Can a company license its preference and rating data to AI companies?
Often yes, and many companies hold it without calling it that. Labs tune models by showing them two answers and which one a person preferred, or an answer and the score a reviewer gave it. Every time a supervisor scored a support reply, an editor picked one headline over another, or a manager rewrote a draft, the company recorded a judgment of the same kind.
Judgments a company records every day
| QA scores on customer replies | A supervisor rates an agent reply on a scorecard and writes what should have been said. |
| Edits to drafts | The first draft and the version a senior person sent, side by side. The edit is the preference. |
| A or B choices | Two proposals, designs or subject lines, with the one chosen and the reason in a comment or ticket. |
| Customer satisfaction ratings tied to a reply | A score the customer gave right after an answer, linked to the answer itself. |
| Review queues with accept and reject | Content moderation, claims review or listing approval, where each item got a decision and a reason code. |
What turns a rating into usable data
The rating must be linked to the exact thing that was rated. A score with no copy of the reply it judged is useless. Next, the rubric: what a five meant, and whether it changed over the years. Then the rater: role, experience and how often raters agreed with each other.
Written reasons multiply the value. "Too long, missed the refund question" teaches more than a number.
Where these records sit
QA scores usually live in a quality tool beside the help desk, edits in document version history or a content system, and A or B choices in tickets, email threads or meeting notes. Exporting means joining the rating to the item through an ID, which is easier while the systems are still running.
Count how many rated items there are, over how many years, with how many raters, and what share has a written reason.
Privacy and fairness checks
The person who was rated and the person who rated both appear in the record. Staff names are replaced with stand-ins, and scores used in performance reviews are treated with care, since employment rules may limit their use. Customer details in the rated item are removed as they would be in any ticket.
Ratings that are left out
| Scores without the item they rated | A number with nothing to compare it to cannot be used. |
| Ratings tied to disciplinary records | Scores that fed a dismissal or a dispute stay out on counsel advice. |
| Ratings on customer content they own | Reviews of customer-written material follow the terms the customer accepted. |
More on this: how personal details are removed, who carries the risk and how you get paid.
An example, invented for illustration
Take a subscription software company with a hundred and twenty support staff. Its QA team scores four percent of tickets each week on a ten-point rubric and writes a short note. Over six years that is a large set of replies, each with a score, a note and, for low scores, the reply the reviewer would have sent. Agent and customer names are replaced, and the rubric versions are documented by year.
Questions owners ask about this
Our rubric changed several times. Is old data still usable?
Yes, if you record which version applied to each rating. Buyers can treat each period on its own or map the versions to each other.
Is a star rating from customers enough?
It helps when it is tied to one specific reply. A general rating of the company or the product says little about which answer was better.
Do the raters need to be experts?
Their role should be clear. A supervisor with years in the job is worth noting, and a buyer may weight ratings from senior staff more heavily.
Didn't find your question? Ask us directly.
Keep reading
Related kinds of records
- Support tickets: The whole thread, wrong guesses included, with customers replaced by stand-ins.
- Client deliverables and drafts: Each draft beside its final version with the comments that changed it, kept only where the client agrees.
- Evaluation datasets: Exam questions, graded work and checked test cases from real jobs, kept private so they can measure a model.