Case Studies · Global streaming platform, localization program
Subtitle rejections were running at 30 percent
30→5%
QC rejection rate in two months
Systemic change
A systemic fix: documentation standards, a vendor feedback loop, and a response procedure.
The situation
The client's content operations team was rejecting a large share of the subtitles delivered for episodic and feature titles, peaking at roughly 30 percent. The reasons clustered around inconsistent terminology, misspelled character names, and missed localization instructions, most of them tied to the Key Names and Phrases lists and the glossary of approved localized terms for each show.
Many of my assigned titles were affected. Each rejection meant rework, a delayed delivery, and a response that was often late because no one owned it. The pattern was putting launches at risk and eroding the client's confidence in the account.
The cause
The easy explanation was that linguists were being careless. Before accepting that, I set up shadow sessions with the linguists who were being flagged, across two vendors and several languages, five sessions in all, and watched how they actually worked.
They were following the KNP lists, the glossary, the style guide and the dialogue lists. The problem was upstream. Updated versions of those documents were being uploaded into the content folder with the change log and date written inside the document, not in the filename. A linguist who had already downloaded a copy had no way of knowing a newer one existed unless they went looking. Stale guidance was producing "errors" that were really version mismatches.
Action
I proposed an amendment to the client's Timed-Text Guide, the shared source-of-truth document that linguists and managers across vendors rely on. The amendment added a section on where updated project materials live and required the update date in the filename of every guide revision, so a stale copy is visible at a glance.
I worked with both vendors on an accountability loop built from the client's existing 30-day rejection report, which names each linguist. A flagged linguist reviewed the updated guide. A repeat flag triggered an improvement plan with their vendor. No improvement the following month meant removal from the project, with a path back if the vendor could show satisfactory work elsewhere.
Alongside this, I set up a rejection-response procedure for my workstreams: each rejection received an owner and a response deadline, tracked on a weekly scorecard.
The result
QC rejections fell from roughly 30 percent to under 5 percent within two months. Late or missed SLA responses dropped by about 85 percent within 30 days of the response procedure going live. The client's content operations and localization teams responded positively, and linguists reported fewer misunderstandings and higher productivity because the instructions were finally the same for everyone.
The vendor adopted the centralized documentation approach as its standard for that program, with monthly audits and quarterly feedback sessions to keep the guidance current.
Defects addressed
- —Reference documents carried no version in the filename. Every guide revision is now dated in its filename, and superseded versions are archived out of the working folder.
- —Quality problems were being attributed to people before the work was observed. Shadow sessions are now the first step when a quality metric moves.
- —Rejections had no owner and no clock. A response procedure with named owners now exists before the first rejection, not after the thirtieth.
Lessons learned
Shared reference material needs a versioning standard and a single known location, written into the onboarding guide so every new linguist, vendor or team member starts from the same place. Quality metrics should trigger observation of the work before any judgment of the workers, and every quality program should ship with a response procedure already in place.
This is the kind of situation I consult on under Program Management.
