Incoming Inspection Checklist: Build One Per Part With AI

To build an incoming inspection checklist for a purchased part, list every characteristic on the drawing and specification, classify each as critical, major or minor, choose the gauge for each, set a sample size per delivery, and write the accept or reject rule. AI can draft that table from the drawing; an engineer must verify the critical dimensions before the checklist is used.

The point of a per-part checklist is that the person on the goods-in bench no longer has to interpret a drawing at 4:45 on a Friday. They follow one page: measure these six features, with these gauges, on this many pieces, and here is what to do with a failure. Writing that page for every part by hand is what stops most small manufacturers doing it. AI removes most of the typing, but it makes characteristic-level errors that look convincing, so the method below puts the engineer's sign-off exactly where a misread dimension would hurt.

Follow me on Instagram@sagnikteaches

What the checklist has to contain

One page per part number and revision. Each line is a characteristic, and every line answers the same seven questions. If a line cannot answer one, the checklist is not ready to use.

Connect on LinkedInSagnik Bhattacharya

1. Which characteristic? A dimension, a surface finish, a thread, a material, a marking, a visual condition, a certificate.
2. What is the requirement? The nominal value and tolerance, or the written specification clause, copied exactly.
3. How serious is a failure? Critical (safety or function fails, or the part cannot be assembled), major (likely to cause a failure or rework), minor (cosmetic or low risk).
4. How is it measured? The named gauge or instrument, and its resolution.
5. How many pieces? The sample size for this delivery.
6. What counts as a pass? The accept and reject numbers.
7. What happens on a fail? Quarantine, contact, containment, return, concession.

Two further items sit at the top of the page, not in the table: the part number with its drawing revision, and the date the checklist was approved and by whom. A checklist built from revision B of a drawing and used on revision C is a common and silent failure.

Subscribe on YouTube@sagnikteaches

The worked part: a machined aluminium bracket

This example is illustrative: a small manufacturer buys a machined bracket, part B-4410, which holds a sensor to a frame. The drawing (revision C) shows these requirements, which the engineer has transcribed into plain text:

PART B-4410 REV C  Bracket, sensor mount
Material: aluminium alloy bar, 6000 series, as per spec sheet MS-12
Overall length: 84.0 +/- 0.2 mm
Overall width: 32.0 +/- 0.2 mm
Thickness: 6.0 +/- 0.1 mm
Mounting holes (2): diameter 5.5 +0.1/-0.0 mm, spacing 60.0 +/- 0.1 mm
Sensor bore: diameter 12.00 +0.05/-0.00 mm
Thread: M4 x 0.7 tapped hole, 8 mm deep, 6H
Surface: anodised, black, thickness 10-15 micrometres
Edges: deburred, break edges 0.3 mm max
Marking: part number etched on rear face
General note: all dimensions in mm, ISO 2768-m unless otherwise stated

The task for AI is to turn that into a checklist, keeping the figures exactly as written and asking about anything uncertain. Here is a prompt that does that:

Build an incoming inspection checklist from the part description below. Use a table with these columns: Characteristic, Requirement (exactly as written), Class (Critical/Major/Minor), Gauge or method, Sample size, Accept/Reject, On failure.
Rules:
- Copy every figure exactly; do not round or convert units.
- Classification: mark as Critical anything that affects fit to the sensor, mounting, or safety; Major anything that affects assembly or function; Minor cosmetic items. Show your reasoning in one line for each Critical.
- Choose a gauge for each characteristic and state the resolution it needs. The gauge's resolution should be no coarser than one-tenth of the tolerance band. If a tolerance is tighter than a handheld gauge can reasonably check, say so and suggest the right method.
- Do not invent dimensions or specifications. Where something is not stated, list it under "MISSING FROM THE DRAWING".
- Sample sizes: leave as "TBD per sampling rule" for now.

PART DESCRIPTION: [paste]

What a good draft looks like, and what to fix

A typical draft (illustrative) produces a table something like this. Read it with the drawing next to you, because the corrections are the point.

Now the engineer's corrections, which are the reason the checklist is not allowed to go live straight from the AI:

• Thickness class. AI called it Minor. The sensor's clamp relies on a 6.0 mm slot, so with a tolerance of 0.1 mm the engineer upgrades it to Major.
• The hole spacing method. A calliper on hole edges measures edge-to-edge, not centre-to-centre, and the AI's suggestion adds a hidden error equal to the hole diameter tolerance. The engineer specifies a fixture gauge with two pins, or measures with pins in the holes. The tolerance of 0.1 mm is also tight for a calliper with 0.02 mm accuracy, so the method matters.
• The material line. The AI left material out, because the drawing says "as per spec sheet MS-12". The engineer adds a line: certificate of conformity for the bar, checked against the sheet, on every lot.
• "General note ISO 2768-m". The AI wrote it in a note but did not apply it to unstated tolerances such as the 0.3 mm break. The engineer decides which dimensions it covers.

An engineer who finds those four in ten minutes has saved more time than the draft cost, and has also learned what to put in the next prompt. The gauge rule in the prompt ("resolution no coarser than one tenth of the tolerance") is a commonly used rule of thumb; your own quality procedure may set a stricter one.

Choosing the sample size

Sample size is where small firms either over-inspect (every piece, every time) or guess. A defensible approach is to write a rule once and let the checklist refer to it, rather than choose a number per delivery.

Published sampling standards, ISO 2859-1 and its close counterpart ANSI/ASQ Z1.4, give a sample size by lot size and an accept and reject number for each quality level you set. As an orientation, published reproductions of the ISO 2859-1 table give a sample of 50 (code letter H) for a lot of 281 to 500 pieces at general inspection level II. You choose the quality level (called the acceptable quality level) per class of defect, and read the accept and reject numbers from the standard's tables. Because those tables are copyrighted, buy or consult the standard itself rather than asking AI to reproduce it from memory, since a language model can easily misstate a table value.

A simple house rule many small manufacturers use instead, for illustration:

Those numbers are an example of the shape, not a recommended plan, and your engineer or quality adviser should set the real ones. The principle that matters is the ladder: critical features get the heaviest sampling and a zero-defect rule, because one failure of a safety feature is one too many. A check on the bracket's 12 mm bore, which the sensor depends on, therefore uses 100% go/no-go on the first delivery, then moves to a sample when three consecutive lots pass.

Tightened, normal and reduced checks

Write the switching rule into the checklist, so the person on the bench does not have to remember history. The sampling standards have their own, more detailed switching rules; a simpler house arrangement for illustration: start every new part or supplier on tightened (larger sample), move to normal after three clean deliveries, and move to reduced after ten, falling back to tightened after any critical failure or two major failures in a row. Ask AI to put the rule into a tracking sheet with columns for date, lot, result and current level, with a formula that suggests the next level. Check the formula on invented test rows before trusting it, since a sheet that quietly keeps a supplier on reduced after a failure defeats the purpose.

Certificates and paperwork that arrive with the parts

Many incoming checks are about documents, not dimensions: a material certificate, a heat-treatment record, a coating report, a certificate of conformity. This is a good fit for AI, because the job is matching text and numbers across pages. A prompt that works: "Compare this certificate with the purchase order and the specification sheet. List: material grade stated versus ordered, heat or lot number, quantity, date, and signature or stamp present. Flag every mismatch and every field you cannot read."

For the bracket, the certificate says alloy "6061" while the specification sheet MS-12 allows two named grades (illustrative). The AI flags that 6061 appears on the allowed list, notes that the lot number on the certificate ends in 78 while the delivery note shows 87, and says one line is blurred. A person reads the original: the lot number is a transposition on the delivery note, which the supplier corrects, and the material is accepted. Without the cross-check the transposed number would have broken traceability months later, when a customer asked which batch a bracket came from.

Record the certificate result on the same page as the dimensional results, with the lot number, so one sheet tells the whole story of a delivery.

Handling failures

The "on failure" column is the part that most checklists leave blank, and bench staff then improvise. Fill it in with a small number of standard actions:

• Critical fail: quarantine the whole lot, label it, stop production use, and inform the engineer and purchasing the same day. Request the supplier's corrective action in writing.
• Major fail beyond the accept number: quarantine, ask the engineer whether sorting 100% is worth it, and decide between return, rework at supplier cost, or a documented concession.
• Minor fail: record it, accept if within the sampling rule, and tell purchasing for the scorecard.

A concession, accepting a part that does not meet the drawing, must be signed by someone with authority and recorded. AI can draft the notification to the supplier and the non-conformance record from the inspection result; the decision belongs to a named person. Here is a prompt and a realistic output (illustrative):

Write a short non-conformance notice to a supplier. Facts: part B-4410 rev C, lot 2214, 120 pieces received, 20 inspected for the sensor bore (12.00 +0.05/-0.00 mm) with a go/no-go gauge, 3 failed the go end (bore too small). Action taken: lot quarantined. Ask for: cause, containment, corrective action, and replacement or sorting plan by [date]. Factual, no blame, under 150 words.

A usable result states the numbers, says the lot is quarantined, and asks four questions. What a person adds: the purchase order number, the photograph of the gauge result, and an instruction on whether the supplier collects or the firm returns the parts. Because the notice goes outside the company, someone reads it before it is sent.

Four other parts, four different problems

A rubber gasket. There are almost no dimensions that matter except the cross-section and the hardness. The checklist is mostly a certificate check (material and hardness from the supplier), a shore-hardness test on three pieces, and a visual check for flash. AI is useful here for drafting the certificate-checking lines, but the engineer sets the hardness tolerance.

A printed circuit board assembly. Dimensional checks give way to a functional test and a visual standard for solder joints. The AI should not be asked to invent acceptance criteria for solder; point it at your adopted workmanship standard and let it extract the lines, and have the engineer approve the extraction.

A cut-to-length steel tube. Length and straightness matter, wall thickness matters, and a mill certificate matters. The checklist samples length on 20% of the lot, wall thickness at both ends of five pieces, and requires the heat number on the certificate to match the number stamped on the bundle. The AI check is a cross-reference of numbers between the certificate and delivery note.

A packaging carton. The critical feature is the internal size, because a carton 3 mm short means products do not fit. The checklist measures three cartons per pallet with a steel rule, and the engineer sets a "fit test" with a real product on the first lot.

The revision mismatch that every part passed

Picture a small firm (an illustration) that built its bracket checklist from revision B of the drawing, where the sensor bore was 12.00 +0.08/-0.00 mm. Engineering then released revision C, which tightened the upper limit to +0.05 mm because the new sensor fits more closely, but neither the checklist nor the purchase order was updated, so the supplier kept machining to revision B. The bench used the checklist faithfully, bores of 12.06 and 12.07 mm passed, and the assembly team found the sensors loose in the bore two weeks later. Nothing was wrong with the AI, the gauge or the person: the checklist did not say which revision it covered, and nobody had a rule to update it when the drawing changed. The fix was a header line (part number, revision, date, approver) and a monthly step where purchasing lists every part whose drawing changed, so the checklist is rebuilt the same day.

A second pattern is a checklist so long that the operator skips lines. If a table has more than about twelve rows, cut the minor items to a rotating sample, or split it into two pages, one for receiving and one for first-article approval at the start of production.

A repeatable workflow, start to finish

1. Collect the inputs. The current drawing, any specification sheets it cites, and the purchase order terms. 10 minutes.
2. Transcribe the drawing's requirements into plain text. By the engineer, or by AI reading the drawing with the engineer checking every figure. 15 to 30 minutes per part.
3. Run the checklist prompt. 5 minutes.
4. Engineer review. Check every critical dimension against the drawing, correct classifications, replace any gauge that cannot do the job. 15 minutes.
5. Add the sampling rule and the failure actions. From your standing documents. 5 minutes.
6. Trial on a real delivery. Have the bench operator use it and mark every point of confusion.
7. Approve and file. Part number, revision, date, approver, and a link from the purchase order.

Once the pattern is set, a new part takes well under an hour instead of a morning, and the first-draft quality improves as you add rules to the prompt. Keep the prompt in a shared document, and add a line each time review finds a new kind of error.

Where this connects to the rest of your quality system

The checklist is one end of a loop. The purchase order should state the drawing revision and the inspection requirement, which the purchase order automation tutorial can build in. The result of each inspection becomes data for a supplier scorecard, so that late deliveries and failed lots shape who gets the next order. And if volumes grow to the point that bench checks are the bottleneck, camera-based inspection is the next question, though it suits repeatable visual checks and not dimensions the camera cannot see. For the wider picture of where a small manufacturer starts, see AI quality control for small manufacturers.

Keeping drawings and specifications safe

Drawings are often confidential and sometimes covered by a customer's confidentiality agreement or export controls. Before pasting one into a chat tool, check what your contracts allow. Business plans of the major assistants do not train on your content by default, according to their own pages, but contractual limits apply regardless. Where in doubt, transcribe the requirements into text and leave out the drawing image, title block and any customer identification, which is usually all this task needs.

Proving the checklist works

Judge the checklist on outcomes, not tidiness. After a month, look for: the number of non-conformances caught at goods-in compared with those found later in assembly or by customers (a rising catch at the door is the good direction); the time per delivery on the bench (it should be steady or falling, not growing); disagreements between operators using the same checklist (any disagreement is a line that needs rewording); and the number of times the engineer had to correct an AI-built checklist (this should shrink as the prompt improves). If a defect reaches assembly that the checklist should have caught, trace it to a line, classify the cause as a missing characteristic, a poor gauge or a sampling gap, and fix the template, not just that checklist. The goal is a set of single pages that someone new could pick up on Monday and use correctly, with an engineer's initials on every one.

Questions from small manufacturers about incoming inspection

Can AI read an engineering drawing accurately?

It can list the dimensions, tolerances and notes it sees, but it misreads small text, stacked tolerances, datum references and hidden views often enough that every extracted figure needs checking against the drawing. Treat its list as a draft for an engineer, and never as the drawing's authority.

How many parts should I inspect from each delivery?

It depends on lot size, how critical the feature is and the supplier's record. Published sampling tables such as ISO 2859-1 and its close counterpart ANSI/ASQ Z1.4 give sample sizes by lot size and quality level. For safety-critical features many firms inspect every piece or set a zero-defect rule.

Do I inspect every characteristic on every delivery?

No. Inspect the characteristics that matter to fit, function and safety, and rotate the minor ones. A supplier with a clean record can move to reduced checks, and one with a recent failure moves to tighter ones. Write those rules into the checklist so the operator does not have to decide.

What if the supplier sends its own inspection report?

Use it as evidence, not a replacement, until the supplier has earned trust. Verify the critical dimensions yourself on the first few deliveries and compare them with the report. If they agree consistently, you can reduce your own checks on the features the report covers.

Further reads

• Importer Product Compliance Checklist: AI Flags Missing Documents — The paperwork side for imported parts and products.
• How to Check AI Is Doing Good Work, Not Just Fast Work — How to check AI's work is good, not just fast.
• Can AI Read Delivery Notes and Update Stock Automatically? — Receive goods into stock from delivery notes.
• How to Match Supplier Invoices to Purchase Orders Automatically — Hold payment for lots that failed inspection.
• How to Compare Supplier Quotes Side by Side With AI — Weigh inspection cost when comparing suppliers.
• Packaging Specification Sheet: AI Drafts Specs Buyers Can Order From — A field-by-field packaging specification sheet, drafted with AI from your product data, with a filled example and a pre-release checklist.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: published sampling-plan standards (ISO 2859-1 and ANSI/ASQ Z1.4) as general references, and the structure of standard first-article and receiving inspection practice.