Construction documents are full of names. The owner’s name sits in every title block and change order header. The site address is on every daily log. Superintendents sign off with their cell numbers. Before any project record leaves your systems for an AI buyer, those details need to go, and they need to go in a way that keeps the records useful.
Here is a process a contractor can run with its own staff and ordinary tools.
Why clean before the buyer does
Each buyer that explains its process mentions de-identification. Scale AI puts the step on your own systems ahead of extraction; micro1, Nyne and Labelbox Prism describe doing it themselves. So why bother?
- You know your projects. A buyer’s tool will catch “John Smith”. It will not know that “the Elm Street job” or “Phase 2 at Riverside” names a client.
- You see exactly what leaves. Reviewing your own cleaned set is the only way to be sure.
- Your contracts may require it. Confidentiality clauses rarely care who removed the details, only that they were protected.
Step 1: build the token map
Start with a spreadsheet that lists every identifier and the token that replaces it. One map, used everywhere.
| Identifier | Token pattern | Example |
|---|---|---|
| Owner or client | OWNER-nnn | OWNER-014 |
| Project name or number | PRJ-year-nnn | PRJ-2021-007 |
| Site address | State and building type | Ohio, two-story retail |
| Your staff | Trade and role | PM-03, SUPER-11 |
| Subcontractors | SUB-trade-nn | SUB-ELEC-04 |
| Design firms | ARCH-nn, ENG-nn | ARCH-02 |
Tokens beat blanks because they keep the links. If OWNER-014 appears on an RFI, a change order and a schedule, a model can still follow one project from question to cost to delay. The map itself never leaves your building.
Add project nicknames, street names and the short names your team uses in emails. Those are what automated tools miss.
Step 2: clean each record type
Each record hides identifiers in different places.
RFIs
Clean the header (to, from, project), the body text and any pasted email chain. Attached sketches carry title blocks; most firms drop attachments and keep the text.
Change orders
Replace the owner block and contract number, delete signature lines, and check the description field, where estimators often name the client’s staff.
Daily logs
Replace the header address, crew lists and visitor lists. Inspector names count. Weather and work-in-place text are usually clean.
Schedules
Activity names are the catch: “Smith kitchen rough-in” names a homeowner. Search activity names and file names against the map.
Payment records
Delete bank, card and account numbers outright. No buyer needs them, and most firms leave pay applications and lien waivers out entirely.
Step 3: strip what you cannot see
Files carry details that never appear on the page:
- Document properties: author, company, last modified by.
- Photo metadata: GPS position, device, timestamp.
- Revision history and tracked changes in Word and Excel files.
- Comments in PDFs.
- File and folder names, which often include the client and street.
Export to a clean format where you can, and rename every file using the project token.
Scanned paper is the hardest case. A scanned daily log is an image, so text search will not find the superintendent’s handwritten name in the margin. Either leave scans out of a first license or run them through text recognition and review every page by eye before they join the cleaned set.
Step 4: check a sample by hand
Automated search catches patterns. People catch meaning.
- Search the cleaned set for every name on the map, plus patterns for phone numbers, emails and ZIP codes.
- Pull a random sample from each system and record type.
- Have a project manager who ran those jobs read the sample in full.
- Fix the misses, update the map, and run the search again.
- Record who checked what, and when.
A sample that comes back clean twice in a row is a reasonable stopping point for a first license.
Where buyers say they de-identify
| Buyer | Published process |
|---|---|
| Scale AI | Cleans on the seller’s own systems, pre-extraction; seller signs off each package |
| micro1 | De-identifies and transforms sensitive information; companies may review representative samples |
| Nyne | De-identifies before release; performer unnamed |
| Labelbox Prism | Handles ingestion, normalization and removal of personal data on its side |
Taken from the buyers’ own sites at our November 7 check. If a contractor we introduced goes on to sign with one of these four, a referral fee may come our way from it; the table would read the same either way. If you are weighing whether to sell construction data at all, the field reference is the shorter version of this post to hand to your project managers.
The best-cleaned record set is one your client could read without recognizing their own building.