It is the Monday after a long weekend and the quality desk has a problem that looks like an avalanche. Over three shifts, the line flagged 312 suspect parts across four defect types: a cosmetic scratch on the housing, a dimensional miss on a bore, a contamination spot, and a handful of "operator unsure" holds nobody wants to touch. Priya, the quality engineer, has a customer shipment leaving at noon and a containment decision to make before then. The old way is to read every reject tag by hand, sort them into piles, and hope she gets the urgent ones first. By the time she finishes sorting, the morning is gone and the disposition work has not even started. This is exactly where AI earns its keep on the floor: not by deciding what to do with each part, but by reading 312 messy tags in two minutes and handing Priya a clean, sorted, summarized stack so she can spend her morning on the decisions only she can make. The trap, and the whole point of this lesson, is the line between those two jobs. AI can triage. AI cannot dispose. The customer audits the disposition, not the prompt, and the day Priya forgets that is the day "the model said scrap it" becomes the worst sentence in her containment report.

Triage Versus Disposition: The Line That Matters

The most important thing in this entire lesson is a definition, so we will go slow on it. Triage is sorting and summarizing: reading the reject tags, grouping the 312 parts by defect type and severity, pulling out the urgent ones, and writing a plain-language summary of what came in. Disposition is the decision: use as-is, rework, scrap, return to supplier, or deviate. Triage organizes the problem. Disposition resolves it, and disposition is a quality record the customer can audit, that carries legal and contractual weight, and that under standards like IATF 16949 (the automotive quality management standard) and AS9100 (the aerospace equivalent) must be made and signed by an authorized person.

AI is genuinely excellent at triage. It can read unstructured operator notes, normalize the spelling and the shorthand, classify a part into a defect category, estimate severity from the description, flag the ones that mention a customer-critical feature, and produce a ranked queue in seconds. The playbook reports that 47 percent of manufacturers now use AI in quality, up from 33 percent the prior year, and triage is one of the reasons: it turns hours of sorting into minutes. For Priya facing 312 parts, that is the difference between starting disposition at 10 a.m. and starting it at 7:15.

AI is genuinely dangerous at disposition, and not because it is incapable of producing a recommendation. It is dangerous precisely because it will produce one confidently. Ask a model "should I scrap this part" and it will answer, because answering is what it does. But that answer is untethered from the customer's contract, the deviation history, the cost of scrap versus rework, the engineering judgment about whether the bore is salvageable, and the question of who is authorized to sign. A disposition is not a text-generation problem. It is an accountable engineering decision. The model can draft the words of a disposition. It cannot make the disposition, and a quality engineer who lets it is signing a record she did not actually make.

It helps to name why the model cannot cross this line even in principle. The model has no access to the cost of scrapping a part versus reworking it on your line this week, no view of how many deviations the customer has already accepted this quarter before they start asking hard questions, and no standing to commit your plant to a contractual representation that a part conforms. Those are not facts in a note it can read; they are judgments and authorities that live with the human and the quality system. A disposition that uses-as-is a nonconforming part is, in effect, a signed statement to the customer that the part is fit for its purpose. No tool that predicts text can make that statement on your behalf, which is exactly why the standards require a named, authorized signer.

AI triages the parts. The human disposes of them. The customer audits the disposition, not the prompt.

What Good AI Triage Actually Looks Like

Let us build Priya's morning the right way, step by step, because the value is real and worth doing well. She exports the 312 reject records, each with an operator note, a defect code if one was entered, a station, a timestamp, and a part number. The notes are exactly as messy as you would expect from three shifts and a thin crew: "scrtch on top corner," "bore looks small didnt gauge," "black spot near weld," "?? holding for QE."

Her triage prompt does five things, and notice that none of them is a decision. First, classify each record into a defect family (cosmetic, dimensional, contamination, unknown) and say which words in the note drove the classification. Second, flag any record whose note mentions a customer-critical characteristic or a safety feature. Third, group the records and give counts per family. Fourth, rank the families by likely urgency, stating the reasoning. Fifth, and this is the load-bearing instruction, for any record that is ambiguous or that lacks the data to classify, mark it "needs human review" rather than guessing. That last rule is what keeps the triage honest. The "?? holding for QE" parts land in a review pile, not in a falsely confident bucket.

Two minutes later Priya has a structured summary: 180 cosmetic scratches concentrated on second shift at station 4, 70 dimensional holds on the bore, 22 contamination spots near the weld, and 40 flagged "needs human review," of which 6 mention the customer-critical sealing surface. That last group is where she starts. The AI did not tell her the sealing-surface parts were the priority because it knows quality. It surfaced them because she told it to flag customer-critical mentions, and she verified the flags by reading those six tags herself. The triage compressed three hours of sorting into the time it took to get coffee, and every part of it is checkable.

Notice a second, quieter benefit hiding in that summary. The triage did not just sort the parts; it revealed a pattern a tired engineer reading tags one at a time might have missed. One hundred eighty cosmetic scratches concentrated on second shift at a single station is not 180 random defects. It is a signal pointing at a likely upstream cause: a fixture, a handling step, or a tool on station 4 on that shift. Triage that counts and groups turns a pile of individual rejects into a Pareto (the ranked bar chart of defect drivers) that hands the continuous-improvement team its tallest bar for free. The disposition still belongs to the human, but the triage just did half the root-cause team's morning, and it did it as a byproduct of sorting the avalanche.

Triage is a hypothesis, not a verdict

Here is the discipline that separates a professional from someone who gets burned. Treat the AI's triage as a set of hypotheses to verify, not a set of facts to act on. The model said 180 parts are cosmetic scratches. Before Priya treats that as true, she spot-checks. She pulls ten of the 180 and reads the notes against the classification. If the model lumped a hairline crack (a structural defect) in with cosmetic scratches because the operator wrote "small crack/scratch," she catches it on the spot-check and corrects the rule. The cost of the spot-check is ten minutes. The cost of dispositioning a structural crack as a cosmetic scratch and shipping it is a customer containment, which the playbook notes can cost more than a month of the entire training program. Ten minutes against a containment is not a close call.

The Disposition Draft and Its Five Traps

AI can help with disposition in exactly one safe way: it can draft the document once the human has made the decision. Priya decides, based on the drawing, the customer contract, and her engineering judgment, that the 180 cosmetic scratches are use-as-is under the existing cosmetic acceptance criteria, the 70 bore holds go to 100 percent regauge and rework, the 22 contamination parts are scrap, and the 6 sealing-surface parts are held pending engineering review. Now AI is useful again: "Draft the disposition records for these four groups in our standard format, citing the acceptance criteria I specify for the cosmetic group, and leave the authorization signature and date blank for me to complete." That is a drafting task, and the model is good at it. But even here, five traps wait.

Trap one: the invented acceptance criterion. Ask the model to justify a use-as-is and it may produce a plausible-sounding cosmetic standard ("scratches under 0.5 mm in non-cosmetic zones are acceptable") that is not your customer's actual criterion. The playbook is explicit that invented specs and fabricated procedures are a primary AI failure mode. Every acceptance criterion in a disposition draft must be traced to the real drawing or customer spec, not to the model's general sense of what a cosmetic standard usually says. If you cannot quote the source line, the criterion does not go in the record.

Trap two: the confidently wrong defect family. If the triage misclassified a defect and the disposition draft inherits that classification, the error propagates into a signed record. A contamination spot mislabeled as cosmetic could be dispositioned use-as-is when it should be scrap. The defense is that the human verifies the classification at disposition time, not just at triage time, especially for anything heading to use-as-is, which is the riskiest disposition because the part ships.

Trap three: the missing customer notification. Many customer contracts require notification or approval before a deviation or a use-as-is on a nonconforming part. An AI draft focused on the internal disposition will happily omit the contractual notification step because nothing in the prompt told it the contract exists. The human owns knowing that the contract requires a customer waiver, and the AI draft is incomplete until that step is in it.

Trap four: the blended batch. The model may draft a single disposition covering parts that should be split. If 65 of the 70 bore holds are reworkable but 5 are out of tolerance beyond rework, a blanket "rework" disposition scraps nothing and ships five bad parts after rework that cannot save them. The human segments the batch; the AI drafts each segment once the human defines it.

Trap five: the absent signature treated as done. The most dangerous trap is the quietest. The AI produces a complete-looking disposition record, and under time pressure it gets filed as if the decision were made, when in fact no authorized person ever reviewed and signed it. A disposition is not made until an authorized human signs it. The draft is not the decision. Keep the signature block empty in every draft so that an unsigned record is visibly unfinished, never mistakable for a closed one.

Keeping the Decision Human and Audit Ready

The reason all of this matters is the cardinal rule of this entire program: the customer audits you, not the vendor. When a customer quality auditor sits down with Priya and pulls the disposition records for last Monday's 312 parts, the auditor is not going to accept "our AI classified them." The auditor will ask who decided the use-as-is, on what acceptance criterion, traced to which drawing revision, with what customer notification, signed by whom, on what date. Every one of those answers must be human and documented. "The model flagged it" is never a sufficient answer to an auditor. The accountability for an AI-touched quality decision stays with the plant and the human who signs the record.

This is why the audit trail has to capture both halves of the work and keep them distinct. The triage record shows what the AI did: the classifications, the flags, the counts, the source notes it read. The disposition record shows what the human decided: the disposition per group, the acceptance criterion with its source citation, the customer notification status, and the authorized signature and date. An auditor reading both can see exactly where the machine helped and where the human took responsibility. That separation is not bureaucracy. It is the structure that lets a plant use AI for speed in quality without ever letting AI make a quality decision it is not allowed to make.

Consider the cost math one more time, because it justifies the discipline. Priya's plant runs roughly 1,200 parts per shift on this line. A single defect escape last quarter shipped 4,000 parts before detection and became a containment. Against that exposure, AI triage saved her about three hours on Monday and let her start the real work before 8 a.m. The verification steps, the spot-checks, the source citations, and the human signatures added maybe forty minutes across the morning. Three hours saved, forty minutes of insurance spent, and not one quality decision delegated to a tool that cannot be held accountable. That is the trade the disciplined quality engineer makes, and it is overwhelmingly in her favor.

The skeptic's checklist before any disposition is signed

Before Priya signs anything, she runs a short checklist that turns the principles above into a habit. Is every acceptance criterion in the draft traced to a real, quoted source on the current drawing or customer spec? Has the defect classification been verified by a human for every part heading to use-as-is? Is the customer notification or waiver status correct and documented where the contract requires it? Has the batch been segmented so no out-of-tolerance part rides along on a blanket rework? Is the signature block empty until an authorized person actually signs? If any answer is no, the record is not ready, no matter how clean the AI draft looks. The clean draft is the temptation. The checklist is the control.

Building the Habit on a Thinning Crew

The deeper reason this workflow matters is the same talent cliff that drives the whole program. With 85 percent of manufacturers saying staffing shortages are hurting product quality and the most experienced inspectors retiring, the quality desk is thinner and greener than it has ever been. A green quality engineer facing 312 parts and a noon shipment is exactly the person most tempted to let the confident AI recommendation stand, because it relieves the pressure. That temptation is the risk. The same shortage that makes AI triage valuable makes the triage-versus-disposition discipline essential, because the person most likely to be in the chair is the one least likely to have the twenty years of judgment that would make them squint at a wrong recommendation.

So the habit you are building is not just personal. It is the thing that lets a plant put a less experienced engineer in the quality seat and still produce audit-grade dispositions, because the workflow itself enforces the line. AI does the reading, the sorting, the summarizing, and the drafting, all of which a green engineer can supervise. The human does the deciding and the signing, with a checklist that makes the verification explicit rather than relying on instinct the new engineer does not yet have. The captured discipline, written into the workflow, is what stands in for the experience that walked out the door when the veteran retired.

There is one more reason to keep the line sharp, and it is about trust. An operator or engineer who gets burned once by acting on a confident but wrong AI recommendation will stop trusting the tool entirely, and then the plant loses the real triage value too. The playbook makes this point about vision systems and false rejects: an operator burned by a false alarm will disable the green light. The same dynamic applies at the quality desk. If you let AI overreach into disposition and it produces a bad signed record, the lesson the team learns is "do not trust the AI," and they throw out the genuine speed benefit along with the misuse. Keeping AI firmly in triage, where it is excellent, is what preserves trust in AI for the long run. Used in its lane, it earns its place. Pushed out of its lane, it loses it for everyone.

Key Takeaways

• Triage and disposition are two different jobs. Triage is sorting and summarizing, which AI does excellently. Disposition is the accountable decision (use-as-is, rework, scrap, return, deviate), which a human must make and sign. AI triages; the human disposes.
• Good AI triage classifies, flags customer-critical and safety mentions, counts, ranks, and explicitly marks ambiguous records "needs human review" instead of guessing. It turns three hours of sorting into minutes, which matters most on a thin, green crew.
• Treat triage output as hypotheses to verify, not facts to act on. A ten-minute spot-check that catches a structural crack misfiled as a cosmetic scratch is trivial against a containment that can cost more than a month of the training program.
• AI may draft a disposition record only after the human has decided. Watch the five traps: invented acceptance criteria, inherited misclassification, missing customer notification, blended batches that should be split, and an unsigned draft mistaken for a closed decision.
• Every acceptance criterion in a disposition must trace to a real, quoted source on the current drawing or customer spec. Invented specs are a primary AI failure mode, and a use-as-is is the riskiest disposition because the part ships.
• The customer audits the disposition, not the prompt. Keep the triage record (what the AI did) and the disposition record (what the human decided, with source citations and an authorized signature) distinct, so an auditor can see exactly where the human took responsibility.
• Run the skeptic's checklist before signing: sourced criteria, verified classification for use-as-is, correct customer notification, segmented batches, empty signature block until an authorized person signs. The clean AI draft is the temptation; the checklist is the control.
• Keeping AI in its lane preserves trust. Let it overreach into disposition and produce one bad signed record, and the team learns to distrust it and throws out the real triage value. Used for triage, where it is excellent, AI earns its place on a thinning quality desk.