Case studies: measured results in engineering teams
Two real cases of AI under engineer control. At a custom-machine SME, automating certificate gathering and completeness checks cut compliance-package compilation time by roughly 80%. At a large international life-science group, writing structured regulatory summaries — which took from 11 to around a hundred hours per document — was brought down to about 30 minutes, a proven result. In both cases: the model proposes, code verifies, the engineer decides.
01Compliance packages: roughly 80% less time at a custom-machine SME
The context. An industrial SME designs custom machines for the pharmaceutical and nuclear sectors. Each market release requires a regulatory technical file — CE marking, ATEX certificates, IMDS declarations, PPAP items, machine documentation. Gathering these parts takes 2 to 5 days per bill of materials, and the company produces around forty packages per year.
The problem. The work isn't intellectual, it's logistical: finding the right version of a certificate from each manufacturer, checking that no mandatory item is missing, repeating the exercise in the language of the destination country. Qualified engineering time absorbed by repetitive gathering, at the expense of design.
The method. Roughly 80% of certificates and data sheets for standard components proved retrievable automatically from manufacturers' public sites. An agent gathers these parts; code — not a model — verifies deterministically the presence of every mandatory item, reference consistency, and traceability to source; the engineer arbitrates the cases flagged as incomplete.
The result. Roughly 80% less time on package compilation, with every item keeping a link to its source to stay auditable before a notified body.
02Regulatory summaries: from 11 h to 30 min at a large life-science group
The context. A large international life-science group produces structured regulatory summaries from source reports — the kind of documentation volume any engineer in pharma, chemicals or medical devices will recognize. A single summary takes around 11 hours of a qualified writer, and up to a hundred hours for the heaviest files.
The problem. Transposing a source report into a stable target format is repetitive but sensitive: a wrong value, or a gap filled by invention, can compromise the file. This is exactly where a language model alone fails, because it hallucinates on what is missing.
The method. The model generates the summary relying only on the source report; code verifies every produced value against the source and flags missing information rather than inventing it; the expert no longer writes — they review the sensitive points and the gap markers.
The result. Production time brought down to about 30 minutes per document — a proven result, roughly 95% less time, with value-to-source traceability intact.
03Frequently asked questions about these case studies
Are these figures real results or targets?
They are results observed at real clients. The roughly 80% reduction in compilation time was measured at a custom-machine SME; the move from 11 hours to 30 minutes per regulatory summary is a proven result at a large life-science group.
Why aren't the clients named?
The cases are described by sector and type of organization rather than by name, to preserve confidentiality while staying transposable: an engineer in a comparable context can recognize their own situation.
Does the AI make the final decision?
No. In both cases the AI prepares and gathers, deterministic code verifies objective rules, and the engineer or expert keeps the final judgment and responsibility.