P

Privado AI

The model drives a core capability, on a product the vendor also sells without it. Wren is a named AI agent that runs the privacy assessment cycle: watching internal tools for activity to review, checking it against the customer's playbook, researching the question, reading documents, contracts, interview transcripts, code and connected apps, generating evidence, filling in the assessment and routing it to approvers.

That is model work at the center of the assessment product. But the pricing page confirms the platform includes a base assessment module that can be bought without Wren. The web auditor, app auditor and code scanner are separately priced products that work by live scanning, static analysis and simulated consent journeys, not models. A large, separately sold part of what Privado offers does not depend on a model.

Grounding is real and the retrieval path is described, but nothing is measured. The vendor names the sources Wren reads to fill in an assessment: imported documents such as product requirement documents, technical specifications, support documents, interview transcripts and notes; contracts, from which data processing terms are extracted; source code through a source control connection; connected apps through API connections; and internet research, with a cookie agent that checks cookie categories by searching in real time.

The output includes evidence excerpts traced to their source, and gaps Wren could not fill are flagged rather than filled. There is no accuracy rate for completed assessments, no evaluation, test set or error analysis, and nothing on how often an evidence excerpt actually supports the conclusion drawn from it. The outputs are assessments and data maps, not legal assertions citing authority.

A real division of labor is published, with review points, but the boundary is not settled. Wren triages incoming activity and starts the right assessment or record of processing based on risk and on the playbook and privacy policies the customer's team defines, so the constraint is the buyer's own document. Assessments go to approvers, gaps Wren cannot fill are passed to a person in chat, and risks it finds are tracked alongside those raised by privacy staff.

That is a workflow with a person at named points. What is missing is where the boundary sits. The vendor says Wren eliminates manual threshold assessments and immediately gives privacy guidance for low risk activity, so it does answer some questions without review. Nothing defines low risk, what Wren may finalize unattended, or what happens when an assessment it filled in turns out wrong after a decision was made on it.

Named customers and named people in named roles, but no figures. Three deployments are published with attributed quotes and linked case studies. At HP, a Privacy Innovation and Assurance Leader describes seeing personal data across the technology stack and automating manual controls such as privacy reviews. At HERE Technologies, a Director and Head of Product Trust says compliance reports keep pace with data flows as the product changes.

At Headspace Health, a chief information security officer describes building privacy into software development to prevent accidental sharing or tracking. Those are real statements from identifiable people. What is missing is measurement: no figure for review time, assessment volume, issues found or cost, no dates, and no method behind any claim. The customers page with the full case studies was not opened and may say more.

The published commitments are substantial and backed by real enforcement, but separation and model providers are not addressed. The subscription terms make Customer Content, including source code, data elements and the reports generated from them, the customer's Confidential Information. They set aside the agreement's standard five year limit for that material, so the duty lasts until an exception applies. Privado's use is confined to what is needed to provide the services, and all intellectual property in the generated reports belongs to the customer.

Enforcement is real: a penalty of twelve months' fees for any confidentiality breach, which does not release Privado from performance or replace the customer's damages claim, and a liability cap that falls away entirely for confidentiality and privacy breaches. The security page adds role based access control and an internal procedure that blocks employee or administrator access to user data except for limited support needs, with staff under confidentiality agreements.

Two things are missing: nothing addresses separation between customers or between matters within one account, and nothing says what the model providers behind Wren may keep, since none is named.

A disclaimer sits in the terms while the product gives guidance, and nothing addresses where output stops and legal judgment begins. The vendor says Wren eliminates manual threshold assessments and immediately gives privacy guidance for low risk activity. That means a model tells a business stakeholder that an activity does not need a data protection impact assessment. Nothing published says such guidance is not legal advice, that privacy counsel or the data protection officer stays responsible for the decision, or what happens if the call was wrong.

The agreement has related but different provisions. Section 5 records the customer's acknowledgment that it relied on its own skill and judgment to check the software's fit for its intended use, and section 9 disclaims fitness for a particular purpose. Those concern suitability and warranties, not the advice line.

No governance position was found for the vendor's own AI, on a product whose agent drafts data protection impact assessments. Nothing names who is accountable for Wren's behavior, describes what is tested before it ships or changes, reports any test result, says whether output varies across legal regimes, languages or document types, or offers a model card. Two things come close but are not governance. The promise that customer data and code are never used to train models is a data use commitment, and the SOC 2 program with continuous control monitoring is security assurance.

The gap matters because of what the agent does. An assessment Wren fills in becomes the customer's record of its own compliance reasoning, shown to regulators and auditors, and the buyer has no published basis to judge how that reasoning is produced or checked. The product documentation site was not opened and may say more.

Retention, deletion, access control, subprocessors and incident practice are all published in specific terms in the data processing addendum and security page. Annex II of the addendum sets out the measures: background checks on new employees, annual security training, written confidentiality and security terms with every vendor, role based access control, and a password policy with credentials held in AWS Cognito. It also covers change management with review before production, TLS in transit and strong encryption at rest, vulnerability assessment and penetration testing twice a year, and backups every fifteen minutes to an encrypted S3 bucket, restore tested twice a year.

Retention has periods. Personal data is deleted when an administrator deletes the user and within six months if the customer leaves, and clause 12 requires its return or deletion within at least thirty days of the end of the agreement. Clause 11 covers incidents: breach procedures, notice without undue delay unless the breach is unlikely to put rights at risk, and help notifying the authority and data subjects. Annex II lists each subprocessor with purpose, data center locations and products supported.

The liability position is published in full, including an insurance schedule. The indemnity at section 12 goes beyond intellectual property: it covers third party claims from Privado's breach of law, from infringement or misappropriation by the software or the customer's lawful use of it, and from Privado's breach of confidentiality. It has three named exclusions and the standard remedies of modifying, procuring, or terminating with a refund.

Section 13 caps liability at twelve months of subscription fees. Section 13(b) lifts both the cap and the exclusion of indirect damages for privacy and security breaches, confidentiality breaches, indemnities, and death or personal injury, so the indemnity sits outside the cap. Section 11(f) adds a fixed penalty of twelve months' fees for any confidentiality breach, without releasing Privado from performance and without prejudice to damages.

Section 14 requires $1,000,000 each of general liability, technology errors and omissions, cyber and privacy liability, and umbrella cover, with four years of tail on claims made policies. Section 5 warrants material conformance to the documentation and no material loss of functionality. Annexure B commits to 99.9 percent uptime with service credits from 5 to 30 percent, and to support response and resolution goals with a named escalation path.

Integrations into the systems where this work happens are documented for each product, with what moves, in which direction and what the customer must set up. Wren connects to Jira, Confluence, Linear and procurement tools to catch activity needing a privacy review, and pushes development tickets back into Jira and Linear. Data maps are built from source code through a source control connection, from third party apps (customer data platforms, tag managers, CRM, contract management, HR, marketing, procurement and database tools) through API connections, and from imported documents.

The OneTrust integration has a stated direction: the data inventory, assessments and risks in OneTrust are updated automatically from Privado. Setup effort is published per module. The web auditor needs only URLs and scan locations, the app auditor needs the app store URL or the APK and IPA files, and neither needs implementation. Wren takes a few days to connect tools, set up playbooks and import documents, and the full platform usually takes one to three weeks. Documentation for each connector sits on the documentation site, which was not opened.

Regions are stated; the deployment model is not. The data processing addendum's subprocessor annex gives a data center location for each entry. Amazon Web Services, the main cloud host where all Privado applications run and all product data sits, is listed for the United States, the European Economic Area and India, and MongoDB, the main product database, for the same three. So a buyer can tell which countries its data may be in.

Transfers out of the EEA use the 2021 Standard Contractual Clauses, set out in full in Schedule 1, under Irish law with the Irish Data Protection Commission as the authority, and clause 8.8 restricts onward transfers. Nothing covers deployment: no tenancy model, no single tenant or self hosted option, no way to choose or pin a region, and nothing separating where processing happens from where data is stored beyond the transfer mechanism.

The web and app auditors let a customer choose where a scan runs from, but that is where the simulated user appears, not where data is held.

There is a badge and a monitoring tool, with no scope, date or report route. The footer and security page show a SOC 2 mark with the AICPA logo, and the security page says Privado uses Drata to monitor more than one hundred internal security controls continuously, with automated alerts and evidence collection. Nothing lets a buyer check it: no type is stated, so Type I and Type II are not distinguished; no trust services criteria are named; no auditor, examination period or report date appears; and there is no route to a report, questionnaire response or portal, since there is no trust center.

Continuous control monitoring is real practice, but it shows a program, not an attestation. The security page says it was last updated on 17 February 2022.

The vendor sells a named AI agent and publishes a subprocessor list, and no model provider appears on it. Annex II of the data processing addendum lists five subprocessors with purpose, data center locations and the products each supports: Amazon Web Services for infrastructure, MongoDB as the main database, SendGrid for email, Intercom for support and Amplitude for product analytics. None is a model provider, and no model, model family, provider, hosting arrangement or inference location for Wren is named anywhere read.

Wren clearly runs on a model, since the pricing FAQ says customer data and code are never used to train AI or ML models. So either an unlisted third party sees the documents, contracts, transcripts and source code Wren reads, or the models run inside the AWS setup already listed, and nothing published says which. No commitment to notify customers of changes is given. The subprocessor mechanism at clause 10 offers an objection route with no advance notice period, though the incorporated Standard Contractual Clauses require thirty days' notice of changes for EEA transfers.

A rate card is published with figures, units, minimums and what each product includes. The web auditor starts at $600 per website per month billed annually, with a four website minimum and bulk pricing by page available. The app auditor is $800 per app per month billed annually, with iOS and Android counted as two apps. Wren is $4,200 per month billed annually for up to 500 assessments. Only the full platform is priced on request, and the page says what it adds.

Each tier lists its features, and volume discounts are disclosed. Annexure A of the subscription terms defines the units: a developer is a user who has committed to a scanned repository in the last ninety days, a website is each unique URL entered, an app is each platform version, and a scan is one run against one website from one location, so the same site from three locations is three scans. The agreement adds fees in advance, a twelve month initial term, nonrefundable fees and 1.5 percent monthly interest on late payment.

Coverage is described in substance across buyers, legal regimes and systems, and the limits are left open. Buyers are addressed by function and confirmed by customer quotes: privacy teams for assessments and records of processing, engineering for code scanning, and security leaders. The three named customers are a technology manufacturer, a mapping company and a digital health provider. Regulatory coverage is named: CCPA, CIPA and GDPR, more than fifty prebuilt compliance checks, dedicated GDPR and CCPA pages, and support for the GPP, TCF and TAG consent frameworks.

The systems covered are listed: websites, mobile apps, internally built software, third party apps and business processes. No limit is stated. No jurisdiction, sector or company size is ruled out, the compliance checks have no stated boundary, and nothing says what a law firm, rather than an in house privacy team, would do with the platform.