AI-assisted LCA modelling and one-click EPD generation

Most of the time in an LCA project goes to two places: linking inventory lines to the right ecoinvent activity, and writing the report text. Both are pattern-recognition work — exactly the kind AI is good at. The critical design question is: what should AI do, and what should it never do? In tools where that answer is set up wrong, "automatic LCA" becomes a black box that collapses in verification.

What the assistant does: suggest, check, interpret

clca's AI assistant works in three modes. Suggest: type "zinc-coated sheet, 0.8 mm" and it lists candidate activities from ecoinvent 3.12 with geography fit; the choice is yours. Check: it scans your inventory and flags suspicious patterns — a line that looks like tonnes entered as kg, a raw material with no transport defined, a preliminary result outside the sector band. Interpret: when calculation finishes it summarises the contribution analysis in plain language; "61% of GWP comes from the A1 steel input" is worth half an hour of process-tree browsing.

What the assistant does not do: the boundary line

• It never finalises activity selection on its own — every suggestion awaits user approval
• It never fabricates inventory data — missing data is flagged missing, not filled by guess
• It never alters results — the calculation engine is deterministic and out of the assistant's reach
• It makes no methodological decisions — system boundary, allocation and scenarios belong to the expert
• Every number entering verification has a human-approved source

This boundary is not arbitrary: ISO 14044 and EPD verification require every modelling decision to be justifiable. "The model suggested this activity, we approved it, and here is the rationale" is a defensible sentence; "the model picked it, we don't know why" is not.

One-click EPD: from template to verifier package

Once the model is complete, EPD generation is one click: pick the program operator (ITB, IBU, EPD International or ECOnsult) and clca produces the PDF from that operator's preconfigured template — cover, mandatory declaration blocks, EN 15804 result tables, scenario tables and operator-specific disclaimer texts all in place. Report generation locks the inventory; the numbers sent to the verifier cannot drift from the numbers in the model.

The net gain: weeks off the calendar

In practice the gain is measured in two places. In inventory setup, suggestion mode visibly shortens activity search time — especially on hundred-line recipes. On the report side, template automation zeroes out the formatting and copy errors of hand-built tables; the verifier's first-round finding count drops, and at 1-3 weeks per round the calendar shortens directly.