A 100,000-file company, built to break the product
Doppelgone for Work has no customers yet, so we built one. This is our own test, run by us on a Microsoft 365 tenant we own. It is not a case study and nobody here is a real person or company. We publish it so you can see how we test, and what we measured.
The setup
| The firm | Lowveld Engineering (Pty) Ltd. Invented. No real company, no real person. |
| People | 18 active staff, 2 who have left (their OneDrives are still full) and 1 outside auditor as a guest. |
| Places | 20 OneDrives and 9 SharePoint sites: Operations, Finance, HR, Legal, Projects, Sales, Marketing, Archive and a private Board site. |
| Files | At least 100,000, about 30 GB: Office files, PDFs, images, a few hundred large videos and small noise files. |
| Where they come from | Real public-domain documents (contracts filed with the US SEC, US government PDFs, open data, public-domain images and books) rendered into real Word, PDF, Excel, PowerPoint and image files, plus documents our own generator wrote. |
| The copies | We planted them, so we know the right answer: exact copies across sites and people, “Copy of”, renamed files, v1 to v2 to FINAL chains, Word files saved as PDF, and look-alike traps (the same contract template for different clients). |
| The mess | Retention labels on about 200 files, files edited after the first scan, 5% of files moved, renamed or deleted before the second scan, private permissions and anonymous links, very long paths, unicode names. |
What we did
- Ran the free audit on the whole tenant, as a customer would.
- Checked every suggestion against the answer we planted, and read by hand the suggestions on real documents that have no label.
- Changed the tenant (edits, moves, renames, deletions) and scanned again, to test the second scan.
- Ran a clean-up on the approved groups, moved files to the recycle bin, checked they were gone, then undid it and checked they were back.
Results
Measured on our test tenant between 27 and 29 October 2026; we did not estimate any of them. Older versions and near-identical documents are never pre-selected: a person confirms each one, so they are not in the precision rows.
| Measure | Result |
|---|---|
| Files scanned | 101,823 |
| Bytes scanned | 25 GB |
| Scan time (minutes) | 7.5 |
| Second scan, after the changes (minutes) | under 1 |
| Whole audit, request to report (hours) | 8.3 (bound by Microsoft’s read limit for one app) |
| Duplicate groups found | 20,954 |
| Space that can go safely | 3.7 GB (14.9% of the storage) |
| Space that needs a look | 449 MB |
| Pre-selected removals checked against the planted truth | 422 copies (identical files and same text, other format) |
| Of those, wrong | 0 |
| Look-alike traps merged by mistake | 0 in the sample |
| Report page: server answers / fully drawn (milliseconds) | 96 / 5,836 |
| AI cost of the whole audit (USD) | 1.69 (40.2M input tokens, second run) |
| Files restored by undo, and checked back | 36 of 36 |
What this does not show
A fake firm is tidier than a real one in some ways and messier in others. These results say how the product behaves on a large, deliberately difficult tenant. They do not say how many customers use it, or what a real company will save. The calculator uses an assumption for that, and the free audit measures yours.
Last reviewed 27 October 2026.