A 35-person UK engineering team had 18 years of reports and specifications scattered across shared drives. A secure, retrieval-based AI assistant indexed 18,400 documents, answered with cited sources and cut search time per engineer from 5.2 to 1.5 hours a week.
Key numbers
A UK engineering design team of 35 people had accumulated 18 years of reports, test results, specifications and design guidelines across shared drives. The knowledge was valuable, but finding it was slow. Engineers spent hours each week searching folders, asking colleagues and sometimes repeating tests that had already been done, simply because they could not find the earlier result.
The situation
A time study showed that engineers spent an average of 5.2 hours a week searching for information. Some of the most experienced engineers spent even longer, answering colleagues' questions about past projects. Any solution had to protect confidential client information and respect existing access permissions.
What changed
The team built a secure engineering knowledge assistant, applying the governance principles in AI governance for engineering firms:
- Documents were cleaned, de-duplicated and indexed in a secure environment, with superseded versions identified.
- A retrieval-based assistant was built to answer questions using company documents, always citing the source documents for each answer.
- Access controls mirrored existing folder permissions, so engineers could only retrieve documents they were already allowed to see.
- Before launch, accuracy was tested with 200 real questions from engineers, and the results reviewed by senior staff.
- Users could flag unhelpful or incorrect answers, feeding continuous improvement.
What the data showed
Average time to find an answer fell sharply for every type of question: material specifications from 18 to 3 minutes, past test results from 34 to 6 minutes, design guidelines from 12 to 2 and supplier history from 26 to 5 minutes.
Results
Search time per engineer fell from 5.2 to 1.5 hours a week, a saving of about 3.7 hours each, or 71%. In testing, 94% of answers cited the correct source document. In total, 18,400 documents were indexed. Engineers reported spending more time on design and less on searching, and fewer tests were repeated unnecessarily.
Why it was trusted
Citations made every answer checkable, and the assistant was designed to say when it could not find an answer rather than invent one, avoiding the risk described in ChatGPT for engineers. Respecting existing permissions meant confidential client information stayed protected.
Lessons for your operation
- Measure how much time engineers spend searching. It is usually more than expected.
- Clean and curate documents before indexing them.
- Require cited sources for every answer.
- Mirror existing access permissions.
- Test accuracy with real questions before launch.
