Use caseContent migration
Procurement is leaving OpenText: 3.8 million files, over half untyped, many in three copies. A straight copy would hand that mess to the AI.
The content is measured before it movesEach file gets a class and its properties on the wayA person reviews every file the AI is unsure about
3.8 million files read in place. 2.4 million carry no category at all, and 410,000 are exact duplicates of another file.
source · discovery scan, Content ServerThe AI is below the threshold · a person confirms or corrects
A search index fed with untyped duplicates returns three versions of the wrong contract, and an assistant grounded in them cites the draft instead of the signed copy. A straight copy moves the problem into the new repository.
The cleanup rides along with the move. The pipeline classifies each file, extracts the properties its class requires and flags what is redundant, and anything below the threshold goes to a person. The files land with a type, a supplier and one main copy.
If you keep one thing from this page, keep the order: measure the content, define what a usable file looks like, fix each file in the pipeline, and send the doubts to a person. Tomas measures, Kevin builds the pipeline, and Keisha settles the doubts before she asks the first question.
A quality profile comes first: it tells the team which stages the pipeline needs and how much review to plan.
Tomas Lindqvist, the content architect, runs a discovery scan on the procurement workspace of the Content Server. It reads formats, volumes, categories and folders in place, and nothing moves.
More than half the files sit in folders called Misc, Scans or Old. The profile becomes the baseline the project will be measured against.
Tomas writes the target model with the procurement team: eight document classes, each with its required properties. A supplier contract needs a supplier, a contract number, a start date and a signed flag.
A file that cannot get its required properties is never loaded with blanks. It goes to review.
Four stages sit between extraction and loading, and each one writes what it did on the file.
Kevin Brandt, the migration engineer, adds a classification stage after extraction. The AI reads each file and proposes one of the eight classes with a confidence level.
Above the threshold Tomas set for this use, the class is applied. Below it, the file goes to the review queue with the AI's proposal attached.
Once a file has a class, the next stage extracts that class's properties: supplier name, contract number, dates, amounts. A business rule normalizes the supplier name to the spelling procurement uses.
Exact duplicates were counted in the profile. The AI also flags redundant content the scan cannot see: the same contract scanned twice, a Word draft next to its signed PDF. It flags personal data in a scan at the same time.
The rule keeps the signed copy as the main document. Every other copy is set aside and listed in the report, with its original path.
Keisha Brooks, from procurement, opens the review queue every morning. Each file shows the AI's proposal and what is missing. She confirms SCAN_0412_0004 as a supplier letter, picks the supplier, and marks two old fax covers as not worth keeping.
After each batch, the team checks the result against the baseline, then tries the content the way users will.
Kevin runs the Scans folder in batches. The counts per stage show where files leave the main path: set aside as redundant, sent to review, excluded by a reviewer. Every exit has a reason that can be listed and exported.
Tomas opens the quality report. It uses the same measures as the profile in the first act, so the two read side by side: files with a class, required properties filled, redundant groups settled. He sends it to the AI project lead with the files still open in review.
Keisha tries the result the way a buyer will. She asks the assistant for the handling rate signed with Haulmark Freight. The answer cites the signed contract and its page, and reads only what she is allowed to see.
When a file leaves the pipeline, one of three things happens to it: it lands, it waits for Keisha, or it stays behind.
The AI is above the threshold and every required property has a value.
Next: the file lands in its class, with its trace in the report.A low confidence level, or a required property it could not find.
Next: a reviewer confirms or corrects, and the file runs again.A copy of the main document, a draft, a fax cover.
Next: listed in the report with its path and the reason, decided by a person.It deletes nothing. It does not write the classes, and it does not apply a class it is unsure of. Across the nine steps on this page, every doubtful file is settled by a person, and the decision is recorded on it.
The AI classifies, extracts and flags. The threshold and the required properties are the team's, so the rule holds when the volume is millions of files.
The Scans folder, 1.1 million files on its own, is where these counts come from. Your content and your rules will set yours.
Where these numbers come from. They are counted on the scenario described on this page, with example data. They are not a customer result.
Before migrated content feeds an AI project, the AI lead, the content architect and the people who know the suppliers each have to trust it. What they get:
The AI project needs clean content, and the migration is about to copy the old mess.
Design a model for files that carry no metadata at all: that is your brief.
If the assistant quotes the wrong contract, it lands on your desk.
With one folder. Take the one that worries you most, profile it, and run it through the pipeline as a pilot. One folder is enough to watch the numbers move.
The source is read in place during the profile, and nothing lands in the new repository until the batch report has been read.
Next step
Point us at a part of your Content Server that nobody can search. We will profile it and show how much the pipeline can classify and enrich, and how much a person would review.