Data Processing Agreement
Controller ("Customer"): [Your organization]
Processor: Auspex Streamline S.L., C.I.F.: B56341829, Calle Velarde 13, 4B, 35010 Las Palmas de Gran Canaria, Canarias, Spain ("Aelo")
This Data Processing Agreement ("DPA") forms part of the Terms of Service between Controller and Processor for the use of Aelo — AI Communication Analyzer ("Services").
1. Definitions
- "Data Protection Laws" — GDPR (Regulation (EU) 2016/679), and applicable national data protection laws including CCPA, UK GDPR, LGPD.
- "Personal Data" — Any information relating to an identified or identifiable natural person, processed by Processor on behalf of Controller.
- "Sub-processor" — Any third party engaged by Processor to process Personal Data on behalf of Controller.
- "Standard Contractual Clauses" ("SCCs") — The clauses approved by European Commission Implementing Decision (EU) 2021/914.
- "Pro Trial" — The features of the Pro plan that a Controller on the free (Pay-as-you-go) plan receives once, for 14 days from its next analysed call or chat (Terms of Service, section 9.6). While its Pro Trial lasts, the processing this DPA describes for the Pro plan and above also applies to that Controller; when the trial ends, that processing stops for new conversations, and what it already produced is retained and deleted under section 9.
2. Scope and Purpose of Processing
| Processing Activity | Data Categories | Purpose |
|---|---|---|
| Audio retrieval from a Controller-designated URL | Audio recordings (voice data) hosted at an address the Controller supplies; the URL and its hostname | On Controller's instruction given through the public API, Processor initiates an outbound HTTPS connection to the host the Controller designates and downloads the recording for the processing activities below. The Controller selects the host and controls what it serves; Processor resolves the hostname through Cloudflare's DNS-over-HTTPS resolver before connecting |
| Audio transcription | Audio recordings (voice data), caller metadata | Convert speech to text using ASR providers |
| AI analysis | Transcripts, conversation metadata | Quality scoring, sentiment analysis, key point extraction |
| Evaluation of the quality of the analysis | Stored transcripts of samples of conversations, with pattern-based PII redaction applied (names are not removed), and, depending on the measurement, their stored analysis or answers and the wording of the criterion or instruction being measured | Measure how accurately the analysis answers: an AI model labels the transcripts of a sample without seeing the analysis’s answers (blind labelling) or checks the stored analysis against its transcript, and the analysis a candidate model or wording produces is compared with the current one. Made with the LLM sub-processors listed in the Sub-Processor List (OpenAI; Anthropic from 2026-12-03). The results are used to decide changes to the analysis, are not shown to Controller and are not used to make decisions about data subjects |
| Customer questions and decline reasons | For a Controller with the Voice of Customer section with its customer questions and decline reasons (Pro plan and above): transcripts with pattern-based PII redaction applied; for each conversation analysed after this was introduced, the questions the customer asked (a one-sentence summary written by the model, the customer's verbatim words from the redacted transcript, for a call the second at which they were said and for a chat the number of the message they were written in, the question's subject from a fixed list, whether the employee answered it and at which stage it was asked) and, for a sales conversation, how it ended and, when the customer declined, the reason from a fixed list, in the customer's words with a verbatim quote and whether the customer stated it or the model inferred it | Collect what customers ask and why they decline across the Controller's conversations for the Voice of Customer section, in a request separate from the analysis, not charged for by the Processor; the results are stored with the record and shown in that section to the Controller's users who may see it |
| FAQ draft | For a Controller with the Voice of Customer section with its customer questions and decline reasons (Pro plan and above), on a Controller user's request: the most frequent question subjects of the period and projects the user selected, the number of conversations each was asked in, and up to 40 of the stored one-sentence summaries of those questions, masked again for phone numbers, emails, card numbers and IBANs before they are sent (a summary that may contain an identification number is left out); no transcript, quote, record identifier or person | Write, in a request separate from the analysis, a draft FAQ for the Controller's website — questions and answer templates for the Controller to complete and verify — in the language set for the Controller's organization; the draft is shown to the requesting user and not stored; not charged for by the Processor; at most 10 drafts per organization and 5 per user in 24 hours |
| Conversation themes | For a Controller with conversation themes (Pro plan and above, together with the Voice of Customer section with its customer questions and decline reasons): the names (up to 60 characters) and short descriptions (up to 80) of the Controller’s active themes, as its owner or administrators wrote them, at most with phone numbers, emails, card numbers and IBANs masked, sent with the request that finds customer questions, together with the transcript with pattern-based PII redaction applied; on the request of the Controller’s owner or an administrator, for a starter set: the distinct pairs of a short conversation-topic phrase (from the stored analysis) and the call category of the Controller’s conversations of the last 90 days, up to 1,500 pairs and 60,000 characters, without demo conversations; once a week, for a Controller whose set has an active theme and that had newly analysed conversations in the last 7 days (and, for a Controller connected to Uspacy, while its Uspacy processing instruction is active): the suggested-theme phrases stored for its conversations of the last 30 days, up to 300 distinct phrases, without demo conversations. The phrases are checked again for phone numbers, emails, card numbers and IBANs before they are sent (a phrase in which the check finds one, or that may contain an identification number, is left out); no transcript, quote, record identifier or person is sent in the starter-set and weekly requests | Assign to each conversation one to three themes of the Controller’s set (or that none fits) and, where the conversation has a subject the set lacks, a short suggested theme (up to 60 characters), in the request that finds customer questions, which is separate from the analysis; propose, on request, up to 25 themes with descriptions as a starter set that the Controller’s owner or administrators edit and approve (only approved themes enter the set); and, once a week, group the suggested-theme phrases into at most five candidate themes, those that came up in at least 10 conversations and 3% of the period’s conversations being shown to the Controller’s owner and administrators, with counts and up to three example conversations, to accept or reject. A suggested theme, name or description the model returns in which the same check finds such data is not stored. A new starter-set proposal at most once in 24 hours per organization (an hour after a failed attempt). Not charged for by the Processor; conversations analysed before a theme entered the set are not processed in bulk |
| Deal state | For a Controller with the deal state (Pro plan and above), for each newly analysed call or chat linked to a deal or lead in the Controller's connected CRM (Bitrix24 or Uspacy): the transcript with pattern-based PII redaction applied (applied again before it is sent), a summary of the stored analysis, the previous state of the deal or lead (for a Bitrix24 or Uspacy deal that has no state of its own yet, converted from a lead that has one in the same project, the lead's state as of the deal's first conversation), the sales methodology of the deal's pipeline in the connected CRM (a Bitrix24 pipeline or a Uspacy funnel) with the Controller's company rules for it (pattern-redacted, names survive), the number of days since the previous conversation and, for a Bitrix24 deal or lead, its stage, amount and currency, and, for a Uspacy deal, its stage identifier; read from the Controller's Uspacy, to find the deal's funnel and stage: the deal's current stage identifier and the portal's list of deal funnels with their stages (no snapshot of a Uspacy deal is stored with the state, and its title and amount are not read; the stage records of the deal-stage progression evaluation are kept separately); read from the Controller's Uspacy for a deal or lead and stored with its states, but not sent to the AI model: the identifier of its responsible employee, only where the instruction accepted by the Controller is version 1.17 or later; read from the Controller's Uspacy for a deal and stored with its states, but not sent to the AI model: of the leads linked to the deal, only their identifiers and whether Uspacy marks a lead as converted are used, and the identifier of the one converted lead is stored; read from the Controller's Bitrix24 and stored with the stage, amount and currency as a snapshot, but not sent to the AI model: the title of the deal or lead, its pipeline, whether its stage is final and its responsible employee's identifier (a stage identifier can carry its pipeline's number); read from the Controller's Bitrix24 for a deal and stored with its states, but not sent to the AI model: the identifier of the lead the deal was converted from. The resulting state (see section 3) and, for the CRM writes below, its text | Keep the current state of each deal or lead — status, chance of closing, reason with a verbatim quote, next action, commitments, objections, qualification, interest — and its history across the conversations linked to it, built by the AI model in a request separate from the analysis (not for a short call or on re-analysis; not charged for by the Processor), and shown in the Controller's deal queue and deal card. Write the state back to the Controller's CRM: a comment on the timeline of the deal or lead after each new state, in the CRM that receives the Controller's call reports (Bitrix24, or Uspacy while the Controller's Uspacy processing instruction is active), only for a conversation held within the last 7 days and unless the Controller switches it off; and, on a Controller user's request, a to-do in the Bitrix24 deal or lead, assigned to its responsible employee, carrying the next action, its reason, its due date and, except when it is created from the Aelo tab in the Bitrix24 card, a link to the deal in Aelo, or, in Uspacy while the Controller's Uspacy processing instruction is active, the same as one planned task (a CRM activity) in the deal or lead, assigned to its responsible employee where the latest state knows one and otherwise without a responsible employee set by Aelo (Uspacy then assigns it by its own default), after reading the activities of that deal or lead only to find a task already added for the same state |
| Draft message to the customer | For a Controller with the deal state (Pro plan and above), on the request of a Controller user who may see the texts of the deal's state: the current stored state of the deal or lead — status, chance of closing, the reason for the status, the next action, points for the next contact, what the latest conversation changed, commitments, objections by reason code, what is known of budget, decision maker, need and timeline, and the customer's interest — with those texts masked again for phone numbers, emails, card numbers and IBANs, and the sales methodology of the deal's pipeline (the standard one and the Controller's company rules for it, masked again); no transcript, quote, conversation or record identifier, title of the deal or lead, amount, stage or responsible employee | Write, in a request separate from the analysis, one short follow-up message to the customer in the analysis language of the deal's project; the draft is shown to the requesting user, who can edit and copy it, and is not stored or written to the Controller's CRM; not charged for by the Processor; at most 5 requests per minute, 60 per hour and 200 per 24 hours per organization and 3, 30 and 60 per user |
| Criteria (the Controller's own questions and the Processor's built-in ones) | The Controller's criteria (name, question, guidance on what counts, answer values — the name is stored only; the question, guidance and answer values are sent to the AI model as the Controller wrote them, at most with phone numbers, emails, card numbers and IBANs masked; the description of a criterion a Controller’s user writes for a draft wording and the reasons the user gives for answers marked as wrong, sent the same way and not stored) and transcripts with pattern-based PII redaction applied; for each answer, the value (yes or no, a 0–100 score or one of the Controller's values, or that no confident answer was possible), verbatim quotes from the redacted transcript, the model's short written explanation and, when known, the identifier of the employee who handled the conversation. Built-in criteria are worded by the Processor (with the Controller's jurisdiction where it applies) and are the only place where the checks of what was or was not said (for example, whether the company was named, a next contact was agreed, call recording was disclosed or the employee used profanity) are made — the analysis itself no longer performs them; the Controller pauses them and narrows where they apply in its criteria settings, within the limits the Processor sets, and the Processor may pause one (where the Processor has switched built-in criteria off for the Controller, they are not shown and the Controller has none of these checks); they are not charged for by the Processor and are not asked for a short call that receives no analysis, or when an already analysed conversation is re-analysed (a failed analysis that is retried asks them as usual) | Answer the Controller's questions about each conversation within a criterion's scope, in a request separate from the analysis, for display in the call panel, the PDF report, the records filter and the dashboard report; and, on the Controller's request, put a draft wording to up to 20 recent conversations within its scope, showing the answers without storing them; on the request of a Controller’s user, write a draft wording from the user’s own description of a criterion, and propose an edit of a draft wording from the answers of its check the user marked as wrong — with the user’s optional reasons and, for each marked conversation (up to 10), a fragment of its redacted transcript around the quoted lines, redacted again before it is sent — showing the draft and the proposed edit to that user without storing them. Built-in criteria: make the checks of what was or was not said, which the analysis no longer performs, and show their answers in the call panel and the PDF report, the dashboard, the attention flag, alert rules and the report written to the Controller's CRM |
| Semantic search | Segments of call transcripts and chat conversations, and the text a Controller's user types into search, both with pattern-based PII redaction and contextual masking applied (Section 4, “Pattern-based PII redaction”); the numerical vectors computed from the segments, each with the identifiers of the record, project, organization and transcript it belongs to, the record type (call or chat) and the segment's length, and — when known — an identifier of the employee who handled the conversation (their user ID in Aelo, in the connected CRM or telephony, or the value supplied through the public API) and the record's date | Let the Controller's users find records by meaning. Segments and search text are sent to the LLM sub-processor to compute the vectors; only the vectors of the segments are stored (Cloudflare Vectorize), without transcript or chat text, and the text shown with a result is read from the stored transcript, masked the same way. The vector of a search query is used for that search only and is not stored |
| AI assistant (interactive questions and answers) | A Controller user's question about a call record or about a CRM deal, lead, contact or company (with pattern-based PII redaction and a check by the LLM sub-processor for Personal Data it missed applied before storage), and the earlier questions and answers of the same conversation; the transcript and analysis of that record, or of the records linked to that CRM entity that the user is permitted to see, together with the entity's name and, for a deal, its outcome in the CRM where known (won, lost or still open); or a question about all the records the user is permitted to see, with the filters chosen (the period, record type, category and customer name are also sent with the question; projects and agents are not), its embedding (used to find matching records in the search index and not stored) and the transcripts and analyses of up to 12 of those records; and the generated answer with the excerpts it cites | On the request of a Controller's user, generate an answer about a single call record, about a CRM entity across its linked records, or about all the records the user is permitted to see, grounded in their transcripts and analyses via the LLM sub-processor, and keep the conversation for the user to return to; before the question is stored, the question, before the answer is generated, the transcripts and analyses it is generated from with the earlier questions and answers of the conversation, and before it is stored, the generated answer with the question, are sent to the LLM sub-processor in separate requests to find Personal Data the pattern-based redaction missed, which is masked |
| Acoustic voice metrics | Audio recordings (voice data) | Derive objective voice-delivery metrics (speaking rate, pauses, pitch variation, voice stability) via Aelo's self-hosted acoustic service |
| Data storage | Audio, transcripts, analyses, metadata | Secure storage for Controller's access and retrieval |
| CRM delivery | Analysis results, scores, key points; the call transcript — attached to the CRM call record where the Controller switches that on, posted to the Smart Process item timeline where delivery targets one, and written into a Controller-designated CRM field only where the Controller switches that on (off by default); for a Controller with the deal state, the comment about the state of a deal or lead and the to-do for its next action described under "Deal state" | Deliver insights to Controller's CRM system |
| Stage-based scorecard selection | CRM deal stage ID, pipeline/funnel ID, stage name, and stage phase (from Controller's connected Bitrix24 or Uspacy; where the CRM gives no phase — in Uspacy only a funnel's closing success or fail stage gives one — the phase Controller set for that stage or, failing that, the phase the AI model (OpenAI) determined from the names of Controller's pipelines and stages, which are sent to it, only while stage-based evaluation is on and without the stages Controller has given a phase itself or any data subject or conversation data, when a stage is new or has been renamed, or an earlier attempt gave no answer for it — the request then contains the names of all of Controller’s pipelines and stages) | When Controller connects Bitrix24 or Uspacy and enables stage-based evaluation, select the evaluation scorecard matching the deal's current pipeline stage |
| Deal-outcome tracking | CRM deal/series identifier, pipeline ID, stage ID, outcome status (open/won/lost), close timestamp, sync timestamp (from Controller's connected Bitrix24 or Uspacy; in Uspacy, won or lost on a funnel's closing success or fail stage); for a Controller that also has the Voice of Customer section with its customer questions and decline reasons (Pro plan and above), on a lost deal the code of the reason the customer gave in the most recent conversation of the deal's current loss cycle that ended in a decline (a code from a fixed list, without the customer's words or quote), the version of the code list, and the identifier and time of that conversation, taken only from conversations held after this was introduced | When Controller connects Bitrix24 or Uspacy (while its Uspacy processing instruction is active) and enables stage-based evaluation, periodically sync each linked deal's outcome status to show a deal-outcome funnel — call volume by won/lost, broken down by average call-quality score — in Controller's dashboard |
| Deal-stage progression evaluation | CRM deal stage history (transitions/timestamps) from Bitrix24, or, for Uspacy, which keeps no stage history, the stage and pipeline/funnel identifiers recorded at each scheduled re-check with the times the deal was first and last seen on that stage; derived progression result (advanced/held/regressed) stored per call record | When Controller connects Bitrix24 or Uspacy (while its Uspacy processing instruction is active) and enables stage-based evaluation, once per call compare the deal's stage at the call against its stage ~7 days later (or at deal close) to derive whether the deal advanced, held, or regressed, shown as a per-call progression indicator in Controller's dashboard; for Uspacy the later stage is the one the deal was seen on when the ~7-day window ended, and a call whose window ended at a time the Processor had not seen the deal at (before stage-based evaluation was switched on, or between two re-checks) gets no result |
| Org-structure import | Department tree (identifiers, names, parent links, sort order); department membership per employee, including employees with no Aelo account yet (stored as the portal's employee ID, email when available, and name, until their account is created); and the department head named by the portal, including a head with no Aelo account yet (stored as the portal's identifier and name) (from Controller's connected Bitrix24 or Uspacy; for Uspacy, only while Controller's Uspacy processing instruction is version 1.22 or later) | When Controller runs the org-structure import, reproduce Controller's department hierarchy inside Aelo — including employees and a department head who do not yet have an Aelo account — so a department head can open the call records of everyone below them in the tree, independently of project assignment. When such an employee's or head's Aelo account is later created (for example when an invitation is sent to them, or through their first sign-in inside the connected CRM's embedded app), Aelo automatically links them to their stored portal identity and, when it can identify a new head's account by the portal's own identifier, grants that head access to the department's records, unless Controller’s organization owner has manually assigned a different head. A head assigned by hand in Aelo always takes precedence over the portal's value and survives later imports |
| Coaching feedback | Notes about an employee written by the Controller’s supervisors, administrators and owner (optionally naming one of the employee’s calls), coaching sessions they schedule with an employee (the scheduled time, a status and notes, optionally naming a call and, for a session scheduled from an alert, that alert), annotations on a passage of a conversation’s transcript and reviews of a conversation’s analysis, each with its author and time; for the notification of the employee, the author’s name, the kind of feedback, a link to the conversation and, for a call, the moment — never the feedback text and no data of the other party to the conversation | Store the feedback and show it to the Controller’s users permitted to see it and to the employee concerned; for a note, an annotation or a review (never for a coaching session), notify that employee in the app and, unless the employee has switched these emails off, by email through the email sub-processor; and, where the employee is linked to a user of the Controller’s connected Bitrix24 portal, by a Bitrix24 notification to that user with a link that opens Aelo inside the portal (one email and one Bitrix24 notification per author and conversation within 30 minutes). Not sent to the AI model |
| Email reports and digests | Recipients' email addresses — including addresses the Controller configures for a project that need not belong to platform users — and the content derived from calls and chats that each email carries, listed per email in the Sub-Processor List (Resend): quality metrics, AI-assigned categories, masked caller references, model-written alert reasons and, in role digests, named employees' quality scores. No transcripts or audio. Account, invitation, billing and minute-usage emails are Aelo's own, as controller (Privacy Policy B.1, B.4), and are not part of this activity. | Send analysis reports to the addresses the Controller configures for a project, and alerts, record notifications and digests (digests on by default, switchable by each user) to the Controller's platform users, through a transactional email provider (Resend) |
| Alerting | Analysis metrics and thresholds; personal email and webhook notifications may include a masked caller reference (first name with surname initial, or the last 3 digits of the phone number) and an AI-generated reason when present. Telegram receives only business alerts, in a contact-free version with neither the caller reference nor the AI-generated reason; no other notification type is sent to Telegram. Project team channels receive a reduced summary with no identity for either the other party to the call or the platform user associated with the record, no phone number, transcript excerpt or AI-generated reason. | Notify Controller's platform users about significant events: to each individual user's own configured channels (email, a personal Telegram chat, a personal webhook), and — independently — to a project-level team channel (a Telegram group, Controller's own connected Bitrix24 group chat, or a Controller-configured webhook) Controller sets up |
| Onboarding communications | Not Customer Personal Data processed on the Controller's instruction. Aelo's own onboarding and marketing emails go to a person who gave separate marketing consent, with Aelo as controller (Privacy Policy B.5); what is sent to the marketing-automation provider (SendPulse) is listed in the Sub-Processor List and Privacy Policy B.5, and is sent only for that person | None on the Controller's behalf: listed so that the Controller can see every provider that receives a user's data, and is not part of this activity |
| Application error tracking and monitoring | Platform user's name, email address, and account identifier; the connecting IP address at the network level; page and API request URLs; breadcrumbs of recent navigation and network activity; error messages and stack traces, which can incidentally include Personal Data being processed at the time of the error; session replay recordings for a sample of user sessions (approximately 10%) and for every session in which an error occurs (on-screen text and form inputs masked, media blocked, by default) | Detect, diagnose, and resolve errors and performance issues in the Aelo web application and its API layer. Not applied to the background call-processing pipeline |
Processor shall not process Personal Data for any purpose other than as documented in this DPA and Controller's instructions.
3. Data Categories and Data Subjects
Data Subjects: Call and chat participants (customers, prospects), call agents (employees, contractors), platform users (staff with Aelo access), recipients of analysis reports at email addresses Controller configures for a project, and — where Controller runs the org-structure import — employees appearing in Controller's imported department tree, including employees and a department head who do not yet have an Aelo account. Employees from the connected portal with no matching Aelo account are displayed to Controller's administrator by name so the gap can be resolved; Processor stores their portal employee ID, email (when the portal provides one) and name, together with their department membership, until an Aelo account for them is created, at which point Processor automatically links the two and deletes this interim record. Processor does not create an Aelo account for them on its own.
| Category | Examples | Sensitivity |
|---|---|---|
| Voice data | Audio recordings of calls | High (potential biometric) |
| Conversation content | Transcripts, spoken words | High |
| Communication metadata | Phone numbers, call duration, timestamps | Medium |
| Analysis results | Quality scores, sentiment, key points | Medium |
| Coaching feedback | Supervisors’ notes about an employee, coaching sessions scheduled with an employee and their notes, annotations on a conversation’s transcript and reviews of its analysis, with their author and time | Medium |
| Customer questions and decline reasons | Per conversation: up to 10 questions with the model's summary, the customer's verbatim words, subject, whether answered and stage; for a sales conversation, its outcome and the decline reason with the customer's words and a verbatim quote | Medium |
| Conversation themes | The Controller’s set of themes (names, short descriptions, active or archived) with a record of who changed the set and when; per conversation: one to three themes of the set or a mark that none fits, and a short suggested theme where the set lacks the subject; for a starter set the Controller asked for, the themes the model proposed and the list of topic phrases with call categories that was sent; for the weekly suggestions, each candidate theme (name and description written by the model), the number and share of conversations, the suggested-theme phrases it was grouped from and the identifiers of up to three example conversations | Medium |
| Deal states | Per deal or lead, one state after each linked conversation, kept as its history: status, chance of closing, what changed, the reason in the model's words with the customer's verbatim quote from the redacted transcript (and, for a call, its second), the next action with due date and reason, points for the next contact, commitments (who, what, by when, status), objections (reason from a fixed list, status, the customer's last verbatim quote), what is known of budget, decision maker, need and timeline with a short fact, interest, the methodology applied, identifiers of the source conversations and, for a Bitrix24 or Uspacy deal converted from a lead, the identifier of that lead; whether the CRM comment was posted and the CRM to-do created, with the to-do's identifier | Medium |
| CRM deal snapshot | Stored with each Bitrix24 deal state: the deal's or lead's title, amount and currency, pipeline and stage, whether the stage is final, the responsible employee's identifier in the CRM, and when it was read | Low to medium (business data about the deal; a title can contain a person's or company's name) |
| Criteria answers | Per criterion (the Controller's own and the Processor's built-in ones): yes or no, a 0–100 score or one of the Controller's values, verbatim quotes from the redacted transcript, the model's short explanation, and the employee's identifier when known | Medium |
| Search vectors | Numerical representations (embeddings) of redacted transcript and chat segments, with record, project, organization and transcript identifiers and the record type, and — when known — the employee's identifier and the record's date; no text | Medium (derived; not designed to be read back as text, but may reveal some of the content they were computed from) |
| Voice-delivery metrics | Speaking rate, pauses, pitch variation, voice stability | Medium (derived acoustic measurements; not voiceprint/biometric identification) |
| User account data | Names, email addresses of platform users; the action log of their security-relevant actions (with the IP address and browser user agent of the request for security events and changes to a project’s quality weights) and the access log of changes to their access (see section 9) | Standard |
| Report recipients | Email addresses Controller configures for a project to receive analysis reports | Standard |
| CRM org-structure data | Department identifiers, names, parent links and sort order; which employees belong to which departments, including employees with no Aelo account yet (portal employee ID, email when available, and name); the department head named by the portal, including a head with no Aelo account yet (portal identifier and name); and the pairing of a portal user ID with the matching Aelo account, established by the portal's own employee ID or by email address (for Uspacy, only by the employee's own link from the portal; an employee matched by email address is stored as a department member with the portal user ID) | Standard (employment/organisational data about Controller's own employees; determines who may open which call records) |
| CRM deal-stage data | Deal stage ID, pipeline/funnel ID, and stage phase (e.g. discovery, negotiation) read from Controller's connected CRM or, where the CRM gives none, set by Controller for the stage or determined by the AI model from the names of the pipelines and stages in that CRM (the phase is kept per stage with a hash of the stage’s name, not the names); also includes the deal's stage history (transitions/timestamps) read to derive a per-call progression result (advanced/held/regressed) and, for a CRM without stage history (Uspacy), the Processor’s own record of the stage and pipeline/funnel identifiers with the times the deal was first and last seen there | Low (business-process metadata about the deal, not the data subject) |
| CRM deal-outcome data | Deal/series identifier, pipeline ID, stage ID, outcome status (open/won/lost), and close/sync timestamps read from Controller's connected CRM; on a lost deal, where the Controller has the Voice of Customer section with its customer questions and decline reasons, the decline-reason code of its most recent declined conversation, with the code-list version and that conversation's identifier and time | Low (business-process metadata about the deal, not the data subject); Medium for the decline-reason code, which is derived from what the customer said, although it is only a code from a fixed list without the customer's words |
| Application diagnostic data | Error messages, stack traces, page and API request URLs, breadcrumbs of recent navigation and network activity, connecting IP address, session replay recordings; platform user's name, email address, and account identifier, captured by the application error-tracking sub-processor | Medium (can incidentally include Personal Data being processed at the time of an error; also includes platform user identity data and connection IP address) |
4. Obligations of the Processor
- Documented Instructions. Processor shall process Personal Data only on documented instructions from Controller.
- Confidentiality. All personnel authorized to process Personal Data have committed to confidentiality.
- Security Measures (Art. 32). Processor implements appropriate technical and organizational measures:
| Measure | Implementation |
|---|---|
| Encryption in transit | TLS 1.2+ for all API communications |
| Encryption at rest | Cloudflare R2 SSE (AES-256-GCM, active by default); Neon database native encryption; provider API credentials (speech recognition, AI analysis, email delivery) and the credentials of connected telephony and CRM integrations additionally sealed by the application with AES-256-GCM and bound to their own record |
| Access control | Role-based access (RBAC) with organization isolation |
| Tenant isolation | Organization-level data segregation; cross-org detection |
| Authentication | Interactive access: multi-factor via Clerk; session management. Programmatic access (public API): bearer API keys, without Clerk or multi-factor, stored as SHA-256 hashes with the plaintext never retained; each key is bound to a single organization, carries separate read and write scopes, supports an optional expiry, and is revocable with effect from the next request |
| Security logging | Production logs capture denied authorization attempts and selected security-relevant events. Coverage does not extend to a customer-facing history of every action. |
| Pattern-based PII redaction | For new transcripts processed through the current application repository write paths, supported phone, email, payment-card and valid IBAN patterns are redacted on write and independently before LLM processing; a contextual masking is applied in addition to these patterns to the stored transcript sent again to the LLM sub-processor — to build a deal state, to answer an AI assistant question or to check a draft criterion wording — and to transcript and chat segments sent to compute search vectors, the text a Controller's user types into search and an AI assistant question matched against the search index: it masks a passport or other identity-document number, a tax number, a date of birth, bank account details and an address that follows a label such as “address” and a colon, a dash or “is” (a labelled address also where it continues after a line break within one turn), and an address written as a street word from a fixed list (“street”, “avenue”, “calle”, “avenida” and the like, in English, Spanish, Ukrainian and Russian), the street’s name, a comma and a house number (such as “calle Mayor, 7” or “avenida Diagonal, 12”); an address or a date of birth without such a label or form is not detected (for example “my address 123 Broadway”, with no colon, dash or “is” after the label, or “Hauptstraße 12”, with no comma); for the stored transcript sent again, a labelled value split between two turns is masked too, and a transcript in which such a value cannot be masked is not sent; search vectors computed before release 26.50.0 were computed after the pattern-based redaction only; for the deal state, the pattern-based redaction and the same contextual masking are applied to the analysis summary, the previous state and the model's answer; the questions a Controller's user asks the AI assistant receive the pattern-based redaction and the same contextual masking before they are stored, and only the stored text reaches the LLM, and the generated answer is masked the same way before it is stored; in addition, each question before it is stored, the transcripts and analyses an assistant answer is generated from with the earlier questions and answers of the conversation, and the generated answer itself, are checked by the LLM sub-processor in separate requests for Personal Data the pattern-based redaction missed (addresses, dates of birth, identity and tax numbers, card and account details, phone numbers and emails, also where split between turns or spelled out in words), and what it finds is masked before the question is stored, before the answer is generated and before it is stored; a question whose check fails is not accepted, a record or earlier questions and answers whose check fails are left out, and when the check of the answer or of every record fails, no answer is given. The control does not detect all Personal Data, does not redact original audio and was not applied retroactively: earlier transcript, preview, metadata and word-timestamp rows were not rewritten and remain subject to the Customer's configured retention period. |
| Rate limiting | Sliding-window rate limiting, bucketed per authenticated user or API key. On the public API an additional limiter buckets by source IP address and runs before the key is authenticated |
| Webhook security | Webhook authentication: per-installation secret, HMAC signature or sender IP allowlist, depending on the provider |
- Data Subject Rights. Processor provides a workflow to record, review and track access, erasure, rectification and objection requests. Approved requests are currently fulfilled through the applicable product and support procedures; the DSR workflow does not yet execute them automatically.
- DPIA Assistance. Processor assists Controller in conducting Data Protection Impact Assessments.
- Deletion. At Controller's choice, Processor shall delete or return all Personal Data after end of Services.
- Audit Rights. Processor allows Controller audits and inspections to demonstrate compliance.
- Infringing Instructions. Processor shall immediately inform Controller if, in its opinion, an instruction from Controller infringes Data Protection Laws (Art. 28(3), final paragraph).
5. Obligations of the Controller
- Obtaining consent from call participants where required by applicable law
- Informing Data Subjects about recording, transcription, and AI analysis
- Maintaining records of consent where applicable
- Responding to Data Subject rights requests through the tracked workflow and the applicable product or support procedure
- Configuring the appropriate consent mode and jurisdiction settings before processing
- Ensuring a valid lawful basis for all data submitted to the Service and, where transcripts may contain special categories of personal data (Art. 9 GDPR), establishing an applicable Art. 9 condition
6. Sub-Processors
Controller provides general authorization for the sub-processors listed in the Sub-Processor List. Processor shall notify Controller at least 30 days before adding a new sub-processor. Controller may object on reasonable data protection grounds. If no resolution is reached, Controller may terminate the affected Services without penalty.
7. International Transfers
Processor is established in Spain (EU). For transfers to the US:
| Mechanism | When Applied |
|---|---|
| EU-US Data Privacy Framework (DPF) | US sub-processor is DPF-certified |
| Standard Contractual Clauses (2021) | US sub-processor without DPF |
| Adequacy Decision | Sub-processor in country with EU adequacy decision |
For UK transfers, the International Data Transfer Addendum to the EU SCCs (issued by the ICO) is incorporated.
Exception — Telegram alert delivery. None of the mechanisms above is available for one optional destination. Telegram Messenger Inc., the counterparty for Telegram's Bot API, is incorporated in the British Virgin Islands; it publishes no data processing agreement for that API and offers bot operators no Standard Contractual Clauses, and neither the British Virgin Islands nor the UAE holds a European Commission adequacy decision. Processor therefore makes no representation that a transfer safeguard is in place for this destination. The channel is disabled by default and transmits nothing until Controller or one of Controller's users explicitly connects a chat; only business alerts are transmitted, and the fields are set out in the Sub-Processor List. Controller decides whether to enable it. All other Services and delivery channels are unaffected.
8. Data Breach Notification
Processor notifies Controller without undue delay and in any event within 48 hours of becoming aware of a Personal Data breach. The notification includes: nature of breach, categories and number of records affected, likely consequences, and measures taken.
9. Data Retention and Deletion
| Data Type | Retention |
|---|---|
| Audio recordings | 365 days |
| Transcripts | 365 days |
| AI analyses | Follows transcript |
| Voice metrics | Follows transcript |
| Search vectors | Follow the record: rebuilt from the current transcript when the record is transcribed again (if that build fails, the earlier vectors stay until the next index rebuild) or the index is rebuilt, and removed from the index after the record is deleted — by a user, at the end of the retention period or together with its project — normally within a few hours; while a large volume of deletions is being processed, removal can take several days |
| Metadata | Follows transcript |
| Customer questions and decline reasons | Follow the record: deleted together with it, including at the end of the retention period |
| Conversation themes | The set of themes and the record of its changes: duration of Controller’s account (a theme is archived, not deleted; cascade-deleted when the account is deleted). The themes and the suggested theme of a conversation: follow the record, deleted together with it, including at the end of the retention period. The list of phrases of a starter set: until the set is approved or a new one is requested, and at most 30 days; the proposed themes stay until a new starter set replaces them. A weekly suggestion: at most 30 days after the last week in which it reached the threshold; a rejected one 90 days after the decision, without its phrases, which are erased when it is accepted or rejected; an example conversation is removed when that conversation is deleted. The phrases of a starter set or a suggestion are not linked to a record, so a phrase from a conversation deleted in the meantime remains until these periods end |
| Criteria answers | Follow the record: deleted together with it, including at the end of the retention period |
| Criteria (definitions) | Duration of Controller's account (cascade-deleted when the account is deleted); an archived criterion is kept so that past answers stay readable |
| Coaching feedback | Annotations and reviews: follow the record, deleted together with it, including at the end of the retention period. Notes: follow the retention period — deleted by a routine scheduled cleanup once older than Controller’s retention period, counted from when the note was written; deleted earlier when Controller’s account is deleted, when their author’s user account is deleted or, for a note about an employee with a platform account, when that account is deleted; deleting the call a note names removes only the link. Coaching sessions: follow the retention period — deleted by a routine scheduled cleanup once both their creation and their scheduled time are older than Controller’s retention period, so a session scheduled for a future date is never deleted before it takes place (a session can be scheduled at most 365 days ahead); deleted earlier when Controller’s account is deleted, when the user account of the person who scheduled them is deleted or, for a session with an employee with a platform account, when that account is deleted; deleting the call or the alert a session names removes only the link |
| Assistant conversations | Controller's retention period; also deleted when the record asked about, or any record the answer was built from, is deleted — usually together with that record, otherwise by a routine scheduled cleanup |
| CRM deal-stage data | Follows transcript; the phase determined by the AI model for a stage (with the hash of the stage’s name) is kept until it is replaced or Controller’s account is deleted (the row of a stage removed from the CRM stays until then); the stage records of a Uspacy deal are deleted 90 days after the deal was last seen on that stage, and when Controller’s account is deleted |
| Deal states (with the CRM deal snapshot) | Controller's retention period, counted from the earliest conversation whose facts a state may carry: a state is deleted once that conversation falls outside the period, when its project is deleted, and on termination. Not deleted when a record it was built from is deleted: its text, quotes and links to that record are kept as the Controller's users saw them until the state itself reaches the end of the retention period |
| Deal-state comments and to-dos written to the Controller's CRM | Under Controller's management in its CRM once written; deleting a record or a deal state in Aelo does not remove them |
| Emails already sent (reports, alerts, notifications, digests) | Outside Processor's control once delivered: copies stay in the recipients' mailboxes, and Resend keeps delivery logs under its own retention (pending confirmation); deleting a record does not recall an email sent about it |
| CRM deal-outcome data | Duration of Controller's account (cascade-deleted when the account is deleted). The decline-reason code is kept only while the deal stays lost and the conversation it was taken from exists: it is removed when that conversation is deleted, including at the end of the retention period, and when Processor sees the deal reopened or won; it is replaced or removed when a newer conversation of the same loss cycle ends in a decline or when that conversation's analysis no longer confirms the reason. While stage-based evaluation is switched off, Processor does not re-check the deal, and the code stays until that conversation is deleted |
| CRM org-structure data | Duration of Controller's account (cascade-deleted when the account is deleted); the tree is replaced in full on each import, and an individual not-yet-matched employee's record is deleted sooner, as soon as an Aelo account for them is linked |
| User accounts | Duration of service |
| Action log (security-relevant actions of the Controller’s users: access to records and analyses, refused access, role changes, settings changes, authorised re-analyses), and the separate access log (changes to roles, membership, project assignments, the organisation structure, analytics access grants and account merges, with the identifiers of the acting and the affected user and the values before and after — for a change from the CRM, also the CRM user identifiers linked to users and the CRM identifiers of employees with no Aelo account yet — without names, email addresses or conversation content) | 365 days from the action, and no longer than the user’s account or the Controller’s account exists; entries older than 365 days are deleted daily. When a user’s account is deleted, the entries of the Controller’s organisation are kept for the Controller, until the end of those 365 days, with the user’s identifier, the action and its time, and the IP address and browser user agent are erased; entries not linked to an organisation are deleted together with the user’s account. Access log entries of the Controller are likewise kept after a user’s account is deleted, until the end of those 365 days. On termination the Controller’s entries are deleted with its other data |
| In-app notifications | Deleted 180 days after the notification was created, together with every user’s status of it (whether it was delivered, read or archived); deleted earlier together with the record it concerns (including at the end of the retention period) and when Controller’s account is deleted |
| Notification delivery log (channel, sent or failed, time, error message, recipient user identifier, technical delivery details such as the provider message identifier) | Deleted 30 days after the delivery attempt, whatever retention period Controller sets |
| Ingest activity log (routing decision for each incoming webhook: routed, rejected, skipped or error, with its reason, the matching routing rule, the external identifier of the call or chat and technical details of the decision) | Deleted 90 days after the entry was written, whatever retention period Controller sets; the entry of a call or chat still waiting to be processed is kept until its receipt is processed |
| Ingest receipts (record of each incoming call or chat taken in for processing, used to recognise a repeated delivery: external identifier, vendor, installation, content hash, identifier of the record created, version timestamp, idempotency key; for a call whose recording and details cannot be fetched again from its source by the call’s identifier, such as one from Uspacy, also, while the call is waiting to be processed, the details that came with the call: the address of the recording, the employee’s identifier in the source, the other participant’s phone number, the call’s direction, duration and subject, its link to a CRM deal, lead, contact or company, and whether it was missed or ended without a recording) | A processed receipt is deleted 30 days after it was processed; a receipt still waiting to be processed is kept until it is, and the call details are removed from it when it is processed. Afterwards a repeated delivery is recognised by the record itself, while it exists |
| Application diagnostic data (Sentry) | Governed by the sub-processor's own plan-based retention settings, not Processor's uniform 365-day policy; the exact period is pending confirmation — see Sub-Processor List |
Controller configures retention. The platform default is 365 days on every plan, and the Controller may set its own retention period for its organization within a range of 1 to 365 days (Settings → Organization; organization owner only). Unlimited retention is not available to Controllers — data is always deleted after the configured period. A Controller that sets no value follows the platform default. Processor automatically deletes data exceeding the effective period.
Lowering the period deletes existing data that falls outside the new window; before the change is applied, the Processor shows the number of records, coaching notes, coaching sessions, AI assistant questions and deal states it would delete, and the deletion is permanent. Upon termination, Controller has 30 days to export, then all data is deleted within 30 days. The application diagnostic data row above is one exception — retention there is set by the sub-processor's own plan; the action log row is another — its entries, and those of the access log, are kept for 365 days from the action whatever period the Controller sets; the notification delivery log, the ingest activity log and the ingest receipts rows are others — they follow the fixed periods stated in their rows.
10. Term and Termination
This DPA remains in effect for the duration of the Services agreement and automatically terminates upon termination of Services, subject to post-termination data handling.
11. Liability
Each party's liability under this DPA is subject to the limitations in the Services agreement. Nothing limits liability for breaches of Data Protection Laws to the extent such limitation is not permitted by applicable law.
Document ID: DPA-Aelo-2026-001 · Version: 1.64