Kluup scores named employees, so the people whose work it judges have rights that are not negotiable and are not a paid tier. Everything below is published rather than answered on a call, including the parts that are not settled yet.
roles, and the instructions we act on
We process on your documented instructions, and your instructions are the things you do in the product.
You decide whose calls are recorded, why, and what is done with the result. We act only on the instructions you give through the product: connecting a source, uploading a recording, configuring a scorecard, inviting a user, setting a retention window, deleting a record.
We do not benchmark across customers. No data crosses a tenant boundary for any purpose, including our own improvement. Nothing you upload, and nothing derived from it, is used to train, tune, evaluate or improve anything for anybody but you. That is also why Kluup has no industry benchmark to sell you.
two data subjects, treated differently on purpose
We hold the identifier your CRM uses for that person and nothing else about them. Records are keyed on that identifier and never on a name. There is deliberately no field for a buyer's name. The person lives in a CRM you already have a lawful basis and a subject-access process for.
Reps are named, because the scorecard is about them. That is exactly why the rep's own view, the evidence behind every number and the dispute route are included in every account and cannot be switched off.
From an integration we take the audio and nothing else. We transcribe each channel on our own hardware, hold the transcript in memory for the length of one analysis, and never store it. We never ingest a third party's transcript.
Special category data is not sought, and one residual is worth stating. A lead may mention health, religion or a political opinion unprompted, and a short verbatim quote tied to a scored moment could capture it. The scorecard judges sales behaviour and has no reason to quote such material, but there is no detection step — and there realistically cannot be one that does not also destroy the quotes the product exists to produce.
who else touches your data, and exactly what they get
| Anthropic | Analysis — scoring a call against your scorecard, and the breakdown Receives transcript text only, never audio. Batched requests are deleted from Anthropic as soon as their results are collected, measured at about three minutes rather than the 29-day default. Processed in the United States, a transfer outside the UK under a documented transfer mechanism. |
|---|---|
| Hosting provider | Running the service and the database Receives every record Kluup holds, encrypted in transit and at rest. Not chosen yet. The agreement carries this row with the name left open. |
| Object storage | Holding call audio for the retention window below Receives audio files keyed by an opaque path. Not chosen yet either, and open in the agreement the same way. |
Transcription is not sub-processed. It runs on infrastructure we operate, which is the reason a CRM's transcript is never good enough: we need each channel separately.
defaults per workspace; an Owner may shorten any of them
| Call audio | 90 daysFrom the call date, not from upload, so restoring a backup cannot silently extend it. This one cannot be lengthened past 90 days by anyone, including us. |
|---|---|
| Scored records | 24 monthsThe judgements and quotes behind each score, so history stays comparable when your scorecard changes. |
| Job history | 90 daysWhat happened to a recording, how long it took, what it cost. No transcript. |
| Billing ledger | 7 yearsThe UK statutory floor for accounting records. |
The sweep is scheduled, and each run is recorded, so deletion can be proven by query rather than asserted in a paragraph.
The 90-day window has a consequence, and the product owns it. Re-analysis after you change a scorecard dimension runs from the audio we still hold. A call whose audio has passed the window cannot be re-scored, and the rubric editor says so before you make the change rather than inventing a score for it afterwards.
Deletion is real. You can delete by contact, by rep, or the whole organisation, from inside the product. Each leaves no residue in the database, in storage or in the ledger — except that the billing ledger keeps the money under an opaque reference for the period above, because accounting law requires it.
included for every customer, and not switchable off
Kluup neither makes nor recommends employment decisions. It produces scores, each traceable to the words that justified it, and a person reads them.
If you use a score in a performance review, a bonus calculation, a promotion or a dismissal, that is a decision with legal or similarly significant effect and it is yours. You are responsible for meaningful human involvement in it, for explaining the logic to the person affected, and for giving them a route to contest it. The two rights above exist so that you can.
And the gate verdict describes the lead, not the rep. A call that reached nobody worth selling to is not a poor performance, and the product keeps the two numbers apart so that a manager cannot accidentally read one as the other.
product red lines, not settings
Article 5(1)(f) of the EU AI Act prohibits inferring a person's emotions in the workplace and has been in force since February 2025. How a voice sounds is never scored — not its tone, not its pitch, not how loud it was — and Kluup never infers stress, frustration, enthusiasm or confidence from it, for any customer, in any jurisdiction. This is a red line in the product, not a switch an administrator can flip.
What a recording is used for. Two things. Its words, and when each was said, which is what transcription produces. And, where we can read one, the loudness trace drawn behind the call timeline so you can see where the call was busy and where it went quiet — that trace is there for a person to look at and reaches no score, no scorecard and no rating of a rep.
What that does not cover, stated precisely because the boundary matters. Scoring the content of what was said against your scorecard is not emotion inference — judging whether a discovery question was asked is a judgement about words. Nor are the turn-taking figures: question counts, longest monologue, talk ratio and patience are counted from who spoke when. Where we transcribed your recording that clock is the recording's own; where a call reaches us with no recording they are worked out from the transcript and each figure says so — estimated from the transcript, not measured. Both are arithmetic over words and turns, and neither is capable of a statement about how anybody felt.
Only the audio, transcribed per channel on our own hardware. It is the only way “who said this” stays a fact.
Records are keyed on your CRM's identifier. There is no field for it to go in.
It exists in memory for the length of one analysis and is gone. The audio is what is retained, for 90 days.
enforced in the database, tested on every build
Every record carries your organisation's id, and the database itself refuses to return one organisation's row to another.
The rule is a database policy, not an if statement in application code, so an
application bug cannot leak a row. Every build runs a suite that holds one organisation's
real credentials and attempts to read, update and delete every table, every stored file and
every procedure belonging to another — by guessing identifiers, by forging filters, and by
calling the database directly around the app. Every attempt has to return nothing or be
refused. A single leak fails the release outright.
Reads of a scored call are audited: who read whose call, and when. Our staff do not read your calls, and there is no support path that needs to.
published, because you would find out anyway and should find out from us
Annex III of the EU AI Act treats AI used to evaluate workers as high-risk. Whether Kluup is a high-risk system for EU deployments, and in which role, has not yet been assessed by a practitioner. For a UK-only customer the Act does not apply. For an EU customer it is a live question, and we would rather you had it in writing here than raise it and be told it was already handled.
The prohibition that is already in force — workplace emotion inference — is a red line Kluup was built on and does not go near.
Kluup scans its own storage for personal data by pattern, so it catches email addresses, phone numbers and postcodes — and deliberately never catches a name, because the rubric keeps place names and company names on purpose and a name blacklist would destroy the data the product is built on.
A clean run therefore means “no detectable identifiers”, not “no personal data”. We will not describe it any other way.
Every clause in it is derived from Kluup's own data protection impact assessment and invents no obligation that assessment does not already assume, but the wording has not yet been finalised by a practitioner, and the impact assessment itself is awaiting review and signature. Both are in progress rather than done.
Read the current text — it is published as it stands, not summarised.
The table above names Anthropic, the one that receives your transcript text. Hosting and object storage are still open: the agreement carries both rows with the name marked as still to come, and this page says the same. Neither document is ahead of the other, and nobody has been picked in private.
When they are chosen, both go in the table and in the agreement, and you get notice before the change takes effect.
Kluup's company number, registered office and contact address are not settled, so the footer of every page shows the gap instead of something plausible. That matters most to you: a data-protection officer needs a name and an address to send a request to.
All three appear here before this site goes public. The check that publishes it fails while any one of them is missing, so it cannot go out half-answered.
It is the same text presented in the product when a workspace is created, and it applies to every customer at no extra cost.