E

Ethyca

Without its models, Ethyca is still a data governance platform. Fides is a taxonomy and policy language, Janus manages consent and preferences, and Lethe fulfills data rights requests and deidentifies data. The data inventory can be filled from existing data security tools and catalogs. None of that needs a model. The models power two core features. Helios classifies discovered data with a large language model applied to database metadata.

Astralis runs ongoing agentic risk assessment and enforces purpose based access at runtime, including rewriting an agent's query when its inferred purpose is not allowed. Ethyca positions itself as a governance harness for AI, not as a model product.

Ethyca publishes measured accuracy for its classifier, with the test set and method. An engineering post dated 23 December 2025, by a named author, reports precision, recall and F1 for Helios. Accuracy rose from about 50% at baseline to over 80% on an adversarial benchmark suite and over 95% on benchmarks drawn from real systems. Validation on customer provided datasets is reported at over 90%. Positive AI labels were wrong 5% of the time, against 14% for human labels.

The test set is described: about 2,000 tagging tasks across 43 data categories, benchmark examples for 46 categories, and fully synthetic schemas built without customer data. The labeling method is described too, including a disagreement review that grew the labeled set from 449 to 825 fields. Three failure modes are named and fixed: laziness, shallowness and credulity. Ethyca also shows why naive accuracy misleads: a classifier that does nothing scores 98.6% on a typical enterprise data mix.

The product cites no legal authority, so checks against primary law and citators do not apply. Two limits: no accuracy figure appears on a product page, only in the engineering post, and the category breakdown is shown in screenshots, not text.

What the system does alone is set out, and customers get review surfaces, but not the thresholds. Astralis checks every agent call through an MCP gateway before it runs, rewrites a query whose inferred purpose is not allowed, rechecks the risk register nightly, and drafts assessments and mitigations for approval. Customers set the limits through a purposes taxonomy, policy definitions and the Fides ontology, which they version themselves.

Review happens in the risk register, the assessment report and an audit record of how and why each access decision was made. What is not published: when the gateway blocks rather than rewrites, what happens when an enforcement decision is wrong, and whether a person approves a mitigation before it applies. The classifier post's discussion of human review concerns Ethyca's own evaluation work, not customer oversight of the product.

Customers are named and figures are given, but the two are never tied together. Sixteen customers are named, each with a profile page, including The New York Times, Ramp, SurveyMonkey, WeTransfer, Vercel, Lush, Zola, Casper and American City Business Journals. Platform figures are specific: 744m preferences processed a year, more than 4m access requests, 200+ global brands, $74m+ saved by automation, subject requests cut from 15 hours to 17.2 seconds, 1 PB governed daily and 150,000 policy decisions a second.

None is tied to a named customer, and no method is given. The New York Times profile page was opened. It shows the paper's own business figures (11.7m subscribers, $2.59bn revenue, 139 Pulitzer Prizes) and a chief executive quote about subscription strategy, not about Ethyca. The other fifteen customer pages were not opened.

Confidentiality commitments sit in a contract a buyer can read before signing, but privilege goes unmentioned. MSA section 4.1 licenses Customer Data only to provide the Service and meet Ethyca's obligations. A separate narrow license over Audit Data, three listed fields, is for legal compliance only. Section 4.2 limits use of Customer Data to operational needs such as audit trails and system monitoring, or where law requires.

Section 9 sets mutual confidentiality, with need to know limits and written duties on staff. DPA section 3(f) commits to deletion or return within 90 days of termination. The CCPA schedule bars retention, use or disclosure beyond the business purpose, and bars sale. The classifier reads metadata only and, in Ethyca's words, needs no access to sensitive data. Two gaps: nothing addresses privilege or work product, and no model provider's retention terms are stated, since the inference platform is not named. MSA last modified 10 February 2023, DPA dated 17 October 2024.

Legal advice is plainly excluded in the agreement that governs the product. MSA section 10.4, headed No Legal or Regulatory Advice, says Ethyca does not provide legal, accounting, tax or regulatory services. It does not advise on complying with or interpreting any privacy or security law. Any recommendation concerns how the Service works, is given from a business view, and should not be relied on as legal advice. The customer alone is responsible for identifying and meeting its own obligations.

Section 10.3 adds that Ethyca does not warrant that using the Service will result in legal compliance. Use is limited to Authorized Users under an Order Form for internal business purposes. What is missing: nothing on a supervising lawyer's competence or supervision duties, and no jurisdiction limit on the guidance the product gives. There is no public advice surface, so consumer facing limits do not apply.

Testing of Ethyca's own classifier is real and published, but no owner is named and no findings on uneven output across people are disclosed. Its published evaluation framework defines and defends an accuracy metric and uses an adversarial benchmark suite. It reports metrics by category in a purpose built workbench, reports findings on model size, and explains the precision and recall tradeoffs between more conservative and more creative models.

What is missing is governance structure. No accountable owner is named. No AI policy, model card, governance committee or review board is published. No finding is disclosed on uneven output across populations or data subject groups, as opposed to data categories. Ethyca sells AI governance and publishes solid testing, but no governance structure of its own.

Ethyca publishes contract terms on all five areas: retention, deletion, access control, subprocessors and incidents. Processing is limited to the term of the Agreement. Controls in the product let the customer retrieve, correct, delete or restrict Personal Data, and the Audit Data license that survives covers three listed fields only. DPA section 3(f) commits to deleting or returning all Personal Data within 90 days of termination.

Backups are carved out but must be isolated, kept from further processing and deleted under stated practices. Annex 2 sets out access control item by item. It covers authentication, permissions checked against each data set's attributes, API access by key or OAuth, VPC and firewall segmentation, and intrusion detection. It also covers static code analysis, annual third party penetration testing, a responsible disclosure program, least privilege staff access reviewed quarterly, and background checks.

Annex 1.H names subprocessors with purpose and location: Amazon Web Services for cloud infrastructure in Virginia and Twilio SendGrid for transactional email. Customers are notified of changes and have a 30 day objection right for European and Colorado data. Incidents are notified without undue delay, with information and help for the customer's own notice duties. The agreements are dated 10 February 2023 and 17 October 2024.

The liability position is specific, but nothing covers wrong output. MSA section 8.1 defends and indemnifies against third party claims that the Service infringes a US patent, trademark, copyright or trade secret. It lists six excluded activities and says this is the sole and exclusive remedy. Section 11 caps each party's liability at 150% of fees paid in the twelve months before the act or omission, and at 300% of those fees for the section 8.1 indemnity.

Gross negligence, willful misconduct, payment obligations and the customer's own indemnity are carved out. Section 10.2 warrants that the Service conforms in material respects to the Documentation, with notice, correction and a right to terminate if it cannot be fixed. A buyer can actually use that warranty. What is missing: no indemnity covers wrong output, and section 10.3 disclaims all warranty on the Service and its reports or outputs. It also expressly disclaims that use will result in compliance with law. No insurance is named.

Integrations with the systems Ethyca's work runs on are documented, but not how deep they go. Its pages name Snowflake and BigQuery, AWS, identity providers, website cookies, tags and SDKs, third party apps, and data intake by API, SFTP, manual and web upload. The MSA prices Connector Integrations as a separate line, which points to a connector catalog. Much of this can be checked: the Fides core and its connectors are public on GitHub, the API docs and release changelog need no login, and Astralis offers an MCP gateway that checks agent calls against policy.

The integrations catalog and documentation site were not opened, so what each connector moves, and in which direction, was not established beyond the home page and the agreement. A product bought by privacy counsel connects to warehouses and identity systems, not document or practice management systems, and no such connection was found.

Delivery is stated, with some residency detail. There are two paths: the open source Fides core, which a customer can run on its own infrastructure, and Hosted Fides, the subscription service. For the hosted service, DPA Annex 2 says Customer Data sits in shared, multitenant storage reachable only through application interfaces and APIs, with no direct customer access to the infrastructure. The subprocessor annex places cloud hosting with Amazon Web Services in Virginia, United States.

DPA section 6 says Personal Data goes to Ethyca, Inc. in the United States and to other places where its affiliates and subprocessors operate, under the EU Standard Contractual Clauses. Not published: a list of regions, any choice of region, what differs between the four tiers, or where processing happens as opposed to storage.

Ethyca shows no certification of its own, though its published security assessment is real. No certification badge appears on any page read. The only standards named are SOC 2 Type II and ISO 27001, and the DPA assigns both to Ethyca's data center partners, not to Ethyca. Those cover the hosting provider, not this product. Ethyca's own assurance is assessment rather than certification: annual penetration testing by recognized third parties, with a summary report shared in confidence on request, plus a responsible disclosure program and static code analysis, all in DPA Annex 2.

A trust center at trust.ethyca.com is open without a sales call and says audit reports and security policies are available there. Its content is hosted by Vanta and could not be read, and a search did not recover it. If it names a certification held by Ethyca itself, the finding here would change. The DPA has a drafting slip: its Virginia and Colorado section says ISO 29001 where the European section says ISO 27001.

The models are named in an engineering post rather than on a product page. The classifier was built and benchmarked against models including Qwen's QwQ-32B and DeepSeek-V3. The post finds that accuracy stops improving above about 32B parameters and that models below 18B make errors that cannot be corrected. The design is described in detail: one request per field so fields are not skipped, written reasoning about tagging, prompt caching, prompt optimization rather than fine tuning, and metadata only inputs.

What is missing: no inference provider is named, the post says only cloud inference platforms, and nothing commits to telling customers when the model supply chain changes. Ethyca's contract subprocessor list names only AWS and Twilio SendGrid, so no model or inference provider is on the list customers are notified about. Post dated 23 December 2025.

A pricing structure is set out, with no prices. The site has no pricing page and no figure, and every call to action is a demo request, a free trial request or a contact form. The structure is in the agreement. MSA section 6.1 names four tiers: Fides True, Fides Team, Fides Plus and Fides Enterprise. Connector Integrations are charged separately, and tier changes or added connectors are charged pro rata for the time used.

Section 6.2 sets monthly invoicing in advance on 30 day terms, and fees are not refundable except on Ethyca's breach. Section 6.3 allows price increases only at renewal, with at least 60 days' notice, and only when applied to all customers. The open source core is free to self host. No rate, price band or implementation figure was found.

Named customers show who Ethyca serves, but not where its coverage stops. Its customers page labels customers by industry: publishing (The New York Times, Axios, American City Business Journals), internet software (WeTransfer, SurveyMonkey, Vercel, Podium), fintech (Ramp), retail and consumer brands (Lush, Casper, Away, Parachute, Zola), and marketplaces and services (JustPark, Slice, Snackpass). It targets enterprises, with a separate enterprise page, and says it works as a technical partner rather than selling an off the shelf SaaS tool.

Its workflows show who buys, though the site has no page by role or buyer segment. Not stated: any jurisdiction limit, which regimes are unsupported, or who the product is not for. Practice area and firm segment coverage do not apply to a data infrastructure product bought by in house privacy and legal teams.