Checking AI output

The skills do the drafting. This lesson is about the part that is now your job: checking. Good checking is not re-doing the work. It is knowing where drafts go wrong, looking there first, and recording what you looked at. With practice, checking a skill’s output takes minutes, not hours.

The checker’s checklist

Use this on any AI-drafted environmental analysis, from a learnLCA skill or not.

1. Is it the right job? Read the label and the method block at the top. Is this the right unit, boundary and standard for the decision in front of you? If the method is wrong, nothing below it matters.

2. Does every number have a source? Scan the source column. Any number without one is a finding.

3. Spot-check three numbers at source. Open the cited page, cell or dataset and compare the value, the unit and the exponent. Pick the biggest contributors first.

4. Redo one calculation by hand. Quantity times factor, with the units written out. If the output says “not tool-checked”, redo more.

5. Read the missing-data list. Is anything missing that could change the conclusion? Is the total labelled as excluding it?

6. Read the flags. Expired documents, missing modules, proxies, “unclear” answers. Decide what each one means for your use.

7. Work through the skill’s own checker’s list, and tick each item.

8. Record what you checked (see “Keep an audit trail” below).

Explainer

Check, don’t write

1. Scene 1 of 6

   The old way: write it yourself

   One specialist. One spreadsheet. Every number typed by hand.

   Request→Weeks of work→One report

   More requests means a longer queue.

   The old way: a specialist writes the whole analysis by hand, one request at a time. When more requests arrive, the queue just gets longer.

2. Scene 2 of 6

   The assistant drafts

   BOM+Standard+Database=Draft analysis

   An AI assistant with learnLCA skills does the first draft.

   Methodological choices are confirmed with you, not decided silently.

   The new way: you give an AI assistant with learnLCA skills the bill of materials, the standard and a database, and it drafts the analysis. Choices such as the standard, functional unit and boundary are confirmed with you, not made silently.

3. Scene 3 of 6

   Every number cites its evidence

   Steel, 4 kg[BOM line 1]

   Steel: 7.942 kg CO₂-eq[dataset name and version]

   Paint: supplier value[EPD reg. no., page]

   Packaging weight[missing: flagged]

   In the draft, every number points to where it came from: a line of the bill of materials, a page of an EPD, or a named dataset and version. Anything missing is flagged as missing, never filled with a plausible guess.

4. Scene 4 of 6

   You check

   The checker’s list

   • Right standard and functional unit?
   • Every BOM line included?
   • Each factor matches its source?
   • Arithmetic tool-checked?
   • Flags resolved?

   The human checks rather than writes, working through a checker’s list: the right standard and functional unit, every bill of materials line included, each factor matching its cited source, the arithmetic checked, and every flag resolved.

5. Scene 5 of 6

   Scale without dropping quality

   Drafting is fast. Checking is where the judgement goes.

   Drafts for several drivers, EPDs, the Construction Products Regulation, CBAM and product carbon footprints, flow to one checker who applies the same evidence rules to each. Drafting becomes fast, and the human judgement goes into checking, so a team can take on more work without lowering quality.

6. Scene 6 of 6

   Check, don’t write

   The assistant writes the draft.

   The evidence travels with every number.

   You check it, and you sign it off.

   Summary: the assistant writes the draft, the evidence travels with every number, and a person who understands the method checks it and signs it off.

Failure modes to look for

These are the mistakes that matter most in environmental data, AI or not. Knowing them tells you where to look.

Failure mode	What it looks like	How to catch it
Invented factors	A neat emission factor with a vague source (“typical value”, “industry average”, “ecoinvent” with no dataset or version)	Every factor must name its dataset and version, or an EPD registration and page. No name, no factor.
Unit errors	Grams treated as kilograms, per tonne as per kg, per m² as per kg; 1.2E-01 read as 1.2	Write the units out in your hand check. Look for results that are 1,000 times too big or small.
Silent boundary changes	The method says cradle to gate, but a line for transport or end of life has crept in; or modules added together, including module D	Compare the method block with the lines in the table. Module D is always reported separately.
Stale or expired EPDs	An EPD past its valid-until date, or superseded by a newer version on the register	Check validity on the programme operator’s register, not just the PDF.
Mixed rulebooks	EN 15804+A1 values set against +A2 values; different PCRs compared as if equal	The comparability check in supplier-epd-compare; the standard version in every summary.
Misread tables	A value copied from the wrong column or module, often in wide PDF tables	Spot-check at the cited page, looking at the column header.
Gaps quietly filled	A total that looks complete but the BOM had lines with no factor	Compare the number of BOM lines with the number of result lines; read the coverage statement.
Hard regulatory statements	“You must comply by…” or “your goods are in scope” with no source	Regulatory statements need an official source and a date checked. Skills should hedge; if one doesn’t, treat it as unverified.

Exercise: find the error

A screening output contains these lines. The method block says: declared unit one bench, cradle to gate.

Line	Description	Quantity	Factor	Factor source	Result kg CO2e
1	Steel tube	6 kg	1.986 kg CO2e/kg	Example Database v1.0, “steel tube, at plant”	11.916
4	Powder coating	300 g	2.579 kg CO2e/kg	Example Database v1.0, “powder coating”	773.7
8	Delivery to customer	3.4 tkm (0.0227 t x 150 km)	0.1 kg CO2e/tkm	“typical truck”	0.34

(Hypothetical teaching values.) What’s wrong?

Answer: Three things. Line 4 multiplied 300 by 2.579 without converting grams to kilograms; it should be 0.3 x 2.579 = 0.7737 kg CO2e. Line 8 is delivery to the customer, which is outside a cradle-to-gate boundary: a silent boundary change. And line 8’s factor has no real source (“typical truck”), which is an invented factor. The second and third problems would be caught by reading the method block and the source column; the first by one hand calculation.

Spot-checking at scale

When a team runs a skill on dozens of EPDs or BOMs, nobody checks every number. Check in a way that would catch a systematic problem.

• Always check the hotspots. The few lines that make up most of a total, and the values that drive a decision. An error there changes the answer.
• Always check everything flagged. Proxies, missing data, expired documents, “unclear”.
• Sample the rest. For example, three values per document, or one document in ten, chosen at random rather than the easy ones. Agree the rate as a team and write it down.
• Widen the check when you find an error. One error in a sample means look at more of the same kind: if one EPD’s table was misread, check the others from the same programme or layout.
• Feed errors back into the skill. If the same mistake appears twice, fix the step in the skill (lesson 4) so it doesn’t happen again.

Why this works

Evidence rules make sampling meaningful. Because every number points to its source, a sample check is a real check, not a guess. Without citations, the only way to check is to redo the work.

Keep an audit trail

If someone asks in six months where a number came from, you should be able to answer in a minute. Keep, for each run that feeds a decision or leaves the building:

Keep	Why
The input files (BOM, EPDs, supplier replies) as they were	So the sources can be reopened
The skill name and the date you downloaded it	Skills are updated; you need to know which procedure ran
The output, including the method block and checker’s list	It records what was assumed and flagged
What you checked, by whom, on what date	The evidence that a person reviewed it
Any corrections you made, and why	So the next run, and the next checker, benefit

A shared folder per product or supplier, with a short log file, is enough. It doesn’t need special software.

What you should now be able to do

• Check an AI-drafted analysis with a short checklist, starting with the method.
• Recognise the common failure modes: invented factors, unit errors, silent boundary changes, stale EPDs, mixed rulebooks, misread tables.
• Plan a spot-check that would catch a systematic error when a team runs skills at scale.
• Keep an audit trail that lets anyone trace a number back to its source.

Sources

• ISO 14044:2006 (interpretation: completeness, sensitivity and consistency checks). https://www.iso.org/standard/38498.html
• EN 15804:2012+A2:2019, Sustainability of construction works: environmental product declarations, core rules (module D reported separately). https://standards.cencenelec.eu/

Checked: 2026-09-29