- Recurring reporting
Two Wurqurs, Management Reporter and Reconciler, build the management pack on Sunday night, unattended.
~$20,000 saved a year for a dental group
See howThe knowledge already existed. Finding it just took longer than asking a person.
Librarian and Precedent Finder answer questions against ten years of a civil engineering consultancy's own project files, the way three senior engineers used to, one interruption at a time.
Civil engineering consultancy · 90 employees · three systems
The interruption
Ten years of technical knowledge sat in old project folders across SharePoint: Word files, technical standards, completed project reports. Engineers knew a colleague had solved a similar problem before, but finding the relevant document took longer than asking. So they asked, and the same senior people answered the same questions repeatedly, an interruption at a time.
150 to 250 of those questions came in a month. Nothing was missing. The company already had the answer written down somewhere; getting to it was the problem.
The knowledge already existed. This was a retrieval problem, not a documentation gap.
What we found, and left alone
The knowledge already existed, ten years of it. A colleague usually knew the answer, so people asked instead of searching, and the same senior engineers answered the same questions again and again. Engineering judgment stayed exactly where it was: the system finds prior work, it doesn't decide what to do with it.
What we built
Librarian indexes their SharePoint library automatically and answers general questions against the company's own corpus, with citations back to the source documents. Precedent Finder does a narrower job: finding the comparable project or matter that was already solved, and saying how similar it actually is. Both respect each engineer's existing Entra permissions, so nobody sees more of the corpus than they're already allowed to, and both say plainly when an answer isn't there rather than inventing one.
What changed
finding prior work
15–30 min
Under 30 sec
1,800–3,000
questions a year, answered fast
900 hrs
a year, back to the senior team
~$43,000
a year
What used to cost them roughly $3,600 a month by hand now runs on two Wurqurs for $251/month.
Where the humans stayed
Engineering judgment. The system finds the prior work; it does not decide what to do with it.
What it connects
Where the documents live, the two Wurqurs that search them, and the permission check before an answer goes out.
where it arrives
- SharePointthe ten years of project folders
what runs
- Librariananswers general questions, with citations
- Precedent Finderfinds the comparable project, says how similar
- PostgreSQL + pgvectorthe corpus index both Wurqurs search
where it lands
- Microsoft Entra IDchecks who's allowed to see what, before an answer is delivered
What it took
Six weeks. Four to look at their data, define the load format, work out matching, build the presentation and deploy. Two to shadow-run.
The same idea, somewhere else
A number that used to take a person a day to assemble, now checked and built before anyone opens it.
Every company has an answer nobody can find fast enough.
Tell us the job. We will tell you whether a Wurqur can hold it.