Skip to content
Outrich Data

Data type

License your QA scorecards and evaluations for AI training

QA scorecards and evaluation records can often be licensed for AI training when the reviewed work belongs to your own business. Outrich Data is the direct buyer, and takes them with agent and customer identities removed and scores detached from individuals. Scorecards kept for outsourcing clients need that client's written consent. You approve each dataset before it is delivered.

What counts as qa data

QA data is what your quality team produces when it reviews work. A reviewer picks a call, chat, ticket, claim file or processed document, scores it against a rubric, marks which criteria passed or failed and writes a comment. The record ties a piece of real work to a grade and an explanation.

A full QA program holds more than scores. It holds the rubric and its past versions, calibration sessions where several reviewers scored the same item, disputes an agent raised and how they were settled, and coaching notes. Evaluations you perform under a client's quality program in an outsourcing contract are that client's records and are excluded without consent.

Why AI labs want it

Training a model takes examples. Improving one takes graded examples. A scorecard says this reply was acceptable, this one missed a required disclosure and this one solved the wrong problem, in the words of a trained reviewer. AI labs use that kind of expert judgment to teach models to assess quality and to build automated graders.

Calibration data is especially hard to find. When three reviewers score the same interaction and disagree, the record shows where a rubric is ambiguous and what the settled answer was. Labs use those cases to measure how closely a model's grading matches human reviewers, and where humans themselves diverge.

Removed before licensing

What comes out first.

  • Agent names and employee IDs, replaced with placeholders that cannot be traced back
  • Reviewer and team lead names
  • Customer identifiers in the reviewed interaction and in comments
  • Coaching notes that touch on discipline, attendance, health or personal circumstances
  • Per-agent score histories that would work as a performance file
  • Client names and program names in outsourced work
  • Card and account numbers quoted in reviewer comments

Data title search

What gets checked.

  • Client ownership. In outsourcing agreements the quality program, scorecards and results commonly belong to the client. Those records need written consent.
  • Underlying interactions. A scorecard is only useful with the call or ticket it grades, so the consent and contract checks for that interaction apply too.
  • Employee records. Individual evaluations are performance data. Check employee notices and, for staff outside the US, local employment and privacy rules with counsel.
  • Rubric authorship. Scoring guides supplied by a client, a consultant or a QA software vendor may not be yours to license.
  • Regulated reviews. Audits of collections, lending or healthcare interactions can carry sector confidentiality duties.

Who signsThe head of quality or customer operations sponsors it. The general counsel approves the license and HR confirms the treatment of evaluation records. The QA platform administrator runs the export.

Value

What moves the price of qa data.

FactorEffect on value
Linked source interactionA score attached to the transcript or ticket it grades is generally worth far more than a score alone.
Written reasoningReviewer comments that explain a mark give a model something to learn from beyond the number.
Calibration recordsItems scored by more than one reviewer show agreement levels, which evaluation work depends on.
Rubric detailLine-item criteria with clear definitions are more useful than a single overall rating.
Balanced resultsA realistic mix of passes, partial passes and failures is more instructive than a set of top scores.

Questions

Licensing qa data: what companies ask.

Are QA scores employee performance records?
Tied to a named agent, yes. That is why agent identity is removed and per-person histories are broken up before a dataset is built. The licensed record is the graded example, not a file on an individual.
We only keep the scores, not the calls. Is that enough?
It limits the value sharply. A grade without the work it grades teaches a model little. If the recordings or tickets still exist in another system, they can often be rejoined by interaction ID during preparation.
Our rubric changed three times. Is that a problem?
No, as long as each scorecard records which version applied. Rubric changes are informative in themselves. Mixing versions without a label is the thing to avoid.
We score calls for a client. Can those scorecards be sold?
Not without the client's written consent. Their program and their customers' interactions are their data under most services agreements. QA records for your own operation are reviewed separately.

Find out if your data qualifies.

Eleven questions, about five minutes, no call required. You get an answer on screen and a written follow-up.

esc