M

Minesoft PatBase

The models drive several named capabilities on a platform that works fully without them. The AI features are real and recent, mostly from 2025 and 2026. Document Chat answers questions about a single patent, AI Explain shows where a query overlaps a document, and Smart Images reads reference numerals in drawings. Users can also train alert classifiers, natural language queries are expanded with relevance scores, the Smart Claims viewer separates claim types, and machine translation is built in.

Underneath is a platform in market since 2003 from a company founded in 1996. Its core is a curated collection of more than 180 million publications from over 100 authorities, with family level organization and daily data quality checks. Boolean, proximity, wildcard and command line searching is still sold to professional searchers. Remove every model and a searcher keeps the data, the family structure, the classification tools and the syntax. The AI changes who else can use it and how quickly.

Grounding is real and documented in detail, with no measured accuracy. AI Explain shows the user exactly where their input overlaps a document, highlighting the matching sections and wording, and the vendor says the purpose is to let a searcher judge a result's relevance and see why it appeared. That is attribution at the passage level for retrieval results. Relevance scores come with results, and Document Chat works on a single patent, so its answers have a bounded, visible source.

No precision or recall figure is published for the AI search, no test set is described, Document Chat has no published evaluation, and no failure mode is named. The agreement runs the other way. Clause 11 disclaims any warranty that the Information is accurate or complete, and says the licensor does not guarantee the accuracy of the Product or the Information or accept liability for the consequences of their use.

Users have real control over how the models behave, short of a full control structure. AI Explain lets a user see why a result surfaced, showing the overlap between query and document rather than presenting a ranking as a verdict. AI alerts can be trained in bulk from a user's own folders and search history, so the classifier deciding what reaches them is tuned by their own past judgments. Relevance scores give a graded signal, letting a searcher decide where to stop.

No threshold is stated at which any feature acts without review, and nothing separates AI assisted from conventional operation although both are available. Nothing describes what happens when a Document Chat answer is wrong, and there is no published route for correcting or reporting an output. There is also no general statement that the tools support rather than replace professional judgment. The controls are practical rather than declared.

Scale and history are published, without a named customer or measured outcome. The vendor says more than 90,000 innovators and legal professionals use the platform, that it has served patent professionals since 1996, and that PatBase launched in 2003. The data covers more than 180 million publications across over 100 authorities, support is offered in fifteen or more languages across all time zones, and a public status page is maintained.

None of that shows deployment results. No customer is named on the pages read, no case study section appears in the navigation, no testimonial is attributed to a named person or organization, and no figure is published for time saved, recall improved or cost reduced by any AI feature. Feature announcements describe benefits such as speed and less noise without measurement. The only third party marker located, a company registration number, identifies the company rather than the product.

Several direct commitments are published, some in contract, and privilege is not addressed. On the product pages, customer data is never used to train models, searches and results belong to the customer, and the platform is a private cloud storing data securely under data protection law. In the agreement, clause 19 sets a mutual confidentiality duty lasting five years after termination and limits use of the other party's confidential information to performing the agreement.

It also requires notice before any legally compelled disclosure where the law allows. Clause 20 commits both parties to the UK Data Protection Act 2018 and the GDPR, and clause 4 lets the customer keep an archive copy of retrieved information for twelve months after termination. Privilege and work product are not named. No model provider is identified for the generative features. No retention or deletion position covers the customer's own searches, queries or Document Chat exchanges, which a competitor would most want to see.

A contractual disclaimer sits in the terms, while the marketing describes the product in decision terms and the audience goes beyond lawyers. Clause 11 supplies the Product and Information as is, disclaims any warranty that the Information is accurate or complete, and says the licensor does not guarantee accuracy or accept liability for the consequences of use. Clause 10 adds that features relying on patent office services are outside the licensor's control and that it takes no responsibility for retrieving Information, a pointed disclaimer for a prior art tool.

Yet the product is marketed for freedom to operate analysis, patent due diligence and invalidity screening, all of which produce legal conclusions, to legal teams, IP experts, professional searchers and R&D teams together. Nothing addresses where the tool stops and professional judgment begins, supervision or competence, or any rule of professional conduct. Nothing covers an R&D user acting on a Document Chat answer about claim scope without counsel.

Individual AI commitments are published without a governing framework. Customer data is never used to train models, processing takes place in a private cloud on UK servers, and the data centers are said to meet ISO 27001. AI Explain is presented as a transparency feature, and the autumn 2025 release grouped changes around transparency, so explainability is a design goal. Beyond that, nothing is published. No AI policy or responsible AI page appears in the site navigation; the dedicated AI section of the site was not read and may say more.

No framework is cited, no person or body is named as accountable for model behavior, no testing before release is described and no evaluation result is published. Bias is not addressed. That matters here because the alert classifier learns from a user's own past selections, which by design reinforces what that user already looks at, and nothing published addresses the narrowing that can cause.

Published policy covers most of the ground. The data position is stated plainly and repeatedly: customer data is never used to train models, searches and results belong to the customer, and information is stored securely and privately in a private cloud on UK servers. The data centers are said to meet ISO 27001. Clause 6 of the agreement makes the customer responsible for the security of its User IDs and requires immediate notice of unauthorized use, with an audit right for the licensor on seven days' notice.

Clause 19 carries a five year confidentiality tail, and clause 20 commits to the UK Data Protection Act 2018 and the GDPR. A public status page shows availability. Three things are missing. No subprocessor list is published; RWS is named only as a translation development partner, not as a processor of customer data. No retention period or deletion commitment covers searches, queries or Document Chat exchanges. No breach notification commitment to customers is published; the agreement puts the notification duty on the customer only.

A published agreement covers the commercial points on liability, with nothing specific to AI. The agreement can be downloaded in full without a login and is short enough to read. Clause 11 warrants that the Product will operate materially in line with its description or technical specification, before disclaiming all other warranties. Liability is capped at the license fee paid in the preceding twelve months, with the usual exclusions for lost profits, revenue or business and for indirect, consequential, special, punitive or exemplary loss.

Clause 13 requires the licensor to indemnify the customer against claims that the Product infringes a third party's intellectual property, subject to notice, control of the defense and no admissions, with a matching customer indemnity for use of the Product. Termination is for material breach with fourteen days to remedy, under English law and courts. Nothing addresses the accuracy of a generated answer or risk from a decision based on one, and no insurance position is published.

Integration routes are named as products, without documentation or a named outside system. A Patent Data API is sold as its own product line, so programmatic access is offered. The site also lists an MCP entry among its AI applications, which, if shipped, would let a customer's own assistant call the platform. That page was not read, and the site also advertises an early access program for an agentic tool, so whether MCP is generally available is unclear.

Within the suite, PatBase sits alongside Minesoft Origin, PatDocs and ChemX, and analytics output, including charts, heat maps and clusters, can be exported and shared. No connector catalog is published, no outside system is named as supported, no API documentation was read, and nothing describes how results would reach a docketing or IP management system. The API and MCP pages may say more.

Region and tenancy are both addressed, at least in outline. The vendor states repeatedly that its servers are in the United Kingdom, offered as a reason customer data is secure. That is a direct answer for a European buyer with residency obligations, and an equally clear one for a buyer who needs US or Asian residency. Tenancy is addressed in substance: the platform is described as a private cloud in which a customer's searches and results are its own, and the 2020 rebuild is described as a cloud move led by RWS for resilience and security.

Data centers are said to meet ISO 27001. No deployment options are offered or refused, so nothing says whether single tenant, private or on premise arrangements are available, and no other region is offered. The private cloud description states an outcome, not an architecture, and nothing explains how customers are kept separate. No data center operator is named.

A standard is named, but it is unclear what it covers, and no scope, date or evidence is given. The full disclosure is one sentence: the vendor's security commitment is backed by ISO 27001 certification, ensuring its data centers meet international information security standards. Read strictly, that attaches the certification to the data centers rather than to Minesoft's own security management system, and a hosting provider's certificate does not attest to the vendor's own controls.

Nothing says which is meant. No certificate number, certification body, issue or expiry date, scope statement or route to obtain the certificate or a report is published, and there is no trust center or security documentation page. What exists is practice rather than attestation: the private cloud, the UK server location, and the User ID security and audit terms in clause 6 of the agreement.

The AI features are described by function without saying what performs them. Seven are named: Document Chat, AI Explain, Smart Images, the trainable alert classifier, natural language query expansion with relevance scores, the Smart Claims viewer and machine translation. No model, version or architecture is given beyond references to image recognition and claim recognition, and no provider is named for the generative feature.

That matters most for Document Chat, where a third party foundation model would usually sit. Nothing says where inference runs relative to the UK servers, what any provider keeps of a Document Chat question, or whether customers would be told of a model change. Clause 10 of the agreement reserves the right to change the Product and its features without notice. One supplier is named: RWS, which develops the platform with Minesoft, provides machine translation and is said to be improving it continuously. That disclosure covers translation only.

The unit and structure of charging are published in the agreement, and no figure appears anywhere. A buyer can see the shape of the deal before speaking to anyone. Licensing is per User ID, and extra User IDs can be added on written request, ending with the existing term and charged pro rata from the month added. The term is twelve months unless the Product Terms say otherwise, renewing automatically for one year periods unless either party gives thirty days' written notice.

Invoicing is annual in advance. Late payment carries interest at eight percent above the Bank of England base rate plus collection costs, and the licensor may suspend access without notice while payment is outstanding. Clause 17 says Product Terms prevail over the general Terms in a conflict, so the commercial specifics live there. No price, band, minimum or range is published, and the only route is a demonstration request.

Coverage is described in substance across data, jurisdictions and workflows, and its limits are left open. The vendor cites more than 180 million publications from over 100 issuing authorities, full text for the major jurisdictions, family and publication level organization, and a live feed with daily quality checks. Machine translation, often available the same day as a new document, covers languages for cross border prior art. Standard essential patent work is covered in recent detail: a spring 2026 release added fourteen standards bodies by name, IEEE, ISO, CCSA, OMA, IMT, OneM2M, TIA, ATIS, ITU-T, ITU-R, IEC, IETF, ANSI and TTA, with dedicated search syntax.

Named use cases span search, monitoring, collaboration, due diligence, freedom to operate, competitor intelligence and technology landscaping. They are aimed at IP experts, legal teams, R&D teams and professional searchers, with support in fifteen or more languages. No jurisdiction is identified as thinly covered, no technology area as better or worse served, and no coverage dates or backfile depth are given.