First Article Inspection Report: AI Formats the FAI, You Measure

Let AI do the formatting and let your inspector do the measuring. A person balloons the drawing, AI turns the balloon list into an AS9102-style table with requirements and limits, the inspector enters every measured value from calibrated gauges, and a spreadsheet formula, not AI, decides pass or fail. A qualified person reviews and signs the report.

The line between the two jobs is the point of the whole method. A first article inspection report exists to prove that someone measured the first part and that it matches the drawing. If AI fills any result, even one that "looks right", the report stops proving that. So the working rule is simple: AI never sees, writes or suggests a measured value. It sees the drawing's requirements, builds the structure, and checks the inspector's entries afterwards for the things a tired person misses, such as a value outside limits, a blank cell or a characteristic missing from the sequence.

Follow me on Instagram@sagnikteaches

What the report contains, and who touches each part

First article inspection reports under AS9102 use three forms, and many customers use the same layout or a close variant. Check which revision and fields your customer requires, since their quality clause overrides anything general. The common structure:

Connect on LinkedInSagnik Bhattacharya

Form 3 is where the work lives, and where AI's help is largest. Most of this tutorial is about it.

Subscribe on YouTube@sagnikteaches

Step 1: Balloon the drawing, with a person in charge

Ballooning means putting a numbered circle on every dimension, tolerance, note and specification on the drawing, so that each requirement has one number that follows it into the report. It is dull and it is where many reports go wrong, because anything not ballooned is not inspected.

AI can speed up the list but not take over. If your drawing is a clean PDF and the customer permits it (see below), ask the model for a list of every dimension and note it can find. Treat that list as a checklist for the person ballooning, not as the balloon list itself. Models skip small text, miss tolerances given in a general block, merge two dimensions on the same line and misread symbols such as a position tolerance with datum references. After ballooning, the person counts the balloons and compares the count with the table. Some customers accept drawing sheet and zone references in place of balloons, but many prefer a ballooned copy of the drawing in the FAI package, so ask before choosing.

A rule for ballooning that saves a re-do

Give every requirement its own number, even when it repeats. "4 holes Ø6.0 ±0.05" is four characteristics if each hole is measured, and one with a note ("4 places") if your customer accepts a single entry, usually recording the minimum and maximum readings, with any out-of-tolerance hole listed on its own line. Ask the customer or read their quality clause. Sub-features (a chamfer on each hole, a thread's major diameter and its fit class) count as separate lines if they are separately specified. The balloon list is complete only when a second person, reading the drawing alone, cannot find an unnumbered requirement.

Step 2: Build the Form 3 table from the balloon list

With the balloon list typed in a plain document or sheet, ask AI to arrange it into table form. Feed it only requirement information. Leave the results out.

Below is my balloon list for a machined bracket, drawing rev C.
Build a table with columns: Char No | Drawing zone | Requirement (as written) |
Nominal | Upper tolerance | Lower tolerance | Unit | Type (dimension, geometric,
surface finish, note, thread, material/process). Do NOT add any measured
values or results. Do NOT calculate limits. Copy numbers exactly as given.
If a line is ambiguous, list it under "Questions for the engineer".
1 | B2 | Overall length 85.00 +/-0.10 mm
2 | C3 | Bore diameter 12.000 +0.018/0 mm
3 | C3 | Position of bore to datum A, B: 0.05 mm
4 | D2 | Thread M8x1.25-6H, 2 places
5 | A4 | Surface finish Ra 1.6 max
6 | A1 | Flatness of face A: 0.05 mm
7 | Note | Break all sharp edges 0.2-0.5 mm
8 | Note | Material: aluminium alloy per customer spec

The returned table will have each line parsed into columns, and a short "Questions for the engineer" list ("Item 4: is the thread gauged on both places or sampled?"). That list is useful: ask the engineer before measuring, not after. Read every nominal and tolerance against the drawing, digit by digit, because a transposed figure at this stage runs through every later step.

Let the spreadsheet calculate the limits

The upper and lower limits are simple sums, which makes them a bad job for a chatbot (see why AI is bad at maths and what to check) and a good job for the sheet. Ask for the formulas rather than the numbers. For a row with nominal in column D, upper tolerance in E and lower in F:

Lower limit (G):  =D2+F2     (F holds a negative tolerance, e.g. -0.10)
Upper limit (H):  =D2+E2
Pass/fail (K):    =IF(J2="","",IF(AND(J2>=G2,J2<=H2),"Pass","FAIL"))

Here J is the result column, blank until the inspector enters a value. For the bore, with nominal 12.000, upper +0.018 and lower 0, the limits compute as 12.000 and 12.018. A reading of 12.009 passes; 12.021 fails and turns red. For geometric tolerances such as position or flatness, the limit is a maximum, so the formula checks that the result is at or below it. For purely visual notes, and for features checked with a go/no-go gauge, use a dropdown with Pass and Fail (an attribute result). A note that carries a figure, such as "break all sharp edges 0.2-0.5 mm", still needs a measured value: at least one aerospace customer's FAI guide rejects "OK" or "Accept" for any note that contains a dimension. Test the sheet with three invented values before it goes near a real part: one that is in, one that is out, one exactly on the limit.

A worked report: the bracket, from drawing to sign-off

Say a four-person shop has a first order for a machined aluminium bracket with 14 characteristics (the eight above plus six more dimensions and finishes). The customer requires an AS9102-style report with the first batch of 20 parts. An illustrative sequence, with the time each step takes:

1. Balloon (30 minutes). The machinist numbers the drawing. A second person checks it and finds one missed note about marking.
2. Table (10 minutes). The balloon list is pasted into the chatbot; the table comes back, plus two questions for the engineer. The shop owner fixes one transposed tolerance.
3. Sheet (15 minutes). The table goes into the report workbook with the limit and pass/fail formulas, checked with three test values.
4. Forms 1 and 2 (20 minutes). The identifiers, the material certificate number and the special process rows are typed in by the quality person, each checked against the paperwork.
5. Measurement (about 90 minutes). The inspector measures part 1 with the shop's calibrated instruments, entering each result and the gauge identifier into the sheet as they go.
6. Review (15 minutes). The sheet is read for completeness (below), then signed by the person responsible.

Ballooning, the table and the sheet together take about 55 minutes, and the forms and review another 35. The measuring still takes 90, which is as it should be. Total shop time comes to about three hours, against roughly four when the table and limits were typed by hand (illustrative figures), and the time saved comes from table-building and layout, not from touching a measurement.

How the inspector enters the results

The shop's measurement process is the record, so it is worth doing carefully. Give the inspector a printed or tablet copy of the sheet with the results column open and a column for the gauge or instrument ID beside it. The calibration status of each gauge is checked before use, not after the report is finished. For each characteristic, the inspector measures, reads the value, and types it in the cell to at least as many decimal places as the tolerance, and to the instrument's full resolution where it reads finer. Results go in the units the drawing uses, unless the customer has agreed otherwise. The inspector records an actual reading, not a rounded "good enough".

If results are dictated

Some shops like to speak the figures and have a tool transcribe them. It is quicker, and it brings a specific risk: "twelve oh oh nine" becomes 12.09 or 12.009, and "point zero three two" becomes 0.32. If you use dictation, have the transcript read back aloud against the gauge display, and let the sheet's limit formulas flag any value that is outside by an unlikely amount. A bore reading of 12.09 against a limit of 12.018 should prompt a re-check rather than a "Fail" entry, because it may be a transcription slip. An entry that is corrected must show who corrected it and when.

Let AI check the sheet, not the part

After measurement, AI is useful as a second reader of the finished sheet. Export the table with the results and ask it to look only for administrative faults.

Here is the completed FAI Form 3 table (CSV). Check ONLY the following:
1) any characteristic number missing from the sequence 1 to 14;
2) any row with a blank result, gauge ID or initials;
3) any result outside its stated limits that is not marked FAIL;
4) any result with fewer decimal places than the tolerance;
5) any row marked FAIL with no non-conformance reference.
List problems by row. Do not comment on whether the part is acceptable and do
not suggest any value.

Illustrative reply: "Row 7: result blank. Row 11: reading 25.4 with tolerance given to two decimals. Row 3: result 0.032 within limit but no gauge ID." Each of these is the kind of slip that gets a report rejected by a customer. The model does not know whether the part is good; it knows the sheet is incomplete. Run the same check by a person on the first few reports, to see whether the model catches what you catch, and what it misses.

What goes wrong when AI is allowed too close to the numbers

An illustrative case. A shop asks a chatbot to "complete the FAI for this part", pasting the table with a column of expected values. The model returns a tidy report with plausible results for every row, each inside its limits. The inspector is busy and reads it as a draft, then forgets to replace two rows. The customer's audit later finds results for a characteristic that the shop's own gauge list shows was not measured on that day. Nothing was wrong with the part. The record, however, claimed measurements that nobody made, and that is a serious quality failure whatever the part's condition.

The protections are structural, not good intentions: the AI-facing copy of the table has no results column, the results cells are filled only by the inspector, entries are initialled and dated, and the report's version history shows who changed what. For the wider habit of checking AI work properly, see how to check AI is doing good work, not just fast work.

When a result fails: the bore that came out at 12.021

Suppose the inspector's reading on characteristic 2 is 12.021 against an upper limit of 12.018. The sheet shows FAIL in red, and the sheet's job ends there. What happens next belongs to people:

1. The inspector re-measures with the same gauge, then with a second calibrated gauge or method, and records both readings.
2. The machinist checks the tool offset and wear. Say the boring bar has worn by 0.005 mm per part over the run.
3. The shop opens a non-conformance record, and its reference goes into the last column of Form 3.
4. The person responsible decides the disposition: correct the process and make a new part, or ask the customer for a concession.
5. A new first article part is made, and at least the affected characteristics, and any that depend on the same feature or tool, are re-measured and recorded.

AI can help with one piece of this, the write-up. Give it the inspector's own notes and ask for a plain-language non-conformance description.

Turn these notes into a short non-conformance description for the FAI record.
Do not add causes or fixes I have not stated. Notes: char 2, bore dia 12.000
+0.018/0, first reading 12.021, second gauge 12.020, both over upper limit by
0.002-0.003; boring bar offset checked, tool wear suspected; part quarantined
in red bin; machinist adjusting offset; new part to be measured.

A good reply states the characteristic, the requirement, the two readings, the amount over the limit, the containment action and what is pending, without a cause written as a fact. That wording is reviewed by the quality person before it enters the record, since "tool wear suspected" and "tool wear confirmed" are different statements to a customer.

Partial FAI and drawing revisions

Form 1 carries a field for a full or partial first article inspection, and customers usually require a new report, or a partial one, when something changes: a drawing revision, a new process, a move to another machine or location, or a long gap in production (one aerospace customer's supplier guide sets that gap at more than two years unless the contract says otherwise). A partial report covers only the affected characteristics and states the reason. AI helps here by comparing two balloon lists: paste the old and new lists and ask "which characteristic numbers have a different nominal, tolerance or note between revision B and revision C, and which are new or removed?" The result is a list of rows to re-measure. Check it against the drawing's revision block and the customer's change notice. The decision about whether the change makes a partial report acceptable belongs to your quality person and the customer, not to the comparison.

Keep each revision's report as its own file with the revision in the name. A model asked to "update the old report" will happily carry old results into the new one, which is precisely the failure this method exists to prevent. Start each new report from the blank table with the results column empty.

Confidentiality: read the contract before uploading a drawing

Customer drawings are often confidential, and some are controlled technical data that must not leave approved systems. Before pasting or uploading any drawing or balloon list into an AI tool, check the purchase order, any non-disclosure agreement and the quality clause. If they are silent or unclear, ask the customer in writing. Where use is allowed, prefer a business plan that does not train on your content by default, and strip the customer's name and part number from what you paste if the table does not need them. If permission is refused, you can still use the method with a generic template: build the layout and formulas on an invented part, then copy it into the real workbook by hand. If you are unsure how such a tool reads files, what ChatGPT can and cannot read in PDFs, spreadsheets and photos sets out what breaks.

Reading drawings: where AI helps and where it misreads

There is a tempting shortcut: upload the drawing PDF and ask the model to produce the whole table. It works well on clean, text-based PDFs and badly on scans, dense drawings and anything with small notes. Typical misreads: a general tolerance block applied to the wrong features, a "±" lost in the conversion, a datum reference read as a note, and a ballooned view labelled as the wrong section. A comparison of reading approaches, such as plain text extraction and AI reading, is in AI document extraction vs traditional OCR, and the problem of counting items from drawings is covered in how electricians use AI to count fittings from drawings. The pattern is the same: let AI propose, let a person confirm.

Before you sign: a completeness check for every FAI

• Every drawing requirement, including notes and specifications, has a balloon number and a line in Form 3.
• The balloon count on the drawing equals the number of rows, and a second person confirmed it.
• Limits come from formulas, and the formulas were tested with a pass, a fail and an on-limit value.
• Every result was entered by the inspector, with a gauge ID and initials, and nothing was typed by AI.
• Every gauge was within calibration on the day.
• Any non-conformance has a reference and a disposition decided by the right person.
• Form 1 identifiers (part number, drawing revision, serial or lot) match the drawing and the traveller.
• Form 2 lists the material and special processes with certificate numbers checked against the paper.
• The report is signed by the person responsible, and the version stored is the one sent.

That list is also a good prompt for the end-of-job review in AI quality control for small manufacturers, where the same line between formatting and judgement applies across the shop. If you would rather map this with someone, that is what my AI implementation consultation covers: finding the paperwork steps AI can take over and leaving every measurement and sign-off with your inspector.

Questions machine shops ask about AI and FAI reports

Can AI fill in the measured results on an FAI?

It should not. The results column is the evidence that a person measured the part with a calibrated gauge, and a figure typed by AI is not a measurement. Keep the results column empty in anything the AI sees, and let the inspector enter every value and initial it. AI can then check the entries against the limits.

Is it safe to upload a customer drawing to a chatbot?

Only if your customer's terms allow it. Many drawings are covered by confidentiality or export-control conditions, and some contracts forbid third-party processing. Read the purchase order and any quality clause first, ask the customer if unsure, and use a business plan that does not train on your content, or a local tool, when permission exists.

Does AI need to read the drawing for me?

It can help list characteristics from a clean drawing, but it misreads tolerance blocks, notes and small text more often than a person would expect. Use its list as a draft, and have a person balloon the drawing and confirm that every dimension, note and specification is on the table.

Which version of the standard should I follow?

Follow whatever your customer's purchase order or quality clause calls up, which may be a particular revision of AS9102 or their own FAI template. AS9102-style layouts with three forms are common, but the customer's requirement decides the fields, not a chatbot's memory of the standard.

Further reads

• How to Build a Staff Training Matrix With AI — Record who is trained and authorised to inspect and sign.
• How Much Does AI Visual Inspection Cost a Small Manufacturer? — What camera-based inspection costs a small manufacturer.
• What to Do Before You Buy Any AI Tool: A 10-Point Checklist — Ten checks before any AI tool touches customer drawings.
• How to Stop AI Tools Training on Your Business Data — Keep customer drawings out of model training.
• How to Document Your Processes Before Adding AI — Write your inspection routine down before adding AI to it.
• A Simple AI Risk Register for Small Businesses (With Template) — Log the risks of AI in your quality records.
• AI Supplier Management: Track Prices, Lead Times, and Risk — Track the material and process suppliers named on Form 2.
• AI Tools for Welding Shops: Quotes, Drawings and Job Scheduling — A welding shop owner's guide to AI on the office side: a drawing-based quote check, cut lists, a schedule stress test and where weld sign-off stays human.
• AQL Sampling Plan: Pick Sample Sizes for Incoming Goods With AI — Work out a sample size and accept/reject numbers for an incoming lot, with a spreadsheet lookup you can test against the standard.
• Supplier Audit Checklist: Plan and Write Up Visits With AI — A small maker's audit of a powder-coating supplier: pre-visit questions from the NCR log, a 3-hour agenda, an on-site checklist and a findings report.
• Supplier Corrective Action Request: Draft a SCAR With AI — A SCAR built around one failed bracket lot: evidence pack, containment ask, root-cause questions, due dates and a closure test the supplier can meet.
• WIP Tracking for a Job Shop: A Spreadsheet With AI Status Updates — A two-tab spreadsheet for jobs and operations, the formulas that flag late work, and prompts that draft customer updates from the daily export.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: published descriptions of the AS9102 first article inspection report (three forms: part number accountability; product accountability for materials, special processes and functional tests; characteristic accountability), checked against public summaries of the standard. The customer's quality clause and the current revision of the standard decide the required fields.