Lubinpla

AI Checks of a Certificate of Analysis Against the Specification

Summary

AI can check a certificate of analysis (CoA) against a specification item by item, if the lot, edition, units and methods match. Near-limit results stay with people. The AI must read your current specification and the CoA for the lot you actually received. Any item the CoA leaves out goes to a person as not verified.

This method covers the numeric and pass or fail items on a supplier CoA, compared with a written specification for a homogeneous product such as a lubricant, solvent or additive. In regulated drug supply it does not replace identity testing or supplier qualification. And the precision methods it relies on are not written for non-homogeneous products (ISO 4259-2:2026).

When should you use AI to check a certificate of analysis against the specification?

Use AI for this check when the work is document comparison at volume: many deliveries, many items per CoA, and a written specification for each product. Kognitos describes incoming CoA checks as reading a document and comparing it against something else, such as a specification, a purchase order or the history of earlier lots, and observes that the long tail of routine receipts often gets a glance, a filing and a release (Kognitos, 2026). AI can close that gap because it reads every row of every CoA, not a sample.

Three moments call for it. One is at goods receipt, before a lot goes to production or is resold. Another is just before a distributor forwards a manufacturer's CoA to a customer whose specification differs from the manufacturer's sales specification. The last is when a customer disputes a delivered lot and someone has to show, item by item, what the CoA proved against which specification.

The limit comes from the kind of task. A generative model can produce confidently stated but false content, which NIST calls confabulation, and NIST notes the risk is particularly relevant in domains that require highly contextual or domain expertise (NIST AI 600-1, 2024). Comparing printed numbers with printed limits is different, because the reviewer can check every AI statement against the two documents on the desk. So use AI for that comparison. Acceptance, rejection and questions to the supplier stay with a person who owns them.

What does an AI CoA check need besides the certificate of analysis?

Besides the CoA itself, the check needs four inputs: the governing specification in its agreed edition, the order or contract that names that edition, the delivery paperwork with the lot number, and the acceptance rule for near-limit results. Without the specification text, a model falls back on typical values, and a typical value is not your limit. For why a general chatbot cannot answer reliably from documents it has not been given, see whether ChatGPT can answer from your company's technical documents.

Each input settles one criterion. The specification supplies the item list, the limits, the units and the named test methods. The order or contract fixes which specification edition applies. That matters because a supplier may test against its own sales specification, and Kognitos advises checking that the certificate reports against your specification, not merely the supplier's, because a result within the supplier's limits can sit outside yours (Kognitos, 2026). The acceptance rule says what to do with a result close to a limit; when ASTM D3244-24 is used, it requires the acceptance limit to be determined before testing begins (ASTM D3244-24, 2024).

Regulated supply adds a fifth input: supplier qualification records. Under ICH Q7, a supplier's CoA can be used in place of other tests on incoming material provided the manufacturer has a system in place to evaluate suppliers. Even then, at least one identity test should still be conducted on each batch, except for the materials listed in 7.32 (ICH Q7, 2000, 7.30).

How do you check a certificate of analysis against the specification step by step?

Work through six criteria in a fixed order: lot, specification edition, missing items, units, test method, and margin to the limit. The first two come first because they decide whether any comparison is valid. If the CoA belongs to another lot, or the specification is an outdated edition, every later row means nothing. The table gives, for each criterion, the threshold to apply, what follows if it is met or not, and who decides.

Criterion

Threshold (source)

If met

If not

Lot match

Lot number on the CoA equals the lot on the delivery label and paperwork; a certificate for a different lot is evidence of nothing (Kognitos, 2026)

Continue

Stop; quality staff request the CoA for the delivered lot

Specification edition

Edition and date of the specification match those named in the order or contract

Continue

Stop; the commercial or quality owner confirms which edition governs

Every item reported

Each specification item has a result; for APIs, ICH Q7 recommends the CoA list each test with acceptance limits and numerical results (ICH Q7, 2000, 11.42)

Continue

Item marked not verified; quality decides to request data or test in-house

Unit

Same unit as the specification, or a cited conversion such as 1 mm2/s = 1 cSt (ASTM D445-26, 2026)

Compare values

AI flags the row; no conversion without a cited factor

Test method and edition

Same method as the specification names, in an edition the publisher lists as active

Compare values

AI flags method differs; the quality reviewer decides if the result is usable

Margin to limit

Result clear of the limit by more than the margin agreed from method precision; acceptance limit set before testing (ASTM D3244-24, 2024)

AI marks meets

AI marks near limit; a person applies the agreed rule

The margin to the limit needs the most care. If a CoA value sits closer to the specification limit than the margin agreed from method precision, the AI marks it near limit, and a person applies the rule fixed before testing (ASTM D3244-24, 2024). Every AI flag should cite the CoA row and the specification clause it compared, so the reviewer can confirm it without rereading both documents. A flag with no cited location is a claim to verify, not a result.

Why do CoA units and test methods break an AI comparison with the specification?

Two numbers can be compared only if they measure the same property, by the same procedure, in the same unit. Some unit differences are just naming: ASTM D445-26 states that 1 mm2/s = 1 cSt for kinematic viscosity, so a CoA in cSt and a specification in mm2/s compare directly (ASTM D445-26, 2026). Others need arithmetic. Take a water content reported in percent by mass against a limit in mg/kg, where a slipped decimal turns a pass into a fail. The AI should show the conversion factor and its source next to each converted value, so a person can check it.

Test methods are harder, because no factor fixes them. ASTM D92-24 determines flash and fire points with a Cleveland open cup tester, and ASTM D93-26 determines flash point with a Pensky-Martens closed cup tester (ASTM D92-24, 2024; ASTM D93-26, 2026). These are separate test methods, so a D93 result is not a D92 result. Methods that look equivalent can also differ in scope. ISO 3104:2023 covers liquid petroleum products, fatty acid methyl ester (FAME), paraffinic diesel, hydrotreated vegetable oil (HVO) and biofuel diesel mixtures up to 50 % FAME (ISO 3104:2023), while ASTM D445-26 is written for liquid petroleum products (ASTM D445-26, 2026).

So a test method difference on a CoA is a flag, not a fail. Whether a result by another method can stand in is a technical judgment for the quality reviewer.

How should AI treat a CoA result on the specification limit?

A CoA result on or near the specification limit is undecided. The AI passes it to a person with the margin, the method precision and the agreed rule. JCGM 106:2012 states: "Because of uncertainty in measurement, there is always the risk of incorrectly deciding whether or not an item conforms" (JCGM 106:2012). How the parties share that risk is a business agreement.

The guide sets out the options against the tolerance interval, which is the range of values the specification allows. Under simple acceptance, also called shared risk, the parties accept an item whose measured value lies in the tolerance interval and share the consequences of incorrect decisions (JCGM 106:2012, 8.2.1). Under guarded acceptance, the acceptance limit sits inside the tolerance interval, which reduces the risk of accepting a nonconforming item (JCGM 106:2012, 8.3.2). There is a trade-off. With a binary decision rule, JCGM 106 adds, reducing the consumer's risk always increases the producer's risk (JCGM 106:2012, 9.5.4).

Petroleum products and lubricants have their own practice. ASTM D3244-24 lets two parties set an acceptance limit from a mutually agreed probability before testing (ASTM D3244-24, 2024), and ISO 4259-2:2026 specifies how to determine conformance when supplier and receiver results conflict (ISO 4259-2:2026).

Rounding is the second near-limit trap. ASTM E29-22 says a reference to it is valid only when the absolute or the rounding method has been chosen (ASTM E29-22, 2022). Without an agreement, ASTM D3244-24 recommends rounding by the E29 rounding-off method to the significant digits of the governing specification (ASTM D3244-24, 2024). Take an illustrative acid number of 0.104 against a 0.10 maximum: the AI needs that rule before it can compare the two.

What does a missing CoA item mean in a specification check?

A missing item means that property is not verified, and the check cannot report the lot as conforming on it. That is why the AI should list the specification items first and then look for each one on the CoA. Working in CoA order hides whatever the CoA leaves out. For active pharmaceutical ingredients (APIs), ICH Q7 recommends that the certificate list each test performed in accordance with compendial or customer requirements, including the acceptance limits and the numerical results obtained (ICH Q7, 2000, 11.42).

An item can be missing for several plausible reasons, and the documents alone cannot tell them apart. The supplier may test it periodically rather than on every lot. It may be in your specification but not in the supplier's sales specification, a mismatch Kognitos warns about (Kognitos, 2026). Or it may have dropped out of the CoA template. Each cause calls for a different action, so the AI reports the item as not reported on the CoA and leaves the cause open.

Rows that read only pass or conforms leave a related gap. Kognitos notes that a certificate of conformance attests that a lot meets its specification without necessarily including the underlying results (Kognitos, 2026). A CoA row with no number gives no margin to judge. In US drug manufacturing, each component is tested for conformity with all appropriate written specifications unless the conditions for accepting a supplier's report of analysis are met (21 CFR 211.84(d)(2), eCFR 2026).

What happens when the CoA is checked against the wrong specification edition?

Against the wrong edition, every row can pass while the lot fails the edition in force, or the reverse. Specifications change by edition, and so do the test methods they cite. ISO 3104:2023 replaced ISO 3104:2020 on 7 November 2023, and ISO 4259-2:2026 replaced ISO 4259-2:2017 on 9 June 2026, according to the EVS standards listing (ISO 3104:2023; ISO 4259-2:2026). A customer specification revised to tighten a limit or add an item behaves the same way: the CoA may be correct against the old revision and short against the new one.

The wrong edition usually gets in one of three ways. An AI given a folder of specifications may pick the newest file by name rather than the edition named in the order. A supplier may issue the CoA against its own current sales specification while the contract fixes an earlier customer revision. And ASTM D3244-24 requires the criticality of the specification to be mutually agreed before testing when that practice is used, which presumes both parties refer to the same specification text (ASTM D3244-24, 2024).

That is why the edition check is the second step. The AI reads the edition and date from the specification, the edition cited on the CoA if there is one, and the edition named in the order, and stops if any two disagree. Which edition governs a delivery is a contractual question for the commercial or quality owner, not for the AI.

Worked AI CoA check example (illustrative)

Here a hydraulic oil CoA is compared with a customer specification, and one of five items can be accepted without a person. All values, methods and limits are invented for illustration, and the lot number and specification revision are assumed to match the order. The table shows what the AI reports for each specification item.

Item

Specification (illustrative)

CoA (illustrative)

AI report

Kinematic viscosity at 40 °C

41.4 to 50.6 mm2/s, ASTM D445

46.2 cSt, ASTM D445

Meets; 1 cSt = 1 mm2/s (ASTM D445-26, 2026)

Flash point

200 °C min, ASTM D92

214 °C, ASTM D93

Method differs; person decides

Water content

200 mg/kg max

0.019 % by mass

Converted to 190 mg/kg; near limit; person decides

Acid number

0.10 mg KOH/g max

Not reported

Not verified

Foaming tendency

50 mL max

Pass

No numeric result; not verified

Watch the water content row. Converted, 0.019 % by mass is 190 mg/kg, 10 mg/kg under the 200 mg/kg maximum. Whether that margin is safe depends on the method precision and the agreed acceptance rule, and the illustrative specification states neither. The other rows are simpler. Viscosity meets because the unit difference is only a naming difference with a cited factor. Flash point is flagged because the CoA used a closed cup method where the specification names an open cup method. Acid number is flagged because the item is absent, and foaming because pass is a statement, not a result.

The reviewer gets a short list: one item meets, two need technical judgment and two need data from the supplier. The AI does not say whether the lot passes or fails. The reviewer asks the supplier for the acid number and a numeric foaming result, then decides, with the precision data for the water method, whether 190 mg/kg is acceptable under the contract.

Where does an AI certificate of analysis check not apply?

The check does not apply where the decision depends on evidence that is not on paper. It cannot confirm identity: ICH Q7 recommends at least one test to verify the identity of each batch of incoming material, with listed exceptions (ICH Q7, 2000, 7.30), and US drug manufacturers must conduct at least one specific identity test on a component even when they accept a supplier's report of analysis (21 CFR 211.84(d)(2), eCFR 2026).

Nor can it prove that a certificate is authentic. Kognitos recommends keeping a reference file of each key supplier's CoA format, signature style, layout and contact details and flagging any deviation (Kognitos, 2026). AI can help with that comparison, but a person must own it.

The precision-based acceptance methods have their own scope. ISO 4259-2:2026 was designed for petroleum and petroleum-related products, which are normally homogeneous; it states the procedures can also be applied to other homogeneous products and are not applicable to non-homogeneous products (ISO 4259-2:2026). ASTM D3244-24 applies only to test methods that specifically state their repeatability and reproducibility values conform to its definitions (ASTM D3244-24, 2024).

Finally, the check cannot replace supplier qualification. ICH Q7 recommends full analyses on at least three batches before reducing in-house testing, and a full analysis at appropriate intervals compared with the certificates (ICH Q7, 2000, 7.31). A clean AI comparison says the CoA is consistent with the specification. It says nothing about whether the supplier's laboratory is reliable.

What must a person decide after an AI CoA check?

A person decides acceptance of the lot, every near-limit result, every method substitution, every missing item and any specification edition conflict. These decisions carry contractual or regulatory responsibility, and NIST notes that confabulation risks may be especially important to monitor when generative AI is integrated into applications involving consequential decision making (NIST AI 600-1, 2024). The AI's job ends at a cited comparison. Release, rejection and supplier questions stay with named staff.

A usable AI CoA report has four parts. The first identifies what was compared: CoA lot and date, specification title, edition and date, and the order that names it. The second gives one row per specification item with the CoA value, the limit, a status (meets, near limit, method differs, unit converted, not reported) and where each value sits in its document. The third lists the open questions for the supplier. The fourth lists the decisions left for the reviewer, each with the evidence needed to make it.

Record each reviewer decision and its reason next to the flagged row. To keep near-limit decisions consistent across lots, write the acceptance rule into the specification or contract before testing, as ASTM D3244-24 expects when it is used (ASTM D3244-24, 2024). In Lubinpla, you can attach a CoA in the main agent chat and @-mention a TDS from the Dataroom or the specification in a customer deliverable; the agent compares them item by item, with a source for each value, for you to judge.

Key takeaways

  • Check the lot and the specification edition first; a mismatch voids every later row of the CoA check.

  • List the specification items first, then find each on the CoA, so omissions show.

  • Convert units only with a cited factor, and treat a test method difference as a flag.

  • Agree the near-limit acceptance rule and the rounding method before testing.

  • AI compares documents; a person decides acceptance, identity testing and supplier trust.

Related posts