How we research and score legal AI tools

How do we score a tool?

One rubric applies to every tool we cover. It has eight criteria, each scored from 0 to 10 and weighted by how much it matters to a law firm or legal team in Finland. The weights add up to 100. Each criterion has anchored descriptions of what a score of 0, 3, 5, 7 and 10 means, so a score can be traced back to a description rather than to a feeling.

Four of the criteria we score from documentation break down into items, each worth a fixed number of points. A tool’s score on such a criterion is the sum of its item points, and its profile shows the points item by item, with the facts and sources they rest on. The fifth, estimated price, is scored from what the plan costs on a fixed scale, so a cheaper plan scores higher. When the meaning of a score changes, for example because a weight or an item changes, the rubric gets a new version number.

What does each criterion measure?

Legal task quality 15%

How well the tool does legal research and drafting, as published legal benchmarks measure it.

A published benchmark of the tool itself, if one is less than two years old, otherwise the Vals Legal Index score of the tool's model. The score is the accuracy in % divided by 10, rounded.

Legal task quality: what the scores mean

Score	Meaning
0	Fails most tasks: answers get the law wrong or miss the issue, and drafts cannot be used.
3	Handles simple tasks, but makes material legal errors or misses key issues on most of them. Output needs rewriting.
5	Handles routine tasks acceptably. On complex tasks it misses issues or makes errors a lawyer must catch, and output needs substantial editing.
7	Correct and useful on most tasks, complex ones included, with occasional errors or omissions. Output needs the review any draft would.
10	Correct, complete and well reasoned across the benchmark tasks, at the level of a competent lawyer. Output needs only light review.

Grounding and verifiability 10%

Whether answers cite primary sources, whether the citations support the claim, and how the tool guards against invented content. Documentation describes the safeguards but cannot show whether they work, so this criterion rests on a published legal research benchmark.

The Legal Research Bench score of the tool's model. An answer passes only if it meets every item of the grading rubric, including relevant, supporting sources. The score is the all-pass accuracy in % divided by 10, rounded.

Grounding and verifiability: what the scores mean

Score	Meaning
0	No citations, or citations that are invented or do not support the statement.
3	Some citations, but many are vague, do not link through or do not support the statement they are attached to.
5	Most claims cite an identifiable source, but often a secondary or unlinked one, and some citations do not support the claim.
7	Claims cite primary sources with working links, citation errors are rare, and the tool says when it is unsure.
10	Every legal claim cites the exact provision, decision or passage with a working link. Unsupported citations are not found, and the tool flags uncertainty and gaps in its sources.

Finnish quality 10%

The quality of the tool's Finnish output and legal terminology, and whether its interface is available in Finnish.

The EuroEval Finnish rank score of the tool's model, where 1 is best. The score is 10 minus 3 for every point above 1, rounded, so a rank score of 2 gives 7. EuroEval tests general language, not legal terminology or the interface.

Finnish quality: what the scores mean

Score	Meaning
0	Cannot work in Finnish, or its Finnish output cannot be used.
3	Understands Finnish but writes awkwardly, mistranslates legal terms or carries over another legal system's concepts.
5	Mostly fluent Finnish. Legal terminology is mostly right, with errors a lawyer must fix.
7	Fluent, correct legal Finnish with occasional slips in terminology.
10	Native-quality legal Finnish that uses Finnish legal terminology correctly. The interface is available in Finnish.

Finnish legal sources 15%

Which Finnish legal sources the vendor documents as part of the tool, and whether they are linked to each other. Having a source earns 1 point. Most of the score rests on the links, a knowledge graph of the sources: statutes kept with their earlier wordings, the court decisions that applied them and the preparatory works that explain them. Answering a Finnish legal question means reading a provision in force on the right date together with its case law and preparatory works, and a tool that only searches them as separate documents leaves that work to the user.

A source counts when the vendor documents it as part of the plan we score: its own database, a licensed publisher's content, or a named official source the tool searches. General web search and third-party connectors sold separately do not count. A link counts when the vendor documents it as part of the tool: on the tool's own pages, or on the pages of the vendor's own legal database that the tool answers from. Links in a partner's database count only when the vendor documents that the tool uses them. Documentation cannot show how complete a source is.

Finnish legal sources: what the scores mean

Score	Meaning
0	No Finnish legal source is documented as part of the tool.
3	Finnish legislation, case law and preparatory works, but searched as separate documents: no links between them are documented.
5	For example, legislation kept with its earlier wordings, and case law not linked to it, with no preparatory works.
7	For example, legislation kept with its earlier wordings and case law linked to the statutes it applied, with no preparatory works.
10	A knowledge graph of the Finnish sources: legislation kept with its earlier wordings, and case law and preparatory works linked to the statutes they apply or explain.

Finnish legal sources: points per item

Item	Points
Finnish legislation Up to 4	4 Documented as part of the tool with its earlier wordings, so the version in force on a given date can be read. 1 Documented as part of the tool. 0 Not documented, or only through web search or a separately sold connector.
Finnish case law Up to 3	3 Documented as part of the tool, with decisions linked to the statutes they applied. 1 Documented as part of the tool. 0 Not documented, or only through web search or a separately sold connector.
Finnish preparatory works Up to 3	3 Documented as part of the tool, linked to the statutes they explain. 1 Documented as part of the tool. 0 Not documented, or only through web search or a separately sold connector.

EU legal sources 10%

Which EU legal sources the vendor documents as part of the tool: EU legislation, and the case law of the Court of Justice of the European Union.

A source counts when the vendor documents it as part of the plan we score: its own database, a licensed publisher's content, or a named official source such as EUR-Lex that the tool searches. General web search and third-party connectors sold separately do not count. A source the vendor lists as coming does not count until it is listed as available.

EU legal sources: what the scores mean

Score	Meaning
0	No EU legal source is documented as part of the tool.
3	Part of EU legislation, such as the treaties alone, and no EU case law.
5	EU legislation, with no EU case law.
7	EU legislation and part of EU case law, such as selected decisions.
10	EU legislation and the case law of the Court of Justice.

EU legal sources: points per item

Item	Points
EU legislation Up to 5	5 The treaties and EU regulations and directives are documented as part of the tool. 3 Only part of EU legislation is documented, such as the treaties without regulations and directives. 0 Not documented, or only through web search or a separately sold connector.
EU case law Up to 5	5 The case law of the Court of Justice is documented as part of the tool. 2 Only selected EU decisions are documented. 0 Not documented, or only through web search or a separately sold connector.

Security and data handling 15%

Where customer content is stored and processed, whether it is used for training, how long it is kept and who decides, and whether the plan comes with a data processing agreement, independent security certifications and single sign-on.

Each item is scored for the plan we score. An item the vendor does not disclose scores 0, and the page says so. An item we could not check stays unscored, and so does the criterion until it is checked.

Security and data handling: what the scores mean

Score	Meaning
0	None of the items is documented: no EU storage or processing, no commitment against training, no data processing agreement, no retention terms, no certification and no single sign-on.
3	For example, a commitment against training and a stated retention period, but no EU storage or processing, data processing agreement, certification or single sign-on.
5	For example, no training, a data processing agreement and retention the customer controls, but no EU storage or processing, certification or single sign-on.
7	For example, no training, a data processing agreement, retention the customer controls, one certification and single sign-on, but no EU storage or processing.
10	Customer content is stored and processed in the EU/EEA with no documented exception, and the plan has a commitment against training, a data processing agreement, retention the customer controls, ISO/IEC 27001 and SOC 2 Type II, and single sign-on.

Security and data handling: points per item

Item	Points
EU/EEA storage and processing Up to 2	2 Customer content is stored and processed in the EU/EEA, model inference included, by default or as a contracted option, with no documented exception. 1 EU/EEA storage or processing is offered, but not both, or documented features, subprocessors or models handle some customer content elsewhere. 0 No EU/EEA storage or processing is documented.
No training on customer content Up to 2	2 Customer content is not used to train models, without the customer having to opt out. 1 Customer content is used for training unless the customer opts out. 0 Customer content is used for training with no opt-out, or the vendor does not say.
Data processing agreement Up to 1	1 A data processing agreement, with the vendor as processor, is available for the plan. 0 No data processing agreement is documented for the plan.
Retention Up to 2	2 The customer sets how long content is kept and can delete it, or the vendor keeps no customer content beyond the request. 1 The vendor states how long it keeps customer content. 0 Not documented.
Security certifications Up to 2	2 Both an ISO/IEC 27001 certification and a SOC 2 Type II report cover the service. 1 One of them covers the service. 0 Neither is documented.
Single sign-on Up to 1	1 Single sign-on through SAML or OpenID Connect is available on the plan. 0 Not documented, or not available on the plan.

Workflow fit 10%

How well the tool fits a legal team's existing tools and controls: Word, email and document management integrations, agents and workflows, and admin. We score it from the vendor's documentation.

A feature counts when the vendor documents it as available on the plan we score, public betas included, whoever built the integration. Announced features and private previews do not count.

Workflow fit: what the scores mean

Score	Meaning
0	Works only in its own app: no documented Word, email or document management integration, agents, admin controls or API.
3	For example, an API, an email connection and a connection to general file storage, but no Word add-in or legal document management system.
5	For example, works inside Word, connects to a legal document management system and offers an API, but has no agents or admin controls.
7	For example, Word, email and legal document management integrations and agents, but no admin controls or API.
10	Works inside Word and email, connects to a legal document management system, runs agents or firm-built workflows, provisions users, keeps audit logs and offers an API or MCP connection.

Workflow fit: points per item

Item	Points
Microsoft Word Up to 2	2 Works inside Word documents through an add-in or built-in integration. 1 Creates or exports Word files but does not work inside Word. 0 No Word integration is documented.
Email Up to 1	1 Works inside Outlook or connects to the team's email. 0 Not documented.
Document management Up to 2	2 Connects to a legal document management system such as iManage or NetDocuments. 1 Connects only to general file storage such as SharePoint, OneDrive, Google Drive or Box. 0 Not documented.
Agents and workflows Up to 2	2 Carries out multi-step tasks as an agent, or lets the team build its own reusable workflows. 1 Offers pre-built workflows only. 0 Not documented.
User provisioning Up to 1	1 Provisions users from the organisation's identity provider, through SCIM or directory sync. 0 Not documented, or not available on the plan.
Audit logs Up to 1	1 Keeps audit logs that administrators can review or export. 0 Not documented, or not available on the plan.
API or MCP Up to 1	1 Offers an API or MCP connections for linking the tool with the organisation's own systems. 0 Not documented.

Estimated price 15%

What the plan we score costs per user per month, excluding VAT and including any licence the plan requires. A lower price scores higher. Price says nothing about value for money: a cheaper tool may do less, and the other criteria show what it does.

We use the price the vendor publishes for the plan. When it publishes none, we use a price reported by the press, an analyst or a buyer's guide, or work one out from the vendor's other published prices, and label the score an estimate with its assumption. Where reports differ, the profile shows each of them and says which one we used. Fees that depend on how much a team uses the tool are not included, and the profile says when a plan has them.

The score is 10 × log(500 ÷ price) ÷ log(12.5), rounded, where the price is in euros per user per month: €40 or less scores 10, about €140 scores 5 and €500 or more scores 0, so every doubling of the price costs about 2.7 points. €40 is about what a general AI assistant seat costs with a legal source connector added, and we do not tell lower prices apart, nor prices above €500. Prices in US dollars are converted at the European Central Bank reference rate of 23 October 2026, US$1.1193 to the euro, for the score only. Euro reference rates, European Central Bank

Estimated price: what the scores mean

Score	Meaning
0	€500 or more per user per month.
3	About €235 per user per month.
5	About €140 per user per month.
7	About €85 per user per month.
10	€40 or less per user per month, or free.

Which plan do we score?

Vendors often sell several plans, and they can differ on exactly what this rubric measures: where data is stored, whether it is used for training, and which integrations are included. We score the best plan available to a business buyer in the EU, and the profile names it. A consumer plan is never scored in place of a business plan.

Each item is judged on the facts that apply to that plan:

• An item the vendor does not disclose scores 0, and the profile says so.
• An item we could not check, for example because the page needs a login, stays unscored. So does its criterion, which reads “Not yet scored” until we can check it, along with the range it could still take: from the points of the items we checked up to those points plus the rest at full points. The tool’s overall score is then shown as a range too, and the home page chart places it by the lowest end.
• A feature counts when the vendor documents it as available on the plan. Announced features and private previews do not count.

How do we score price?

We score the price of the plan we score per user per month, excluding VAT and including any licence the plan requires, such as a base subscription. A cheaper plan scores higher, on a logarithmic scale: €40 a month or less scores 10, about €140 scores 5 and €500 or more scores 0. A few euros more costs little, while a tool that costs ten times as much scores about 9 points lower. The rubric below gives the formula.

• The vendor publishes the price. We use it, and the score rests on documentation.
• The vendor publishes no price for the plan. We use a price reported by the press, an analyst or a buyer’s guide, or work one out from the vendor’s other published prices, and label the score an estimate with its assumption. Where reports differ, the profile shows each of them and says which one we used.
• Part of the price depends on use. Fees billed by usage are not counted, because they depend on how much a team uses the tool. The profile says so, and that the real price is higher.
• The tool works inside an AI assistant. A pair such as “Claude + Laki.ai” is scored on the two prices added up, because you pay for both. Because the scale is logarithmic, a low-cost connector changes the pair’s score little or not at all.

For the score only, we convert prices in US dollars to euros at the European Central Bank reference rate the rubric names. Everywhere else prices stay in the vendor’s own currency. The pricing guide compares the prices side by side.

Price says nothing about value for money. A cheaper tool may do less, which is why the other criteria carry most of the weight and show what each tool does.

How do we score the criteria that rest on benchmarks?

We do not test tools ourselves. Three criteria, legal task quality, grounding and verifiability, and Finnish quality, are about the answers a tool gives, so we score them from published benchmarks, the same benchmarks for every tool. Every score names the benchmark it comes from and links to it, on the tool’s profile and under the score chart on the home page.

• Legal task quality comes from a published benchmark of the tool itself, if one is less than two years old. Otherwise it comes from the Vals Legal Index result of the tool’s model.
• Grounding and verifiability comes from the Legal Research Bench result of the tool’s model. An answer passes only if it meets every point of an expert grading rubric, including relevant sources that support it.
• Finnish quality comes from the EuroEval Finnish leaderboard for the tool’s model.

The rubric below shows how each result becomes a score from 0 to 10. Every profile shows the benchmark, the model or product it measured, the result, the date on the benchmark’s page and the date we read it.

Which model we look up follows fixed rules:

• The vendor names the models on the plan. We use the most capable one by list price. If a benchmark does not list it, we use the next model on the plan that the benchmark lists.
• The vendor names models, but not which one answers. We assume one of the named models answers, use the lower of their results when they cost the same, and label the score an estimate.
• A benchmark lists none of the plan’s models. We use the vendor’s newest model that it lists, and label the score an estimate.
• The vendor does not name the model, and no benchmark covers the tool. We assume a low-cost general-purpose model, OpenAI’s GPT-6 Luna, and label the score an estimate. A stronger model would score higher, so these estimates are deliberately cautious.

Every estimate states its assumption next to the score. When a vendor names its model or a benchmark covers the tool, we replace the estimate.

These benchmarks have limits that matter here. The legal benchmarks test US law in English, not Finnish or EU law, and they test a model on its own, not the vendor’s sources and safeguards around it. The language benchmark tests general Finnish, not legal terminology. So a benchmark score shows how capable the engine is, not how well the tool handles a Finnish matter.

A tool’s overall score is the weighted average of the scores of the criteria that apply to it. We show it only once every one of those criteria is scored, and we label it when any of them is an estimate. Three of the criteria behind every overall score rest on published benchmarks, so an overall score can change when a benchmark publishes new results.

How do we score legal sources?

Finnish and EU legal sources are separate criteria, because a Finnish legal question usually needs both, and a tool can be strong in one and weak in the other.

For Finnish sources, having a source is not enough. Answering a Finnish legal question means reading the provision in force on the right date together with the court decisions that applied it and the preparatory works that explain it. So having legislation, case law and preparatory works earns 1 point each, and the other 7 points go to documented links between them: legislation kept with its earlier wordings, and case law and preparatory works linked to the statutes they apply or explain. A tool that documents all three sources but no links scores 3. A tool that documents the full knowledge graph scores 10.

For EU sources, EU legislation and the case law of the Court of Justice each earn half of the score. A source the vendor lists as coming does not count until it is listed as available.

Which criteria apply to which tools?

Not every criterion makes sense for every kind of tool. A translation tool, for example, does not research the law, so it is not scored on Finnish or EU legal sources. The rubric lists which criteria apply to each type of tool. A criterion that does not apply reads “Not applicable”, and an overall score re-weights the criteria that do apply, so no tool loses points for a criterion that does not fit it.

How do we score a tool that works inside an AI assistant?

Some tools have no assistant of their own. A legal source connector adds legal sources to an AI assistant you already use, and the assistant writes the answers. Scored alone, such a tool would be credited or blamed for an assistant it does not control. So we score it only together with each assistant it works in, as a pair such as “Claude + Laki.ai”, and never show it on its own in a score chart or comparison. The pair gets each criterion’s score by a fixed rule:

The tool’s own profile shows each pair’s scores, with the tool’s own points item by item next to the pair’s, and for price the two prices that add up to the pair’s.

Why does a pair score close to its assistant?

A pair’s score can differ from its assistant’s only on Finnish sources, EU sources and price. That follows from how we score, not from anything we found about the connector:

• Answer quality. Task quality, grounding and Finnish quality come from benchmarks that run the assistant’s model on its own, without any legal source connector. They cannot show whether answers improve when the model reads Finnish statutes, case law and preparatory works instead of searching the web. The same limit applies to every tool: the benchmarks do not test a vendor’s own source library either.
• Sources. A pair counts every legal source, and every link between sources, that either tool documents, so a connector shows its full effect on the Finnish and EU sources criteria.
• Security. A pair takes the assistant’s security score, because the assistant keeps your conversations and files. The connector’s own data handling, such as where it stores data and which certifications it claims, is set out on its profile.
• Price. You pay for both, so a pair is scored on the two prices added up.

Where do our facts come from?

We research only from official sources: the vendor’s own pages, documentation, terms and trust centre, official registers such as the Finnish business register (PRH/YTJ) and the global LEI register, regulators, and primary law. Benchmark results come only from the benchmark publisher’s own pages. We use other comparison sites for ideas about structure only, never as a source of facts. We never use a reviewed tool’s own connector or data service to research or verify facts about any tool.

Every fact on a profile carries a label that says how we know it:

Every fact row shows the date we last checked it. Every profile shows “Last verified”, the date of its oldest fact check, and “Last updated”, the date its facts, scores or text last changed. Every price is labelled Official or Reported. We record every gap, whether not disclosed or not yet verified, and re-check it.

How do we keep facts up to date?

A weekly check lists every fact whose last check is more than 90 days old, and we re-check it against its source. When a fact changes, its profile shows the new value and the new date, and any score that rests on it is re-assessed.

Which tools do we cover?

A tool is included when:

• it is marketed to or used by legal professionals in Finland or the EU, or it is a global benchmark that lawyers here compare against;
• it is generally available, not a preview limited to one region;
• the company that runs it can be identified.

Vendors do not pay to be included, and we do not contact them for access or fact checks. Everything we write rests on what they publish.

How do we rank tools?

A list that ranks tools names the criteria it ranks by, right next to the list. It ranks only by criteria that have been assessed for every tool on it, and only tools of comparable types. A tool with an overall score is never ranked against tools without one by that score. No placement can be paid for.