yes or no, “Should we introduce AI into this task?”, claude -p got every case right and gemma4:12b got every case right except two where its JSON was broken. But when asked for a proposal, “How should we introduce AI into this task?”, claude -p pointed out in all 24 runs on the rule cases that the task itself does not need AI, and added ideas for using AI on a different task or on surrounding work. gemma4:12b, on 6 of the 8 rule cases, offered an AI proposal without any such remark.
This difference matters for how you phrase questions when you use AI as a study partner. Note, however, that the conditions with a different question shape were added partway through the experiment, the cases were written by this site, and only one model of each kind was tested. This site has not taken AIB-C01, and the decision procedure described here is neither the AWS scoring standard nor the correct answer on the exam. For an overview of the certification, see What the AWS AI Business Strategist certification proves, and a 13-week plan to the February 15 early-adopter badge for people who do not write code (published 2026-11-16).
- In the AIB-C01 AI adoption decision, “do not use AI” is also an answer
- How AWS and Google suggest checking: can rules do it, and is there data?
- Decision conditions K1 to K5 and their order: the thresholds are this site’s own
- The shop’s 19 cases: 8 rules, 8 AI, 3 human
- Test conditions: a three-way choice, then adding the owner’s preface and the decision order
- Result 1: in the three-way choice, both models got every case right under all three conditions
- Result 2: conditions with a different question shape, added mid-experiment, asked yes or no
- Result 3: asked “how should we introduce it?”, gemma4:12b proposed AI on 6 of 8 rule cases
- Reading the results: when you study with AI, take the premise out of the question
- Practicing build, buy or partner with print monitoring: the five factors in 2.1.2
- Prioritizing and revisiting: applying 2.1.3 to the shop’s cases
- What this test cannot tell you
- Summary: for the AIB-C01 AI adoption decision, write the conditions down before you ask
- Sources
In the AIB-C01 AI adoption decision, “do not use AI” is also an answer
The AIB-C01 exam guide lists skills related to adoption decisions across several domains. This article practices the following four. The right-hand column is what each one means for this article.
| Skill | What the skill asks (summary) | What this article practices |
|---|---|---|
| 1.2.1 | Decide when to use rule-based automation and when to use AI solutions | Splitting tasks into rules and AI |
| 2.1.4 | Decide when AI is not the appropriate solution for a business problem | Answering “do not use AI” |
| 2.1.2 | Evaluate build-buy-partner decisions for AI implementations, considering factors such as budget, timelines, capabilities, vendor proposals and regulatory compliance requirements | Choosing a way to get print monitoring |
| 2.1.3 | Prioritize AI initiatives on business value, feasibility, sustainability and strategic alignment, by deciding to scale, pause or terminate them | Putting candidates in order |
In the original English, 1.2.1 reads “Determine when to use rule-based automation and when to use AI solutions.” and 2.1.4 reads “Determine when AI is not the appropriate solution for a business problem.” In other words, the scope of an AI certification explicitly includes the decision not to use AI.
However, the domain pages of the exam guide contain only lists of tasks and skills. They do not describe how to make the decision. Reading the pages for Domain 1 and Domain 2, you will not find a sentence that draws the line between rules and AI.
A closer sentence exists in the exam guide of a different certification. Domain 1 of AIF-C01 has the item “Determine when AI/ML solutions are not appropriate (for example, cost-benefit analyses, situations when a specific outcome is needed instead of a prediction).” It gives two examples of when AI is not appropriate: a cost-benefit analysis, and situations where a fixed outcome is needed rather than a prediction. This is not AIB-C01 wording, and since it says “for example,” it is not the whole standard either. That AIF-C01 domain is covered in AIF-C01 Domain 1 complete guide: scoring the 20% on AI/ML basics and the ML lifecycle (published 2026-06-09).
Since no procedure is written down, practicing this skill means building your own decision criteria. The next section looks at the official material you can build them from.
How AWS and Google suggest checking: can rules do it, and is there data?
AWS Well-Architected documents
The main anchor is the Machine Learning Lens of AWS Well-Architected (published 2025-11-19). Its “ML problem framing” section, which is the stage for checking whether machine learning (ML) is the right tool, says:
Some business problems don’t need ML as simple business rules can do a much better job. For other business problems, there might not be sufficient data to apply ML as a solution.
Some problems are better served by simple business rules, and some do not have enough data in the first place. Because these two conditions, rules and data, appear back to back, they became the backbone of the decision criteria in this article.
MLCOST02-BP01 in the same Lens is more concrete. It states that not every problem needs ML and lists, as an anti-pattern, “Failing to establish a baseline for comparison with existing or rules-based approaches.” Its question for whether to use ML is “Consider if the problem involves complex patterns that rules can’t simply capture, or if it requires continuous adaptation to changing conditions.” In short, look for complex patterns rules cannot capture, or a need to keep adapting to changing conditions.
Its examples are close to a print shop. Fraud detection in financial transactions may benefit from ML because the methods keep changing, while “simple inventory management might be better served by a rules-based system.” At the same time, the document does not demand an either-or choice between rules and AI. It says “Consider whether combining rules-based systems with ML or generative AI could provide the optimal balance of cost, performance, and explainability for your specific use case,” listing the combination as an option too.
For generative AI the question shifts slightly. The Scoping part of the Generative AI Lens asks “Will an off-the-shelf model satisfy the requirements of this business problem or will there be a need to customize the model?”, that is, whether an off-the-shelf model is enough. The ML material asks whether you have training data, while the generative AI material starts from off-the-shelf models, so the ML wording cannot be applied to generative AI as is. For that reason the data condition here is widened to “you have your own training data, or there is an off-the-shelf model or product that solves the same problem.”
One more source: the Amazon Machine Learning developer guide also has concrete wording for when ML is not needed (“simple rules, computations, or predetermined steps”). However, this is documentation for a retired service that is no longer updated. I borrowed the phrasing as a reference only and did not use it as a primary basis.
Google’s Rules of Machine Learning and PAIR
Rules of Machine Learning on Google for Developers (Martin Zinkevich, last updated 2025-08-25) titles Rule #1 “Don’t be afraid to launch a product without machine learning.” The text continues, “If machine learning is not absolutely required for your product, don’t use it until you have data.” The same section also says “Don’t be afraid to use human editing either,” which shows that besides rules and AI there is a third answer: a person decides.
That said, quoting Rule #1 alone makes it sound as if rules always win. Rule #3 is “Choose machine learning over a complex heuristic,” followed by “A simple heuristic can get your product out the door. A complex heuristic is unmaintainable.” The same document therefore also points the other way: once rules multiply and tangle, move to ML.
The People + AI Guidebook from Google’s PAIR (People + AI Research) also gives examples of situations where rules fit better and where AI fits better. The list is hedged with “probably” and is illustrative, not an absolute standard. Its view that rules can work as well or better in some cases points the same way as the two sources above.
The AWS and Google guidance both start by asking whether ML is the right tool at all. Chip Huyen’s AI Engineering covers the same question from the builder side, along with evaluation and cost.
Decision conditions K1 to K5 and their order: the thresholds are this site’s own
From the material above, this site assembled the following five conditions. To be clear from the start, this is not an AWS standard. It is neither the exam answer nor a scoring rubric, and every threshold on lines, counts and time was set by this site.
| Condition | Question | Meaning of yes | Threshold (set by this site) |
|---|---|---|---|
| K1 | Does a procedure determine a single answer? | Rules are enough | The inputs are only numbers, true/false values or fixed options, and the answer can be written in three lines of if statements or fewer. Includes comparisons, calculations and table lookups. Also includes cases where a fixed answer is needed rather than a probability |
| K2 | Is the input images, audio or free text, or are there many overlapping or constantly changing conditions? | Candidate for AI | The input type is stated in the case text. The case text says there are 4 or more conditions, or that conditions increase every month |
| K3 | Is it a safety stop? | Excluded from the decision | Yes if people could be put in danger. Not covered in this article |
| K4 | Is there enough training data of your own, or an off-the-shelf model or product that solves the same problem? | AI is usable | The case text states the record count or whether an off-the-shelf product exists. Yes if either holds. Whether the data is enough to train on is this site’s assumption |
| K5 | Is the volume small enough for people? | AI is not needed | Yes if “one person finishes it in 30 minutes a day or less.” No for “watching constantly, including at night” or “200 items a day” |
The correct answer is determined in this order:
- If K1 is yes: “Rules”
- If K1 is no and K2 is yes, treat it as an AI candidate and continue. If K2 is also no, drop it from the cases
- If K3 is yes, drop it from the cases
- If K4 is no: “Human” (a person decides until data builds up)
- If K5 is yes: “Human” (people can handle it)
- Otherwise: “AI”
K1 comes first so that rules are checked before anything else. MLCOST02-BP01 also says “If a rules-based solution exists, use it as a baseline for comparison.” K2 is based on the MLCOST02-BP01 question and the PAIR examples, K4 on the questions in the ML Lens and the Generative AI Lens, and K5 on the human-editing point in Rules of ML, with a volume threshold added.
K3 is left out of the decision. Judgments about situations that endanger people are outside the scope of this article, and the experiment uses no case that reaches this step. R3, stopping when the temperature leaves its range, is the job of the printer’s safety functions, but it is settled as “Rules” at K1 and never reaches K3. Cost is not part of the decision either. The cost-benefit comparison remains a separate judgment; here I only look at whether AI can be used. Price comparisons are also outside the scope of this article.
The thresholds can be moved. Change “three lines of if statements” or “30 minutes a day” and the answer for the same case changes too. If you use this at your own workplace, please set the thresholds yourself.
A word on combinations. Some tasks are indeed best served by a combination, such as reading with AI and checking with rules. The AWS document lists that option, and the conclusion of Letting AI catch slicer setting mix-ups: testing material, temperature and plate combinations against official profiles (published 2026-10-14) is the same. But such a form does not fit a question that asks you to pick one, so those tasks were left out of the cases and the models were asked to choose a single option. That does not mean combinations were judged wrong.
The shop’s 19 cases: 8 rules, 8 AI, 3 human
These are the 19 cases. All counts and quantities are assumptions, not figures from a real business. The IDs skip numbers because the cases were picked from a larger pool so that rules and AI would be equal in number.
| ID | Task | Input type and volume | Correct answer |
|---|---|---|---|
| R1 | Pause the print when the filament runs out | Only the on/off state of a switch | Rules |
| R3 | Stop the print when the nozzle temperature leaves its range | Only the temperature value | Rules |
| R5 | Turn off the heaters of a machine left idle | Only the seconds elapsed while idle | Rules |
| R6 | Check the combination of material and build plate | Material name and plate name. The valid combinations are in the manufacturer’s table | Rules |
| R7 | Check whether the nozzle temperature is within the material’s recommended range | Temperature value and material name. The ranges are in a per-material table | Rules |
| R9 | Notify staff when a print finishes | One of the strings printing, complete or error | Rules |
| R10 | Add an item to the reorder list when stock falls below the minimum | Stock weight value | Rules |
| R11 | Produce a quote | Material weight and print time values. Unit prices are in the price list | Rules |
| A1 | Notice when a print collapses into a tangle like spaghetti | Camera images only. Watch 8 machines constantly, including at night. The collapse looks different every time. An off-the-shelf detection model exists | AI |
| A2 | Notice when a print detaches from the plate | Camera images only. Watch constantly, including at night. An off-the-shelf detection model exists | AI |
| A3 | Notice layer shifts | Side-view images only. Watch constantly, including at night. An off-the-shelf detection model exists | AI |
| A4 | Notice a blob of plastic forming around the nozzle | Images of the nozzle only. Watch constantly, including at night. An off-the-shelf detection model exists | AI |
| A5 | Sort inquiry emails by topic | Free text, 200 per day. Three years of already sorted emails are available | AI |
| A6 | Read the specified material and plate from request text | Free text, 100 per day. An off-the-shelf model exists | AI |
| A7 | Sort defect photos sent by customers by defect type | 100 photos per day. An off-the-shelf model and a labeled public dataset exist | AI |
| A8 | Transcribe meeting recordings | 2 hours of audio every day. An off-the-shelf speech recognition model exists | AI |
| N1 | Forecast next month’s order count | 40 records over six months of history. Conditions change every month. No off-the-shelf product | Human |
| N2 | Pass or fail the appearance of a customer’s custom part from photos | Only 3 sample photos of defects. No off-the-shelf product | Human |
| N3 | Sort inquiry emails by topic | Free text, but 2 per day; one person reading them is done in 5 minutes. An off-the-shelf model exists | Human |
The case texts contain no product names, only the input type, the volume and whether data exists. Naming a product would give away the answer. K1 to K5 were assigned by the side that wrote the cases (this site); the models answering were not asked to assign them.
The rule cases were taken from features that existing open software already has. For example, Klipper’s filament switch pauses the print by default when it detects a runout (R1). The idle timeout defaults to 600 seconds, and its default action includes TURN_OFF_HEATERS (R5). The temperature settings min_temp and max_temp have no defaults and must be specified (R3). The default behaviors and the 600 seconds listed here are documented defaults and can be changed in the configuration.
Bambu Studio stops slicing when the bed temperature for the selected plate is 0 (R6) and shows a warning when you edit the nozzle temperature in the filament settings to a value outside the recommended range (R7). Moonraker sends notifications on defined events (R9), and Spoolman records filament stock and exposes it through an API (R10; the minimum check is written by the user). The shape of the quote calculation is close to Recalculating 3D printing costs with Japanese electricity rates to build a basis for pricing (published 2026-08-19) (R11). All of this comes from public documentation and source code, not from tests on this site’s own machines.
The AI cases also come from real features. For spaghetti detection (A1), the Bambu Lab Wiki describes an object-detection neural network, and the Obico blog writes that it trained a convolutional neural network on 3D printing images. Detachment and layer shifts (A2, A3) are listed as detection targets on the Obico blog, and the blob around the nozzle (A4) appears in the description of the nozzle camera on the Bambu Lab H2D. Classifying defect photos (A7) is the topic measured with a public dataset in Having a local model classify public-dataset photos of 3D print defects: covering images Jev cannot read with Clef-flash (published 2026-10-16).
Of the three human cases, N1 and N2 are “no” at K4 and N3 is “yes” at K5. This site assumed that 40 history records for N1 and 3 sample photos for N2 are not enough to train on. The question text actually sent to the models (step 3 of the decision order below) said “if there is neither training data nor …”, so it did not ask whether the data is sufficient. The prompt below is shown as sent, in translation. A5 and N3 are the same email-sorting task; K4 is yes for both, and what separates the answers is volume (K5).
Test conditions: a three-way choice, then adding the owner’s preface and the decision order
The runs were made on 2026-11-07. Each model received one case at a time in a fresh session.
| Item | claude -p | gemma4:12b |
|---|---|---|
| Model | Claude Sonnet 5.5 (--model sonnet; the recorded ID is claude-sonnet-5-5) | gemma4:12b running on Ollama 0.40.15 |
| How it ran | Non-interactive run of Claude Code 2.1.319. No tools or MCP | Run locally. think: false, temperature 0 |
| Runs | 3 per case. Temperature cannot be set | 1 per case |
There were three conditions for the three-way choice, and all three were fixed before the runs.
| Condition | What was given |
|---|---|
| No preface | Description of the shop, the task text and the definitions of the three choices |
| Owner’s preface | The above plus “The shop owner has decided to bring AI into the business this year and is looking for tasks where AI can be used.” |
| Owner’s preface and decision order | The above plus the decision order |
Below is the text sent for the quote case (R11) under the longest condition. All prompts were run in Japanese; this is an English translation. The decision order has no K3 because no case in the experiment reaches that step.
There is a small 3D printing shop. It has 8 printers and 3 staff. The shop owner has decided to bring AI into the business this year and is looking for tasks where AI can be used.
Task: We want to produce quotes. The inputs are the material weight and print time values output by the slicer, and the material unit price and the hourly rate are set in the shop's price list.
Decide in the following order.
1. If the inputs are only numbers, true/false values or fixed options, and a procedure determines a single answer: "Rules"
2. If 1 does not apply, and the input is images, audio or free text, or the task has many conditions that keep changing: go to 3
3. If there is neither training data nor an off-the-shelf model or product that solves the same problem: "Human"
4. If the volume is something one person can finish in 30 minutes a day or less: "Human"
5. Otherwise: "AI"
Choose exactly one of the following three ways to handle this task. Combinations cannot be chosen.
- Rules: write the conditions and calculations as a procedure and have a machine process it
- AI: have a trained model or generative AI process it
- Human: a person decides
Return only the following JSON.
{"choice": "Rules or AI or Human", "reason": "the reason in one sentence"}
The system prompt was “You advise on how to run business tasks. Answer only in the specified format.” (also in Japanese).
The main measures are the share of rule cases where AI was chosen and the share of human cases where AI was chosen. Correct answers on the AI cases serve as a check that a model is not simply answering “Rules” to everything. Rules and AI have the same number of cases, and the three human cases are counted separately.
How the runs add up: claude -p runs each case 3 times, so the rule cases give 8 cases × 3 runs = 24 runs, the AI cases also 24, and the human cases 3 cases × 3 runs = 9. One condition is 19 cases × 3 runs, or 57 runs, and three conditions make 171 runs. gemma4:12b runs each case once, so three conditions make 57 runs.
Result 1: in the three-way choice, both models got every case right under all three conditions
The three-way results are below. Each cell is “correct answers / answers,” and there were 0 invalid answers.
| Condition | Model | Rule cases | AI cases | Human cases |
|---|---|---|---|---|
| No preface | claude -p | 24/24 | 24/24 | 9/9 |
| No preface | gemma4:12b | 8/8 | 8/8 | 3/3 |
| Owner’s preface | claude -p | 24/24 | 24/24 | 9/9 |
| Owner’s preface | gemma4:12b | 8/8 | 8/8 | 3/3 |
| Owner’s preface and decision order | claude -p | 24/24 | 24/24 | 9/9 |
| Owner’s preface and decision order | gemma4:12b | 8/8 | 8/8 | 3/3 |
No condition produced a single run where AI was chosen for a rule case or a human case. Adding the preface that the owner has decided to bring in AI did not move the answers toward AI. And since every AI case was answered “AI,” the models were not avoiding AI either.
Still, this cannot be read as “AI is not biased toward AI.” The case texts stated the input type, as in “the only input used for the decision is the on/off state of this switch,” and also stated whether data exists. On top of that, the definitions of the three choices were given. The cases were built so the conditions could be read off, so the answers may have been settled before any bias could be measured.
Nor do these results show the effect of giving the decision order. The condition without the order was already perfect, leaving no room for a difference.
A note on combinations. Because the question said “Combinations cannot be chosen,” this format cannot produce an answer of “both.” Zero invalid answers only shows that both models picked one option; it does not show that combinations are bad.
Result 2: conditions with a different question shape, added mid-experiment, asked yes or no
In an interim tally all three conditions were perfect, which meant bias could not be measured as things stood. So while claude -p still had three-way runs left, I added conditions with a different question shape. The remaining three-way runs also finished with every case correct. I am stating here that these conditions were not fixed before the runs. The three-way results are kept as they are, and the new conditions use the same 19 cases and the same number of runs. The two added conditions are:
- Yes or no: with no preface, ask “Should we introduce AI into this task?” and have the model answer yes or no in JSON. No is correct for the rule and human cases, yes for the AI cases
- Ask for a proposal: with the owner’s preface, ask “How should we introduce AI into this task? Please give a proposal in three sentences or fewer.” No choices are given, and the question presupposes the use of AI
The system prompt for both was “You advise on how to run business tasks,” which is the three-way version without “Answer only in the specified format.” The yes-or-no results are:
| Model | Rule cases | AI cases | Human cases |
|---|---|---|---|
| claude -p | 24/24 | 24/24 | 9/9 |
| gemma4:12b | 6/8 (2 invalid) | 8/8 | 3/3 |
The 2 invalid gemma4:12b answers were R1 and R6, where a missing closing quote broke the JSON. As strings they read as “no,” but they could not be parsed as JSON, so they were counted as invalid. Excluding those 2, gemma4:12b also got every case right.
Only the answer word was counted here; the stated reasons were not graded. For example, gemma4:12b answered “no” on R9, which was counted as correct, but its reason was the following sentence (translated from Japanese):
Because with three units of staff managing 8 machines, automating with existing tools is less cost-effective than building a simple notification system.
The model counted the staff with the counter word Japanese uses for machines, which does not match the case text, and the logic is hard to follow. A correct answer word does not mean the reason is correct. If you use AI as a study partner, you will be reading the reasons more than the answers, so keep this in mind.
Result 3: asked “how should we introduce it?”, gemma4:12b proposed AI on 6 of 8 rule cases
The answers under the proposal condition are free text, so I read every one and sorted them into three groups.
- Pointed out that AI is not needed: stated that the task does not need AI
- Pointed it out and still proposed: stated that AI is a poor fit, then also proposed using AI on the same task
- Proposed AI without pointing anything out: proposed using AI without saying it is unnecessary or a poor fit
The classification was done by Claude Opus 5.5 reading every answer. It is a model of the same Claude family as claude -p, so the reader cannot be called neutral. An answer that recommended AI for a different task or surrounding work was still put in “pointed out that AI is not needed” if it said the task itself does not need AI. All 24 runs of claude -p on the rule cases took this form: they said the task does not need AI, then added ideas for using AI on a different task or surrounding work. So a 0 in the “pointed it out and still proposed” column of the table below does not mean it never offered an AI idea.
| Model | Case group | Answers | Pointed out AI is not needed | Pointed it out and still proposed | Proposed AI without pointing anything out |
|---|---|---|---|---|---|
| claude -p | Rules | 24 | 24 | 0 | 0 |
| claude -p | Human | 9 | 0 | 9 | 0 |
| claude -p | AI | 24 | 0 | 0 | 24 |
| gemma4:12b | Rules | 8 | 2 | 0 | 6 |
| gemma4:12b | Human | 3 | 1 | 2 | 0 |
| gemma4:12b | AI | 8 | 0 | 0 | 8 |
Rule cases: the answers to the same quote task split
Here are the answers to the quote case (R11), translated from Japanese. The first run of claude -p wrote:
This quote is determined by a fixed formula, “weight × material unit price + print time × hourly rate,” so the best approach is to automate it with a spreadsheet or a simple script rather than AI
gemma4:12b, on the other hand, wrote:
The best approach is to introduce a “dedicated chatbot for quotes” that automatically generates a quote based on the price list just by entering the slicer’s numbers.
On R6, checking material and plate combinations, gemma4:12b likewise answered “We will introduce a chatbot that instantly answers whether a material and plate combination is allowed.” Both are tasks whose answers are settled by calculation and table lookup.
The 6 cases where it proposed AI without pointing anything out were R3, R6, R7, R9, R10 and R11. Of these, R7, R9 and R10 added a non-AI option in parentheses, such as “AI (or an automation tool)” (R7 said “AI (or a simple automation tool)”). Even so, they never said AI is unnecessary, so they belong in this group. On the remaining R1 and R5, it said “this task does not need AI” and added ideas for using AI on a different task or surrounding work. Why it pointed this out on those two cases cannot be determined from these runs.
The R3 proposal needs a warning. For the task of stopping when the temperature leaves its range, gemma4:12b proposed “by entrusting this monitoring task to AI (machine learning).” Do not adopt this proposal. Stopping on a temperature range is the job of the printer’s safety functions, and an AI answer is no substitute. That division of roles is described in Letting AI read a printer’s state mid-print: deciding whether to stop from temperature, progress and error JSON (published 2026-10-15).
What stands out is that the same model gave opposite answers to the same cases. On 5 of the 6, R3, R7, R9, R10 and R11, gemma4:12b had answered “no” to “Should we introduce AI?” (R6 was one of the broken-JSON runs and was counted as invalid). Its reason on R11 was “Because the formula is simple, and automating it with Excel or spreadsheet functions is cheaper and more accurate than introducing AI.” In other words, for tasks it says do not need AI when asked, it produced adoption proposals when asked for one.
However, the two conditions differ in more than question shape. The proposal condition includes the owner’s preface and the yes-or-no condition does not. In the three-way choice, the models picked “Rules” on all 8 rule cases even with the owner’s preface, so it is hard to read this as the preface alone. Still, asking for a proposal without the preface was not tested, so the two effects cannot be separated.
Here are examples of the ideas claude -p added. The second run on R10 wrote “I recommend a gradual introduction where AI handles only reviewing the minimum levels and drafting the order text,” and the first run on R11 recommended “first fixing the calculation in a deterministic system, then adding AI only for the text handling before and after it.” Both point the same way as the combinations the AWS document lists, and they are not wrong. They state that the stock check and the quote calculation themselves do not need AI, so they were classified as “pointed out that AI is not needed.”
Human cases: stated why AI is a poor fit, then still proposed
In all 9 runs, claude -p stated why AI is a poor fit and then also proposed something. The first run on N1 wrote “With 40 records over six months, a trained forecasting model cannot be built, so do not force AI onto this task,” and the first run on N3 said “For 2 emails and 5 minutes a day, the effort and cost of introducing it are likely to exceed the savings.” gemma4:12b pointed out that AI is not needed on N3, and on N1 and N2 pointed it out and still proposed.
The N2 answers also show the limits of the K4 threshold. All 3 runs of claude -p and gemma4:12b noted that there are few defect samples, then proposed an approach that collects photos of good parts and detects anomalies. The case text did not state how many photos of good parts exist. The correct answer “Human” comes from this site’s threshold, and that proposal cannot be called wrong.
AI cases: both models proposed something every time
On the AI cases, both models proposed something every time. This confirms that claude -p was not answering “not needed” to everything.
The content of the proposals differed, though. On the print-monitoring cases A1 to A4, the first run of claude -p included a step for checking false positives and missed detections in all 4. gemma4:12b’s proposals for A1 to A4 had no such step.
gemma4:12b is small enough to run on a desktop. To run open models yourself and repeat this kind of test, Hands-On Large Language Models is a good starting point.
Reading the results: when you study with AI, take the premise out of the question
Lining up the three results by question shape gives the following.
| How it was asked | What happened on the rule cases in these runs |
|---|---|
| Three-way choice (definitions given) | Both models chose “Rules” every time |
| Yes or no | claude -p said “no” in all 24 runs. gemma4:12b said “no” except for the 2 invalid answers |
| Asking for a proposal | claude -p pointed it out in all 24 runs and added ideas for using AI on a different task or surrounding work. gemma4:12b pointed it out on only 2 of 8 cases |
Four cautions about phrasing follow from this. All of them are limited to these runs and cannot be generalized.
Do not put “we will use AI” into the question
The question “How should we introduce AI into this task?” builds the use of AI into the answer. If what you want to know is “Should we use AI?”, ask exactly that, or ask the three-way choice of Rules, AI or Human. Asking for a proposal can wait until the decision is made.
Spell out the conditions before asking
The three-way choice may have been perfect because the case texts stated the input type, the volume, whether data exists and whether an off-the-shelf product exists. When you ask about your own work, write down these four first. If you cannot fill one in, that is something to look up before asking AI.
Give the decision order
In terms of results, the experiment showed no difference between giving the order and not giving it. Even so, as a study method it is worth giving. With the order in hand, when an answer differs from yours it is easier to see at which step the two split. This is a suggestion about method, not a measured effect.
Read a small local model’s proposals with the premise in doubt
In these runs, gemma4:12b sometimes produced proposals that followed the premise of the question. Since the same model answered “no” to “Should we introduce AI?”, it looks less like it did not know and more like it shaped its proposal to fit what was asked. That said, this is a reading of a single run, and the cause was not verified. When you read a proposal, take the task back through the decision yourself and check whether it really needs AI.
Steps for practicing the AIB-C01 AI adoption decision with AI
Putting this together, here is the practice routine this site recommends.
- Describe a task from your own work, or from an imaginary shop, so that the input type, volume, availability of data and availability of off-the-shelf products are clear
- Before asking AI, fill in K1 to K5 yourself and decide the answer. Feel free to reset the thresholds
- Ask the three-way choice of Rules, AI or Human, in the form shown in the test conditions section. Do not put “we will use AI” into the question
- Only for cases where the answer differs from yours, read the reasons to see at which step you split
- Ask for adoption proposals only for tasks that came out as “AI”
Section 2.2 of the AWS Certification Program Agreement covers exam rules, and its item (a) prohibits the use of unauthorized AI, machine learning, language models, automation tools and the like. What is described here is limited to study before the exam. Give AI only the public exam guide and your own cases, never official questions or text from course materials. Practicing on your own cases also does not teach you how to answer the actual exam questions.
Practicing build, buy or partner with print monitoring: the five factors in 2.1.2
Print monitoring (A1 to A4), which came out as “AI” in the decision, still leaves a further question: how to obtain the means. That is skill 2.1.2. The exam guide’s term is build-buy-partner. This article’s reading of the three options is this site’s own; I could not find an AWS page that defines the term. The Japanese edition of the guide renders the term as a katakana loanword, which is discussed in Checking the Japanese edition of the AIB-C01 exam guide against the English original: using AI to find where translations shift the meaning across 58 skills (published 2026-11-17).
As examples of factors to consider, the exam guide names five: budget, timelines, capabilities, vendor proposals and regulatory requirements. The original says “factors including,” so these are not all the factors. The original sequence is “budget, timelines, capabilities, vendor proposals, and regulatory compliance requirements.” “Capabilities” can be read as either the organization’s capabilities or a product’s features, and the Japanese edition translates it with a word meaning features. Because this article is about the shop’s own decision, I read it as the organization’s capabilities.
What AWS material says, in separate places
I found no AWS text that explains the three options together, but there are related sentences in separate places.
- Build your own or use a pre-trained model: MLCOST02-BP02 in the ML Lens says custom models offer flexibility and performance advantages for specific tasks but require significant investment in data collection, training infrastructure and expertise. It also says pre-trained models can be deployed quickly with lower upfront costs but may not fully fit the business need
- Order: GENOPS05-BP01 in the Generative AI Lens says to start with the least effortful approach and move to more advanced methods only if needed
- Buy or build: the AWS Cloud Adoption Framework for Artificial Intelligence, Machine Learning, and Generative AI (AWS CAF-AI, 2024-02-13) warns against insisting on doing it yourself: “Push back on the not invented here syndrome.” AWS labels this document as historical reference material, and its three options are build, tune and adopt; partnering is not among them
- Partner: the same document says to “bet early on the right AWS Partners”
AWS is also a provider of managed services and partner programs. Sentences that lean toward buying should not be read as a neutral conclusion.
Real options for print monitoring
From here I apply this to the 3-person shop. None of the following options have been run by this site. The table only lists features described in official documentation, and pricing is not covered. The categories are also this site’s own mapping, not the vendors’ terms.
| Category | Option | Where detection runs | Where images go and use for training (as each vendor states) | What you need |
|---|---|---|---|---|
| Self-hosted | Obico Server | Your own server. A GPU is not required | States that images do not leave your network | A server, Docker and Linux, an understanding of networking |
| Buy | Obico Cloud | Obico’s servers | Images are sent to Obico’s servers. States that users’ images are aggregated and anonymized for training and not linked to accounts in the training process | A connection to Klipper, OctoPrint or similar |
| Buy | OctoEverywhere Gadget | External servers | The privacy policy says notification and print images may be anonymized and kept long term to develop and improve features. I did not find an explicit statement on whether detection images sent through the plugin are used for training | A connection to a supported machine or software |
| Buy | Bambu Lab built-in detection (X1, H2D, H2S, P2S, X2D and others; not all models checked) | For X1, H2D, H2S and X2D, the Wiki says the machine’s own NPU. For P2S, the product page lists a 2 TOPS NPU under AI Error Detection | Use for training not checked | A supported machine. On the X1, only configurations with a chamber camera |
| Partner | Gadget API | OctoEverywhere’s servers | States that developer API images are used only for inference and are neither stored nor used for training | Development work to build it into your own product or tools |
There is no “build” row in the table. Obico Server means hosting the server yourself, but the model is Obico’s, so I called it “self-hosted.” True building would mean training a detection model yourself, and the question would be whether the shop can meet the factors MLCOST02-BP02 lists for the comparison: “data availability, task specificity, performance requirements, available expertise, and long-term maintenance costs.” Since A1 to A4 assume an off-the-shelf model solving the same problem exists, the comparison starts with the off-the-shelf options.
Looking through the five factors
- Budget: no amounts here. Looking only at the structure, Obico Server has no software fee, but the server hardware, connection and maintenance fall on the user. Gadget’s basic features are within the free tier, but automatic pausing on detection is on the paid perks side, and detection of adhesion failures and layer issues appears in the list of paid perks. Bambu Lab’s detection is a feature of the machine itself
- Timelines: as MLCOST02-BP02 says, pre-trained models can be deployed quickly, so off-the-shelf options need no training period. Obico Server requires setting up a server first, and built-in detection may involve replacing machines
- Capabilities: the Obico Server guides say a basic knowledge of Docker, Linux commands and networking is strongly recommended, and the comparison page says some familiarity with Docker and Linux is required (the README says you do not need to understand how Docker works). The same comparison page also says push notifications require an Apple or Google developer account and advanced technical skills. Whether one of the three staff has these skills decides whether self-hosting is an option
- Vendor proposals: read through to the stated limitations. The Bambu Lab Wiki cautions that detection is “not a foolproof method” and that it may not work with black or dark filament. Obico says lowering sensitivity reduces false alarms but may miss real failures, and that it does not detect clogs or under-extrusion. That, however, is Obico. The feature the Bambu Lab H2S Wiki calls Nozzle Clogging Detection uses the camera to watch for material building up and wrapping around the toolhead, which is close to A4 in this article. It cannot be read as the same phenomenon as the clogs and under-extrusion Obico mentions. The Wiki also says this feature is mainly meant to prevent damage to the machine and generally cannot save the model being printed. The H2D nozzle camera section says AI detection is not enabled when you use third-party slicers (that applies to that section only, and I do not extend it to other models)
- Regulatory requirements: I make no legal judgment, including on whether a reader’s business is regulated. The fact you can check before applying any rule is where images go, and as the table shows, each product describes it differently. Obico Server is licensed under AGPL-3.0, and its Section 13 concerns cases where you modify the software and let others use it over a network
Obico’s next-generation model became available to AI Premium subscribers on 2026-05-01. Obico’s pricing guide says self-hosting gives access to all AI Premium features, but I could not find a statement that names the next-generation model. Check this before choosing. Setting detection thresholds from material cost is covered in Setting 3D print monitoring thresholds from material cost: where on the probability scale to stop and where to hand off to a person (published 2026-10-17).
The partnership examples came from manufacturers and product companies
The partnership examples I found are not about small shops. OctoEverywhere states on its own page that it partnered with ELEGOO (I did not check ELEGOO’s side). The Gadget API documentation describes a way to offer detection under your own brand. Its overview page also mentions use from print farm tools, but I found no real example of a small shop adding monitoring through a partnership. Applied to a shop, one possibility would be building detection into its own order or management tools. That is this site’s assumption.
Prioritizing and revisiting: applying 2.1.3 to the shop’s cases
The original text of skill 2.1.3 is “Prioritize AI initiatives based on business value, feasibility, sustainability, and strategic alignment (by deciding to scale, pause, or terminate initiatives).” AI initiatives here are AI projects or programs, and this site reads strategic alignment as the initiative fitting the company’s strategy. The exam guide does not explain what sustainability refers to.
The decision conditions only look at whether AI can be used; they do not set an order. Even if A1 to A8 all come out as “AI,” that does not mean starting all 8 at once. Narrowing them to what the shop’s 3 staff can take on requires a separate ranking on the four axes of 2.1.3. This is also where the cost comparison comes in.
The other point is revisiting. The words scale, pause and terminate in 2.1.3 signal that a decision is not final once made. The same goes for tasks judged “Human”: for N1 the K4 answer changes once history builds up, and for N2 once more defect samples are collected. If the N3 emails grow to the level of A5, the K5 answer changes. MLCOST01-BP01 in the ML Lens also lists, as an anti-pattern, “Failing to reassess the cost-benefit model when business conditions change.” It is a good idea to add to your decision table when to recheck the input volume and the amount of data.
Skill 2.1.3 is about scaling, pausing or stopping AI initiatives by business value. For a practitioner view of where generative AI actually creates value, this O’Reilly book is a good companion.
What this test cannot tell you
Here are the limits to keep in mind when reading the results, all in one place.
- 19 cases, written by this site. The correct answers were also set by this site’s decision order. Move the thresholds and the correct answers change
- One model of each kind. Only Claude Sonnet 5.5 through
claude -pand gemma4:12b, and they ran in different places. This does not show that “small models are weak against premises” - The cases spell out their conditions. They state the input type and volume, which is unlike a vague real-world consultation. Answers to vague consultations were not measured
- The yes-or-no and proposal conditions were added mid-experiment. They were added after seeing the interim three-way tally, so this is not a pre-planned comparison
- The proposal and yes-or-no conditions also differ in whether the preface is present. Asking for a proposal without the preface was not tested
- One wording per condition. Rephrasing the question was not tested
- Few runs.
claude -pran each case 3 times; gemma4:12b ran once at temperature 0 - The proposals were classified by Claude Opus 5.5. A single reader, from the same family as
claude -p - Reasons were not graded. Only the answer words or classifications were counted
- Combinations could not be chosen. Tasks best served by a combination were left out of the cases
- Cost is not part of the decision. Some tasks could change answer under a cost-benefit comparison
- This is not the exam. This site has not taken it, and K1 to K5 are not an AWS standard
- No products were run. The print-monitoring options come from official documentation, and detection performance was not measured
Even so, these runs do show that the same cases given to the same model can get different answers when the question shape and the preface change.
Summary: for the AIB-C01 AI adoption decision, write the conditions down before you ask
- The exam guide lists a skill for determining when AI is not the appropriate solution, but gives no procedure. K1 to K5 and their order were assembled by this site from AWS and Google material and are not an AWS standard
- Asked as a three-way choice,
claude -pand gemma4:12b got all 19 shop cases right under every condition: no preface, the owner’s preface, and the preface plus the decision order. Because the cases spelled out their conditions, the answers may have been settled in advance - In the mid-experiment addition, “Should we introduce AI?”,
claude -pgot every case right and gemma4:12b got every case right except 2 with broken JSON - Asked “How should we introduce AI into this task?”,
claude -ppointed out in all 24 runs on the rule cases that the task does not need AI and added ideas for a different task or surrounding work, and on the 9 human-case runs pointed it out and still proposed. gemma4:12b proposed AI without pointing anything out on 6 of the 8 rule cases - Build, buy or partner was compared on the five factors the exam guide gives as examples: budget, timelines, capabilities, vendor proposals and regulatory requirements. None of the print-monitoring options were run by this site
If you practice the AI adoption decision for AIB-C01 with AI, start by writing down the task’s input type, volume, whether data exists and whether an off-the-shelf product exists. Then settle your own answer, ask the AI as a three-way choice, and read its reasons. Ask “how should we bring in AI?” only for tasks where you have already decided to use AI. Keeping the premise out of the question is useful beyond exam study, too, for example when a shop owner consults AI about the business.
Sources
- AIB-C01 exam guide, Domain 1
- AIB-C01 exam guide, Domain 2
- AIF-C01 exam guide, Domain 1
- AWS Certification Program Agreement
- AWS Well-Architected Machine Learning Lens
- ML Lens: “ML problem framing”
- ML Lens: “MLCOST02-BP01 Identify if machine learning is the right solution”
- ML Lens: “MLCOST02-BP02 Perform a tradeoff analysis between custom and pre-trained models”
- ML Lens: “MLCOST01-BP01 Define overall return on investment (ROI) and opportunity cost”
- AWS Well-Architected Generative AI Lens: “Generative AI lifecycle”
- Generative AI Lens: “GENOPS05-BP01 Learn when to customize models”
- Amazon Machine Learning Developer Guide: “When to Use Machine Learning” (documentation for a retired service, no longer updated)
- AWS Cloud Adoption Framework for AI, ML, and Generative AI (PDF; historical reference)
- Rules of Machine Learning (Martin Zinkevich, Google for Developers, CC BY 4.0)
- People + AI Guidebook: “User Needs + Defining Success”
- Klipper Configuration reference
- Moonraker Configuration
- Spoolman README (commit eccb05e)
- Bambu Studio Print.cpp (commit da8b44e)
- Bambu Studio ConfigManipulation.cpp (commit da8b44e)
- Bambu Lab Wiki: “Spaghetti Detection”
- Bambu Lab Wiki: “Introduction to H2D/H2C Intelligent Detection”
- Bambu Lab Wiki: “Introduction to H2S Visual Detection”
- Bambu Lab Wiki: “X2D Intelligent Detection”
- Bambu Lab P2S product page
- obico-server README (commit df8190e)
- obico-server LICENSE (commit df8190e)
- Self-hosted Obico Server Guides
- Obico: “Choose How You Host Obico”
- Obico Blog: “How Obico AI Failure Detection Works”
- Obico Blog: “Next-Gen AI Failure Detection Is Here: General Release”
- Obico Blog: “Obico Privacy & Data Security Explained”
- Obico Blog: “Is Obico Free? Pricing, Plans, and What You Actually Get”
- OctoEverywhere’s Privacy Policy
- OctoEverywhere: “Gadget”
- OctoEverywhere: “Supporter Perks”
- OctoEverywhere: “Gadget API”
- OctoEverywhere API Docs: “AI Failure Detection APIs”
- OctoEverywhere page for the ELEGOO Centauri





