P

Prudencia.ai

Prudencia.ai sells access to a legal copilot, and nothing in the offer works without it. Every plan is priced on the copilot. The free plan is counted in interactions, and the Pro plan lists unlimited interaction with the copilot among its features. The how it works page describes four steps on every query. The system identifies the subject and territorial scope, retrieves passages ranked by meaning rather than matching words, checks its citations, and writes an answer linked to its sources.

Drafting, contract risk review and document analysis all run through the same assistant. The legal material underneath is public, from the Official State Gazette and CENDOJ, and no page offers it as a searchable database apart from the copilot.

The ranking page reports the Legal Reliability Index of observatorio.legal for a campaign run from 7 April to 7 August 2026. Fifteen assistants answered 90 items, each run three times, and two blind graders checked 2,774 citations one by one against CENDOJ and the official gazettes, classing each as verified, misattributed, hallucinated or unverifiable. Prudencia.ai scored 90.6 of 100, with an interval of 88.6 to 92.6, a hallucination rate of 2.6 percent and 91.2 percent traceability.

The page says the vendor took no part in designing, running or scoring the test, and does not say who runs the observatory. According to the FAQ, every cited judgment is checked against the collection before it is shown. An unverifiable one triggers a rewritten answer and is dropped if cited again, so a real ruling not yet indexed may be left out. Prudencia.ai's own misattribution rate is not shown.

Section 4 of the Terms says all output must be reviewed, validated and decided by a qualified human professional, and places professional responsibility on the lawyer using it. Section 5 forbids using the product to assist judges, magistrates or arbitrators in binding decisions, to assess a person's credibility, or to predict court outcomes for decision making. The responsible use policy requires the user to check every answer before it goes into a filing, contract, opinion or client message, and drafts arrive for the lawyer to review and sign.

The automatic steps are described: citation checks, rewrites and removal of unverifiable judgments. What happens after a wrong answer reaches a user is not described.

Bufete Cortés Martín Almendro says work time fell by about 25 percent from the first phases of use, and reports no errors in citing case law. Diana Díaz of Diana Díaz Asociados in Palma filed an appeal in two days and says she cannot give a figure, in a case study of 21 April 2026. Guillermo Ramón Bernabeu, managing partner of a firm in Barcelona and Tarragona, reviewed a four hour video in about thirty minutes, by his account of 4 June 2026.

Further studies cover two more lawyers and firms. The tool pages claim more than 25,000 lawyers in Spain, and the August 2026 release says 288 firms use Deep Research. None of the figures comes with a period, baseline or method, and the testimonials on the tool and practice pages carry initials only.

The data processing agreement bars training, binds staff to confidentiality and holds subprocessors to equivalent duties, and section 13 of the Terms adds a mutual confidentiality duty that survives the contract. The compliance page says the platform was designed around the duty of professional secrecy, allows no human access to content and shares nothing with third parties. The advisory firms page says the secrecy duty is covered. None of that secrecy language is in the Terms, and nothing addresses privilege or what OpenAI and Anthropic may retain.

Section 1 of the Terms limits the service to legal professionals and businesses and says it is not directed at consumers. Section 3 calls output auxiliary, indicative and nonbinding, and the legal notice says outputs are not binding legal advice. Responsibility for the work stays with the reviewing lawyer, and the product is built on Spanish law. The assistant page also offers a round the clock triage chatbot for a firm's clients with no disclosure described, and a July 2026 post addressed to citizens carries the free sign up button twice.

The responsible use policy, updated July 2026, lists principles of transparency, human oversight, quality and responsible development. It commits to telling users they are dealing with AI, and says new features are evaluated periodically to make answers more useful, precise and safe. The compliance page classifies the product as a limited risk professional AI system under the EU AI Act, the Terms list prohibited uses drawn from the Act, and a data protection officer is reachable at dpo@prudencia.ai.

The outside benchmark measures accuracy but is not a governance process of the vendor's own. No one is named as accountable for model behavior, and no release testing or error review is described. Nothing reports whether answers are as reliable in labor or administrative law as in civil law, or across Spain's regional legal systems.

The data processing agreement governs what happens to an account after the contract, with total deletion available on demand. Section 5 requires breach notice within 48 hours, section 10 allows one client audit a year on two months' notice, and Annex II names six subprocessors. The privacy policy of 16 January 2026 lists least privilege access, activity logs and breach procedures, and keeps billing data for six years.

The agreement and the privacy policy describe the period after the contract differently. Activity logs are kept under applicable law, with no period stated.

Section 14 of the Terms of February 2026 says Prudencia.ai does not guarantee absolute accuracy, fitness for a particular case or the absence of errors. It makes human validation mandatory, provides the service as is and as available, and calls AI probabilistic rather than deterministic. Section 15 caps total liability in every case at the amount the client paid in the 12 months before the claim. It excludes the client's professional decisions, indirect damage and lost profits, prohibited use and outages attributable to third parties.

No indemnity, insurance or remedy for a defective answer is published, the service level in section 16 covers only the vendor's own availability, and no exception to the cap for intent or gross negligence is stated. A fabricated citation in a filed brief stays with the lawyer.

The assistant page lists a browser and email extension, a connection with LexNET and CENDOJ, and synchronization with a firm's software, without saying what is exchanged, in which direction or how it is set up. LexNET is the courts' electronic filing system. The Enterprise plan includes integration with the client's systems through an API, and no developer documentation is published. Elsewhere the product presents itself as standing apart: the advisory firm and civil law pages call it a complement to practice management software rather than a replacement.

The administrative law page says it does not replace case management or electronic file applications. Documents arrive as uploads of PDF, Word, scanned images and audio, and the subprocessor list names Google for sign in. No Word add in, document management connector or firm single sign on is described.

Annex II to the Terms places hosting, infrastructure and data processing with Amazon Web Services EMEA in Spain. It lists OpenAI Ireland for AI inference with residency in the European Economic Area, and Anthropic, PBC for AI inference outside it under standard contractual clauses. Section 9 of the data processing agreement commits to valid transfer mechanisms. The marketing pages say otherwise. The drafting page says client data never leaves the European Union, the public sector page shows zero administrative data processed outside it, and the how it works page promises no unnecessary international transfers.

The service is a single cloud application. Not stated are which features route to Anthropic, whether a customer can keep processing inside the European Economic Area, and any private or regional option.

No attestation issued to Prudencia.ai is published. No ISO 27001, SOC 2 or Esquema Nacional de Seguridad certificate issued to the company appears on the compliance page, the Terms or the responsible use policy, and there is no trust center or route to request an audit report. The public sector page says the solution is aligned with the GDPR, Spain's data protection law and the Esquema Nacional de Seguridad requirements for unclassified processing.

It shows a label reading 100 percent EU infrastructure, ENS and GDPR, which ties the standard to the infrastructure rather than the company, and names no certificate, certifying body or scope. The compliance page and the pricing FAQ describe 256 bit AES encryption at rest and TLS 1.2 or higher in transit, and the privacy policy lists security audits without saying who performs them. The data processing agreement lets the vendor offer equivalent reports or certifications in place of an audit, and none is named.

The model providers are named in the agreement and the models are not. Annex II names OpenAI Ireland Ltd and Anthropic, PBC for AI inference, beside Amazon Web Services for hosting. Clause 8.2 of the data processing agreement gives a general authorization to add or replace subprocessors, provided the client is informed of changes in a reasonable way, with no lead time or objection right stated. The product pages name neither provider: the privacy policy refers only to language model providers, and a comparison post of 18 September 2026 describes the vendor's own legal retrieval architecture.

Undisclosed are the model family or version, which provider serves research, drafting or transcription, and whether the two run side by side or one backs up the other. Nor is any notice promised when a model changes within a provider.

Self serve prices are published with unit and term, and Enterprise is quoted on request. The pricing page sets out a free plan, Pro for one user and Teams for two to five users, each with monthly and annual prices that exclude VAT and carry no minimum term. Enterprise adds unlimited users, API access, activity reports and branding at a price on request, and nothing says what its custom deployment adds. The Terms make fees payable in advance and nonrefundable once a period starts, and let prices change at renewal after notice.

Prudencia.ai's FAQ names solo lawyers, firms with teams, advisory firms and company legal departments. Separate pages address advisory firms of one to 50 professionals with 30 to 500 recurring clients, and public bodies such as municipal secretariats and regional legal services. Pages for civil, labor, commercial and administrative law list the matters covered, from successions and leases to dismissals and public procurement.

Each also lists what not to expect, such as foreign law beyond the EU succession regulation, a tax only tool or a practice management system. The limits of the source material are stated: not all Spanish case law is published, and bulk access to some collections depends on conditions set by the judiciary's governing council. The August 2026 release counts more than 76,700 judgments. Answers run in five languages, among them Spanish, Catalan and Basque.