A
AnwaltGPT
AnwaltGPT is an AI platform for German law firms, in house legal departments and notaries, combining research, drafting and document work with case files and client collaboration. Its chat answers questions on German law with references to statutes, case law and literature, and Deep Research works through statutes, then case law, then the open web, linking each finding to its source. Beside it sit a contract generator, a gap analysis that finds missing clauses, a document editor, transcription, a legal watchlist and a workflow builder.
Case intake checks every new inquiry against the firm's files and contacts for conflicts, and a client portal gives an access link for each matter. Word and Outlook add ins are offered, and a desktop agent called Cowork works on a folder on the lawyer's computer and returns edits as tracked changes. Models from providers such as OpenAI and Google run through Microsoft Azure in the European Economic Area, with AWS Bedrock or the vendor's own models as alternatives.
Plans are priced per user per month, from a free trial of 30 requests up to 99 euros. AnwaltGPT is made by Mindverse, the trade name of Relativity GmbH of Berlin.
Capability grades
All 15 axes, graded from public sources on the date shown. Hover a grade to see what the letter means on that axis.
AI Centrality
How much of the product is actually AI. Whether the machine learning is the mechanism the buyer is paying for or a feature layered onto conventional software, and whether the vendor is specific about which is which.
The models are what the buyer pays for, and the price list says so: every plan is metered in AI requests, while uploading documents, creating case files and exporting stay unlimited on every tier. The chat answers legal questions with references and Deep Research runs a search in several stages. The contract generator, gap analysis, document editor, transcription and the Cowork desktop agent all produce their output through a language model.
The security page describes the machinery as leading language models from providers such as OpenAI and Google, plus a vector database for semantic search. Around that sits practice software that would work without the models. It covers case intake with a conflict check the intake page calls deterministic, case files, deadlines, a client portal, and export of master data as CSV or JSON for systems such as RA-MICRO and DATEV.
The home page sells the whole as an AI operating system for law firms, and the free tier is 30 AI requests rather than a period of access. The practice layer is real but secondary: it is not counted against a plan's quota, and every product page leads with the AI functions.
Citation Accuracy and Hallucination Disclosure
Whether the vendor publishes measured accuracy on citations and assertions, grounds output to primary sources, and says plainly what its system does when it does not know. Legal has a documented public record of fabricated citations reaching filed briefs, so an untested claim of accuracy is not evidence.
References are promised on every answer, and what stands behind them is neither identified nor measured. The chat page says each answer refers to the relevant statutes, sections and judgments with direct source references. The Deep Research page says results are linked to their source after a search that runs through statutes and regulations, then current case law, then supplementary web sources. AnwaltGPT's home page says answers based on sources from real legal databases clearly reduce the risk of false statements, yet no database, court collection or publisher is named.
The chat page also says the assistant was extensively trained on all German statutes and case law, which describes model knowledge rather than retrieval. No accuracy figure, test set or error rate is published. The nearest thing is an anonymous solo practitioner on the home page saying the references were right in roughly ninety nine percent of cases. Section 8 of the Terms of November 2025 warns that answers can contain inaccuracies, outdated case law or wrong interpretations and obliges users to check all information. Sources are linked, but nothing says which collection they come from or how often they are right.
Autonomy and Oversight Model
What the system decides on its own, what a lawyer must approve, and whether the vendor documents where the review point sits. A tool that drafts under review and a tool that files without one are different products and different risks.
The vendor commits in several places to the lawyer making the decision, and most outputs arrive as drafts to review. The partner brief and the brochure of September 2026 say the legal decision stays with the firm and that the AI prepares work while the firm checks facts, sources, deadlines and results. Drafts need professional approval before use. Cowork, the desktop agent, works on a folder the lawyer points it at and returns edits as tracked changes that can be checked one by one.
Case intake marks conflicts in red before anyone calls back, logs every check, and takes on a matter only after the firm's express confirmation. The security page lists a mandatory notice before an AI generated contract is exported, a visible AI generated label on every answer, and a button to report an inaccurate answer. Section 8 of the Terms leaves review, assessment and decision on all output with the user. The control structure behind those surfaces is missing.
Nothing describes a point at which a workflow or the desktop agent stops for approval, the notice before export warns rather than blocks, and no route after an error is described.
Operational and Outcome Evidence
Named, dated evidence that the product works in production at real firms or legal departments. Case studies with figures and identified customers count. Unattributed testimonials and launch announcements do not.
Users are quoted at length, and none is tied to a named firm with a dated, measured result. Eight case studies are published, and two carry names. Dirk Semper, a lawyer, says he has found no error in the document analysis and still checks every legal statement himself. Klaus Siebert, a partner at an unnamed travel law firm, describes a relevant structural time saving; the figure he gives, complex contracts cut from eight hours to 45 minutes, refers to tools he used earlier.
The other six are anonymous: a case manager, a solo practitioner, a management consultant, a managing director, a commercial manager and a private individual. The consultant says the AI takes about three quarters of an hour of work off each hour he bills, the commercial manager calls it very well suited to people who are not lawyers, and the private individual used it in a consumer dispute with a gym. No case study carries a date.
The About page says law firms across Germany use the product and that the wider Mindverse platform has more than 500,000 users and over 30 enterprise customers. No law firm customer or AnwaltGPT figure is given.
Privilege and Confidentiality Posture
How client confidences are handled: attorney client privilege and work product treatment, segregation of one client matter from another, whether client data trains any model, and what the vendor commits to in writing rather than in marketing.
The marketing promises confidentiality, and the only agreement published tells users to keep client data out. AnwaltGPT's home page says inputs and documents are not used to train AI models and serve only to process requests, and that each firm receives a data processing agreement under Article 28 of the GDPR. It says the product complies with professional secrecy under section 43a of the Federal Lawyers' Act and section 2 of the lawyers' professional code.
The brochure of September 2026 describes a confidentiality agreement that refers to section 203 of the Criminal Code and section 43e of the Federal Lawyers' Act and binds staff and subcontractors without time limit. Neither agreement is published; the security page offers them by email, sent for signature within 48 hours. Section 8 of the Terms of November 2025 runs the other way. Users are asked not to enter confidential, sensitive or personal data, with client data and trade secrets named, and the provider accepts no responsibility for unauthorized disclosure of what users enter.
Separation between matters, what model providers retain, and privilege itself are not addressed in any published document.
UPL and Professional Responsibility Posture
Whether the vendor is clear that it supplies a tool rather than legal advice, who its audience is, and how it addresses unauthorized practice of law, competence and supervision duties, and jurisdiction limits. ABA Formal Opinion 512 is the reference point. Where the advice line is not the duty a product raises, the axis is read through the nearest professional duty it does raise: judicial conduct rules and the reviewing duty for products sold only to courts, and the duty to bill for time actually spent for products that draft time entries.
The advice line is drawn plainly, and its audience shifts from page to page. Section 8 of the Terms of November 2025 says the software serves only as an information and research tool and provides no legal services under the Legal Services Act. It gives no advice, representation or review of individual cases, and users are asked to consult a licensed lawyer for binding advice, wording addressed to someone who is not one.
The footer on every page says the product is intended exclusively for lawyers and other professional legal service providers and not for consumers. The intake widget is described as taking down details rather than advising, with reference to section 2 of the Legal Services Act. The case studies point elsewhere: a private individual describes using it to win a consumer dispute, and a commercial manager calls it very well suited to people who are not lawyers.
Nothing described checks who signs up for the free plan. On supervision, the Terms oblige users to check all information themselves and the brochure keeps the legal decision with the firm. German law is the stated focus, and no page says the product stops at it.
AI Governance and Bias Disclosure
Published governance over model behavior: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.
An AI Act program is published with a product control behind each item, and nothing is said about testing or uneven output. The security page classifies the product as not a high risk system under Annex III of the EU AI Act and lists seven measures, each tied to an article. Users are told they are dealing with an AI system, every answer carries a visible AI generated label, and every AI interaction is logged and kept for at least 24 months.
A mandatory notice appears before an AI generated contract is exported, users can mark any answer as inaccurate, onboarding includes a consent step for AI processing, and users can delete all their data. Section 8 of the Terms of November 2025 repeats the classification and forbids use as a high risk system or for prohibited practices. The security page also describes an information security management system oriented on ISO 27001, which governs security rather than model behavior.
No accountable owner for model behavior, release testing, evaluation results or finding about uneven output across practice areas or kinds of user is published.
AI Safety and Data Stewardship
Retention, deletion, access control, and what happens to prompts and documents after they are processed. Whether the vendor states its subprocessors and its incident practice, or leaves the buyer to assume.
Retention, deletion, breach notice and part of the supplier chain are published in the security materials; the full subprocessor list is not. The security page says requests are processed per session by default and that every AI interaction is logged with a retention period of at least 24 months. Users can permanently delete all their data, including conversations, documents, case files and research, from their account settings.
The brochure of September 2026 says the data processing agreement provides for deletion or return at the controller's choice after the contract ends, unless a duty to keep the data applies. Personal data breaches are notified within 72 hours so firms can meet their own deadlines under Article 33 of the GDPR, and penetration tests run at least every six months with a written report. Access runs on role based permissions in workspaces, and data is encrypted in transit and at rest.
The security overview of 8 September 2026 names Azure, Bedrock and Hetzner; the full subprocessor list comes with the contract documents, sent on request. The privacy policy covers the website alone and names LinkedIn as its only service provider.
AI Liability and Recourse
What the vendor stands behind contractually when its output is wrong. Indemnities, caps, carve outs, insurance, and whether any of it is published or only reachable through a negotiated agreement.
Liability is handled by a short limitation clause, and both the accuracy of output and the data users enter are disclaimed. Section 8 of the Terms of November 2025 says use is at the user's own risk and that the provider accepts no liability for the accuracy, completeness or currency of the information it provides. Liability for damage from use of the service is limited to intent and gross negligence. Injury to life, body or health and breach of essential contractual duties are excepted, and liability for slight negligence is excluded as far as the law allows.
No monetary cap is stated. The same section says AI answers can contain inaccuracies, outdated case law or wrong interpretations and obliges users to check everything themselves. It disclaims responsibility for unauthorized disclosure, loss or misuse of data users enter. No indemnity, warranty on output, insurance or service level is published in the Terms, on the security page or on the pricing page. The data processing agreement, which may carry terms of its own, is sent only on request. A wrong citation in a filed brief stays with the lawyer who filed it.
Practice Systems Integration Depth
How deeply the product reaches into the systems legal work already lives in: document management such as iManage and NetDocuments, Word and Outlook, contract lifecycle management, matter management, e-billing, and court filing systems.
Two Microsoft add ins are described by function, and practice management systems are reached only by file export. The Word add in is installed from the Microsoft add in store for Word 2019, Word 365 and Word Online. It lets a lawyer mark a passage and have it reworded, extended or checked, and opens the chat, templates and gap analysis in a sidebar. The Outlook add in summarizes email threads and prepares replies, according to the brochure of September 2026.
Cowork, a desktop agent for Windows, macOS and Linux, works on a folder on the lawyer's own machine and returns tracked changes. Case intake exports master data as CSV or JSON for external systems, naming RA-MICRO and DATEV, and single sign on through Azure AD is listed for large firms. Depth is missing. No documentation describes what the add ins send, how they are rolled out across a firm or what an administrator configures.
No connection to a document management system, the lawyers' electronic mailbox or SharePoint is named. The brochure lists none, and the sitemap of more than 580 pages carries no help or documentation section.
Deployment Model and Data Residency
Where the software runs and where the data sits. Multi tenant cloud, single tenant, private deployment, on premises, and whether region of residence is a published option or an enterprise conversation.
Storage, model processing, a local option and an enterprise option are each described, with one discrepancy in how precisely the location is stated. The home, workspaces and data protection pages say processing and storage run on servers in Frankfurt am Main. The security page, the security overview and the data flow diagram of 8 September 2026 say hosting is in the European Economic Area. The selected model runs through Microsoft Azure in the European Economic Area, with AWS Bedrock or the vendor's own models as alternatives.
An optional pseudonymization service runs on Hetzner in Nuremberg before a request reaches the model. Each customer runs as its own instance with separate database segmentation. Cowork keeps documents on the lawyer's computer and sends out only model requests, through the AnwaltGPT gateway on Azure in Germany or the firm's own keys for AWS Bedrock, Azure OpenAI or other providers. Local models are a further choice. Enterprise adds hosting in the customer's own cloud and free choice of models. The region used for Bedrock, the route for Google's models and the country of each Azure call are not named.
Security Certifications and Trust Center
Independent attestation a buyer can pull without a sales call: SOC 2, ISO 27001, penetration test summaries, a trust center with current reports and named scope rather than a badge image.
No independent attestation is claimed for the product, and the security program is described in detail. The security page says the operator, Relativity GmbH, runs an information security management system oriented on ISO 27001, with 16 binding security policies. It has an independent information security officer who is an ISO 27001 lead auditor, and an external data protection officer. Penetration tests run at least every six months and are documented in a report.
Service providers must hold at least one certification under ISO 27001, BSI IT-Grundschutz or an equivalent, and processing runs in certified data centers with perimeter protection, access and fire controls, backup power and redundant connections. Being oriented on a standard is not being certified to it. No certificate, certifying body, scope, date or audit report for AnwaltGPT or Relativity GmbH is published, and none is offered on request.
The certifications of the data centers and model clouds belong to those providers. How security is run is described, and no evidence that anyone outside has checked it is offered.
Model Supply Chain Disclosure
Which models sit underneath, whose they are, where they run, and whether the vendor commits to telling customers when that changes. A legal buyer inherits every dependency it cannot see.
Providers and routes are named, and the models and any notice of change are not. The security page says AnwaltGPT uses leading language models, for example from OpenAI and Google, plus a vector database for semantic search, and that the selected model is provided through Microsoft Azure in the European Economic Area. The data flow diagram of 8 September 2026 lists the routes as Azure, Bedrock or models the vendor runs itself.
The security overview of the same date names Azure, Bedrock and Hetzner as subprocessors, Hetzner for a pseudonymization model hosted in Nuremberg. Cowork can use the AnwaltGPT gateway on Azure in Germany, the firm's own keys for AWS Bedrock, Azure OpenAI and other providers, or local models through Ollama, and Enterprise offers free choice of models. An automatic mode sends simple questions to smaller models, without naming any.
No model name or version appears anywhere, and the route for Google's models is not stated. No commitment to tell customers when a model or provider changes is published. The full subprocessor list sits in an annex to the data processing agreement, sent on request.
Commercial Transparency
Whether a buyer can learn what this costs without entering a sales process: published rates, the unit being charged, what sits behind an enterprise tier, and what implementation adds.
Prices, the unit and the quotas are published for every plan except Enterprise. AnwaltGPT's pricing page lists Standard at 19 euros, Premium at 49 euros and Premium Plus at 99 euros per user per month, each including VAT and cancellable monthly, with 20, 50 and 100 requests a day and monthly caps of 500 and 1,000 on the two upper plans. Yearly billing takes 8 percent off, shown as 209, 549 and 1,089 euros. Extra request packs cost 19 euros for 50, 49 euros for 150 and 149 euros for 500, and do not expire.
The free plan gives 30 requests once, with no card and no automatic conversion, which section 3 of the Terms of November 2025 confirms. The page counts a chat message, a contract review, a generated document or an API call as a request, and leaves uploads, case files and exports unlimited. The brochure and the partner brief of September 2026 price Premium at 49.99 euros, which matches the 549 euro yearly figure, against 49 euros on the pricing page.
Enterprise, with custom hosting, model choice and individual integrations, is priced on request, and no implementation or onboarding fee is stated.
Firm and Practice Coverage
Who the product is actually built for. AmLaw, midlaw, small firm and solo, in house departments, government and courts, and which practice areas are supported rather than merely claimed.
Segments are defined by size and type, and the limits of what the product covers are not stated. Solution pages address solo practices of one to four lawyers, midsize firms of 5 to 50, large firms of more than 50 with an international focus, in house legal departments in companies and groups, and notaries, the last for document preparation, commercial register searches and templates. Practice area pages cover more than fifteen fields, among them employment, family, tenancy, contract, criminal, inheritance, corporate, traffic and data protection law, and city pages address firms in two dozen German cities.
The chat page says the assistant covers all essential areas of German law, and the home page sets AnwaltGPT against general chatbots as specialized in German law. The footer limits the audience to lawyers and professional legal service providers, while the case studies include a management consultant, a commercial manager and a private individual. No surface says what the product does poorly, whether other jurisdictions or EU law are supported for a large firm with an international focus, or which practice areas the sources behind the answers actually reach.
3 public documents
The public pages on file for AnwaltGPT, with the recorded signals each one supports and the date it was last read. Open any of them and check the reading against the record.
-
anwaltgpt.de1 signal
Client Data in Training
Read Oct 2, 2026
-
anwaltgpt.de/produkt/workspaces1 signal
Ethical Walls and Matter Segregation
Read Oct 2, 2026
-
anwaltgpt.de/sicherheit1 signal
Prompt and Output Retention
Read Oct 2, 2026
€19 per user/monthEUR, as published, never converted
- AnwaltGPT costs 19 euros a month per person on the cheapest paid plan.
- Premium costs about 49 euros and Premium Plus 99 euros a month per person.
- Each plan gives a set number of AI questions per day.
- You can try 30 questions free without a card.
- Prices include VAT, and large firms ask for a custom price.
Free: 30 requests once, no card and no automatic conversion. Standard: 19 euros per user per month including VAT, 20 requests a day. Premium: 49 euros per user per month including VAT on the pricing page, 50 requests a day and 500 a month. Premium Plus: 99 euros per user per month including VAT, 100 requests a day and 1,000 a month. Yearly billing takes 8 percent off: 209, 549 and 1,089 euros a year. Request packs: 19 euros for 50, 49 euros for 150 and 149 euros for 500, with no expiry. Enterprise: on request, with custom hosting, free choice of models and individual integrations.
Implementation: Not published. The partner brief of September 2026 describes a pilot limited to two workflows and a small team, with no fee stated.
Confidentiality and data terms: Not applicable in the German market. A data processing agreement under Article 28 GDPR and a confidentiality agreement are sent on request.
Note: Prices include VAT and monthly plans can be canceled at the end of each billing period. A request is a chat message, a contract review, a generated document or an API call; uploads, case files and exports are unlimited. The brochure and the partner brief of September 2026 give Premium as 49.99 euros against 49 euros on the pricing page.
Legal Signals
What each signal meansA signal records what public sources say on the date shown. It is not a grade and it is not a recommendation. Where a signal reads Not addressed, it means the index did not locate the material in public sources on that date, which is a statement about disclosure rather than about the product.
Client Data in Training
Can material a lawyer puts into this product be used to train a model?
A public policy or trust page states no training on customer content, with no matching term located in the published agreement.
The home page and the data protection page say inputs and documents are not used to train AI models and serve only to process requests. AnwaltGPT's pricing page says client data is not used for training, and the security page says the platform ships with learning switched off by default. Section 9 of the Terms of November 2025, the only agreement published, says data is used only to provide and improve the service and does not mention training. The data processing agreement, which the brochure says keeps learning off by default, is sent on request.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
A specific retention period is published and the customer cannot change it.
The security page says every AI interaction is logged with a retention period of at least 24 months, and that requests are processed per session by default. Users can permanently delete all their data, including conversations, documents and case files, from their account settings; the page does not say whether deletion reaches the interaction log, and no period is stated for stored conversations short of the user deleting them.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
Segregation is asserted in public materials with no published detail on how it is enforced.
The workspaces page says permissions are fine grained so that each member sees only what they should, and that knowledge bases and assistants can be shared with chosen roles or individual members; the security page says each customer runs as its own instance with separate database segmentation. No help article or documentation describes the permission model, and nothing addresses walls between matters inside a firm.
Third Party Request and Subpoena Notice
If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?
No located term or policy addresses third party requests for customer data.
The Terms of November 2025, the privacy policy, the security page, the three security documents of 8 September 2026 and the brochure do not address requests from courts or authorities for customer data, and no transparency report is published. The data processing agreement, which the brochure says covers confidentiality and support duties, is not published and is offered by email on request.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
Coverage is described by jurisdiction with no identification of the underlying corpus.
The chat page says the assistant covers all essential areas of German law and was trained on all German statutes and case law, the home page cites more than 1,000 legal sources and answers from real legal databases, and Deep Research searches statutes, regulations, case law, literature, the firm's own knowledge bases and the open web. No database, court collection, publisher or license is named, and the knowledge base is said to be updated regularly with no cadence given.
Good Law Verification
Does the product tell you when the authority it just cited has been overruled?
The product instructs the reader to check the citation without performing the check.
Section 8 of the Terms of November 2025 warns that answers can contain outdated case law and obliges users to check all information themselves, and the brochure says the firm checks sources before use. The chat page says the knowledge base is updated regularly with new legislation and case law; no check of whether a cited judgment or provision is still good law is described.
Refusal and Uncertainty Behavior
What does the product do when the answer is not in the corpus?
No located public material addresses what the product does when it cannot ground an answer.
Nothing on the chat, Deep Research and security pages, in the brochure or in the Terms describes what the product does when it finds no source for an answer. The security page describes a button for users to report an answer as inaccurate, which is a route after the fact rather than abstention.
Fabricated Citation Record
Does a public court record exist addressing fabricated or hallucinated legal citations in output from this product?
No court order, opinion or disciplinary record addressing fabricated or hallucinated legal citations produced by this product has been located as of the date shown. This is a statement about the public record on that one subject, not a finding about the product, and this signal is not a litigation history.
The AI Hallucination Cases database maintained by Damien Charlotin records no case naming AnwaltGPT or its maker Mindverse, and no court order, opinion or disciplinary record naming the product appears in the public record.
Bar Guidance Alignment
Has the vendor engaged in public with the ethics opinions its buyers are bound by?
Public materials engage with at least one named ethics opinion.
AnwaltGPT's home page says it complies with professional secrecy under section 43a of the Federal Lawyers' Act and section 2 of the lawyers' professional code, and that its data processing agreement meets the requirements of the Federal Bar and section 2 of that code; the brochure says the confidentiality agreement refers to section 203 of the Criminal Code and section 43e of the Act. No bar association guidance on AI is named, and only German rules are engaged.
Billing and Fee Posture
Does the vendor address what happens to the bill when the work takes an hour instead of six?
Public materials claim time savings without addressing billing or disclosure, and the product sits inside a fee relationship between a lawyer and a client where those savings would change the bill.
The home page quotes a management consultant saying he can bill an hour while the AI takes about three quarters of an hour of the work off him, and the intake page says the same inquiry becomes a case file, deadlines, pleadings and a statutory fee invoice. Nothing addresses how AI assisted work is recorded, billed or disclosed to clients, and the interaction log is described for AI Act purposes rather than as a record for each matter.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
The material exists behind a sales conversation or an executed agreement.
The security overview of 8 September 2026 names Azure, Bedrock and Hetzner and says the full contractual subprocessor list comes with the contract documents, and the security page says the data processing agreement, with the complete list in Annex IV, is sent by email within 48 hours of a request. A security overview, a data flow diagram and a partner brief can be downloaded without a form.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
Some elements of the record are available, short of a document level export.
Every answer carries a visible AI generated label, every AI interaction is logged for at least 24 months, Deep Research reports export to PDF or Word with their sources, and Cowork returns tracked changes. No record for each document of the model used, sources retrieved and human verification for a court filing, and no guidance on disclosing AI use to a court, is published.