AI Legal Index
Reference

The 12 legal signals

The AI Legal Index grades every vendor on fifteen capability axes that resolve to a letter. That shape works for capability and it fails for a class of question the legal buyer actually asks, because those questions do not have degrees. Whether a vendor may train on a client’s material has an answer, not a grade. Whether a court record exists is a fact about the public record.

So the index carries a second surface. 12 signals, recorded rather than graded, each with a published value set, a source basis and a date. They are recorded for every vendor in the index, including the ones where the honest value is that nothing was located.

A 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.

Confidentiality and Privilege

What happens to client material once it enters the product, and what the vendor has committed to in writing.

Client Data in Training

Can material a lawyer puts into this product be used to train a model?

This is the first question a general counsel asks and the one most often answered in a sales call rather than in a contract. The distinction that decides it is not whether the vendor says it respects confidentiality, it is whether the commitment sits in the terms a client can hold the firm to. An opt out buried in an admin setting is a different risk posture from a contractual prohibition, and a marketing page that says "your data is secure" while the terms reserve a license is a third thing again.

Recorded values
  • Never, in the contract

    The published terms prohibit training on customer content. Not a policy page, the agreement.

  • Purpose limited, in the contract

    The customer agreement or data processing addendum contractually limits use of Customer Data to providing the contracted service, and no surface names training either way. The limit is bound, which a policy page is not, but it is not an express training prohibition. If any surface names training in either direction, one of the other values is true and this one is not.

  • Never, in policy only

    A public policy or trust page states no training on customer content, with no matching term located in the published agreement.

  • Opt out

    Training occurs unless the customer turns it off.

  • Opt in

    Training occurs only where the customer has affirmatively enabled it.

  • Permitted, in policy only

    Public material states that customer content trains, refines or personalizes models, with no matching term located in the published agreement. Any de identification, anonymization or aggregation qualifier is recorded in the summary.

  • Permitted, in the contract

    The published agreement expressly reserves a right to train on customer content, with no opt out located. Any de identification, anonymization or aggregation qualifier is recorded in the summary.

  • Terms silent

    A published agreement or policy exists and none of it addresses the question either way, or the document that would answer it could not be read and the summary names the retrieval limit. The summary states which shape the silence takes: an improvement right granted that never names training, or no improvement right granted at all.

  • No agreement published

    No customer agreement, terms of service or equivalent contract is published on any surface located, and no policy page states a position on training. Nothing is granted and nothing is withheld, so a client has no term to hold the firm to. Where a policy page does state a position, the row takes the matching policy value instead and the summary records that no agreement exists.

Prompt and Output Retention

How long does the product keep what a lawyer typed, and can that be set to zero?

A prompt history is a record. Records are discoverable, subject to legal hold, and reviewable in a malpractice action. A firm that has not asked how long its associates’ prompts persist has created a repository it does not know it owns, containing draft reasoning it would never have put in an email. The buying question is not whether retention exists, it is whether the retention window is disclosed, whether the customer controls it, and whether the vendor will say plainly what it produces if that history is ever requested.

Recorded values
  • Customer set, zero available

    The customer sets the retention window and no retention is an available setting.

  • Customer controlled, no zero option

    The customer controls the retention window, by product configuration or by contractual instruction, but zero retention is not stated as available.

  • Disclosed fixed window

    A specific retention period is published and the customer cannot change it.

  • Disclosed without a period

    Retention is acknowledged in public materials with no stated period.

  • Not addressed

    No located public material states how long prompts and outputs are retained.

Ethical Walls and Matter Segregation

Does retrieval respect the firm’s ethical walls, or can the model read across them?

A firm spends real money maintaining walls in its document management system so that the team on one side of a matter cannot reach the other side’s files. A retrieval layer that indexes the whole corpus and answers from all of it defeats that control silently, and the failure surfaces as a conflict, not as an error message. The question is whether the product inherits the existing access model per user and per matter at query time, or whether it built a second permission system that has to be kept in step by hand.

Recorded values
  • Inherits document system permissions

    Retrieval enforces the source system access model at query time, per user, and the vendor documents it.

  • Own model, documented

    The product maintains its own permission model, documented, requiring the firm to keep it aligned.

  • Claimed, not documented

    Segregation is asserted in public materials with no published detail on how it is enforced.

  • Not addressed

    No located public material addresses walls or matter level segregation.

Third Party Request and Subpoena Notice

If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?

Privilege is asserted by the client, not by a software company. If a vendor can be served and can produce client material without telling the firm, the firm loses the chance to move to quash and may never learn that the material left. A published commitment to notify unless legally prohibited, and a transparency report showing what has actually been received, is the difference between a stated principle and an operating practice.

Recorded values
  • Notice committed, report published

    Terms commit to notifying the customer where lawfully permitted, and a transparency report is published.

  • Notice committed

    Terms commit to notice where lawfully permitted. No transparency report located.

  • Discretionary

    Terms reserve discretion over whether to notify.

  • Disclosure addressed, notice absent

    Published terms or policy address disclosure to authorities or in response to legal process, and no commitment or reservation regarding customer notice is located anywhere. The vendor has told the customer that data can leave and has said nothing about whether the customer hears of it.

  • Not addressed

    No located term or policy addresses third party requests for customer data.

Accuracy and Authority

Where the law comes from, whether it is still good law, and what the product does when it does not know.

Primary Law Corpus Provenance

Where does the law in this product come from, and does the vendor have the right to use it?

Two separate risks sit on this one question. The first is coverage: an answer drawn from a corpus missing a jurisdiction or lagging by months is confidently wrong in a way the reader cannot detect. The second is title. Legal publishing has already produced litigation over whether one company may build a research product on another’s editorial content, and a buyer who cannot say where a vendor’s case law came from cannot price the risk that the corpus is enjoined. Naming the source, the license and the update lag is a low bar that most of the market does not clear.

Recorded values
  • Sources named and licensed

    The vendor names its primary law sources and the license or public domain basis for each, with an update cadence.

  • Sources named, basis unstated

    Sources are identified without stating the license or rights basis.

  • Jurisdictions only

    Coverage is described by jurisdiction with no identification of the underlying corpus.

  • Not addressed

    No located public material identifies the corpus behind the product’s answers.

Good Law Verification

Does the product tell you when the authority it just cited has been overruled?

A citation can be real, correctly quoted, on point, and dead. Checking subsequent history is the most routine verification step in legal research and the one a generative product is least likely to have performed, because retrieval optimizes for relevance rather than for currency. A tool that returns a vacated opinion with no treatment flag has produced a failure that looks nothing like a hallucination and carries the same consequence. For registered rights the equivalent question is currency rather than subsequent treatment: whether a patent is still in force, or a mark still live rather than abandoned or canceled. A status determination the vendor surfaces with the record, including status taken from the issuing office and named as such, is its own treatment signal; the licensed value stays reserved for a named commercial citator.

Recorded values
  • Licensed citator, surfaced inline

    Treatment signals come from a named commercial citator and appear with the authority.

  • Own treatment signal

    The vendor computes and surfaces subsequent history itself, with the method described.

  • Prompts the user to verify

    The product instructs the reader to check the citation without performing the check.

  • Not addressed

    No located public material addresses whether authority is checked for subsequent history.

Refusal and Uncertainty Behavior

What does the product do when the answer is not in the corpus?

Every fabricated citation that has reached a filed brief passed through a moment when the correct output was "I could not find support for that." Whether a product is capable of producing that sentence is a design decision, it is testable, and almost nobody publishes it. This signal records whether the vendor documents the behavior and whether it is demonstrable, because a confident answer to an unanswerable question is the single failure mode this profession cannot absorb.

Recorded values
  • Documented and demonstrable

    The vendor documents an explicit no answer path and it is observable in the product or in published evaluation.

  • Documented

    The vendor describes refusal or abstention behavior in public materials.

  • Confidence signal only

    The product exposes a confidence or grounding score without an explicit abstention path.

  • Not addressed

    No located public material addresses what the product does when it cannot ground an answer.

Fabricated Citation Record

Does a public court record exist addressing fabricated or hallucinated legal citations in output from this product?

Courts have now issued a substantial body of orders addressing filings containing citations that do not exist, and those orders are public records. Most name the filer rather than a product. Some name a product. The index records what the record says and nothing beyond it, because the distinction between a lawyer who filed without checking and a tool that misrepresented its own grounding is exactly the distinction a reader needs and exactly the one a summary destroys. An adverse value here requires a citation to the deciding court and the date. Nothing located is the most common value and it is a statement about the record, not about the product.

Recording rule

An adverse value on this signal requires a citation to the deciding court and the date, and the recorded summary describes what the record says rather than what the index concludes. No citation, no adverse value.

Recorded values
  • None located

    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.

  • Record addresses filer conduct

    A court record exists in which the product is mentioned and the court’s findings address the filer’s verification failure rather than the product.

  • Product named in findings

    A court record exists in which the court’s findings address fabricated or hallucinated legal citations in the product’s output directly. Allegations that survive a motion to dismiss are not findings. Citation required.

  • Vendor responded publicly

    A record exists and the vendor has published a substantive response or remediation. Both are linked.

Professional Responsibility

The obligations that attach to a lawyer rather than to software, and whether the vendor has engaged with them in public.

Bar Guidance Alignment

Has the vendor engaged in public with the ethics opinions its buyers are bound by?

The professional obligations around generative AI are now written down. The American Bar Association issued a formal opinion on generative AI in 2024 and a growing number of state bars have followed with their own, and they differ on the points that matter, including client consent and confidentiality. A lawyer stays responsible for compliance whatever the vendor says. What separates vendors is whether they have read the guidance their customers are bound by and published something a firm can hand to its own risk committee, or whether they have left every buyer to do that work alone.

Recorded values
  • Mapped by jurisdiction

    The vendor publishes a mapping of its product to named ethics opinions across more than one jurisdiction.

  • Named guidance addressed

    Public materials engage with at least one named ethics opinion.

  • Generic reference

    Public materials refer to professional responsibility in general terms without naming guidance.

  • Not addressed

    No located public material engages with bar or ethics guidance.

Billing and Fee Posture

Does the vendor address what happens to the bill when the work takes an hour instead of six?

This is the most commercially uncomfortable question in legal AI and the reason it is worth tracking. Ethics guidance is direct that a lawyer bills for time actually spent, so a tool that compresses six hours into one has changed the invoice, not just the workflow. A product that emits a defensible record of what it did and how long it took gives the firm something to disclose. A product that markets hours saved to the buyer while staying silent on the client’s side of that equation has handed the firm a problem with the marketing material attached.

Recorded values
  • Audit record and guidance

    The product emits a per matter record of AI assisted work and the vendor publishes guidance on fee and disclosure treatment.

  • Audit record only

    A usable record of AI assisted work exists with no published fee guidance.

  • Savings claims only

    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.

  • Fee record is the product

    The buyer is the payer, not the biller. The product exists to review, author or benchmark legal fees, so the record of fees is the product itself rather than a byproduct of AI assisted lawyering. Typically sold to in-house teams paying outside counsel.

  • Client cost pass through

    The vendor’s charge reaches the client as a disbursement or case expense, typically advanced by a contingency firm and recovered from a settlement, so the fee question is disclosure of a pass through cost rather than compressed hours. The summary records whether the vendor addresses that disclosure.

  • Outside the fee relationship

    The product does not touch a fee between a lawyer and a client. It operates before an engagement exists, or it is bought by a team that bills no client for the work. Savings claims aimed at the buyer’s own cost are recorded in the summary and do not make the row a savings claim, because no client bill is in the loop.

  • Not addressed

    The product sits inside a lawyer to client fee relationship and no located public material addresses billing, fee or disclosure treatment, with no savings claim published either.

Outside Counsel Guideline Readiness

Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?

Corporate clients have started writing AI terms into outside counsel guidelines, which reverses the usual direction of diligence: the firm now has to evidence its vendor to its own client. The practical artifacts are a current subprocessor list, a statement of which model providers see client content, and a consent or notification pack the firm can forward. Vendors who have assembled that get through procurement. Vendors who have not become a reason the firm loses the matter.

Recorded values
  • Disclosure pack published

    A subprocessor and model provider list plus client facing disclosure material is published or available without an agreement in place.

  • Subprocessors listed

    A current subprocessor or model provider list is published.

  • On request only

    The material exists behind a sales conversation or an executed agreement.

  • Not addressed

    No located public material supports a client side disclosure obligation.

Court Disclosure Support

If a judge’s standing order requires an AI disclosure, can the product produce one?

A number of judges now require filers to disclose whether generative AI was used and to certify that citations were verified by a human. Answering that honestly requires knowing which model produced which passage, what was retrieved, and who checked it. Most products cannot say. A vendor that can export that record has turned a compliance obligation into a two minute task, and one that cannot has left the certifying lawyer to reconstruct it from memory. Where no court sits in a product’s path, the signal is read through the body that does ask: the examiner or regulator a compliance product answers to, on the same test of whether the record can be produced rather than reconstructed.

Recorded values
  • Exportable disclosure record

    The product exports a per document record covering model used, sources retrieved and human verification.

  • Partial record

    Some elements of the record are available, short of a document level export.

  • Guidance or template only

    The vendor publishes disclosure guidance or a template without product support for producing the record.

  • Not addressed

    No located public material addresses court disclosure or verification certification.

Eleven questions this data answers
Which AI legal research tools publish their hallucination or accuracy rates?

Every vendor graded on Citation Accuracy and Hallucination Disclosure, with the published numbers, the documented grounding and the silence all named.

Do legal AI vendors train their models on client data?

The Client Data in Training record across the index: contractual prohibitions, policy statements, reserved rights and the silent majority.

What happens to privilege when a law firm uses a third party AI tool?

Privilege and Confidentiality Posture across the index, with ethical wall enforcement and subpoena notice recorded on every vendor.

What do the bar rules require before a lawyer uses AI on client matters?

The four professional responsibility signals across the index: bar guidance engagement, billing posture, court disclosure support and client AI clause readiness.

Which legal AI vendors publish SOC 2 or ISO 27001 certification?

Attestation a buyer can open, separated from a certification nobody can read and a badge with nothing behind it.

Which AI legal tools integrate with iManage and NetDocuments?

Documented integration depth crossed with whether retrieval respects the ethical walls the document system is enforcing.

Which legal AI vendors offer on premise or private cloud deployment?

Where the software runs, where the data sits, and the model supply chain question a deployment page almost never answers.

Which legal AI vendors disclose which AI models they use?

Models, providers, where inference runs and notice on change, crossed with where the primary law corpus came from and who licensed it.

How should a law firm bill for AI assisted work?

Which products emit a per matter record of AI assisted work, which vendors say anything about fee treatment, and which publish a price to attach to it.

Can legal AI tools produce an AI disclosure record for a court?

The three lines a standing order asks for, a specimen of the record that answers them, and which products can fill each line, crossed with whether the law was checked and whether the product has met a court already.

Which legal AI vendors publish their pricing?

Every vendor graded on Commercial Transparency, the published entry prices on record in the vendor’s own currency, and which vendors let a firm read both the price and the customer agreement before a sales call.

A note on absence

Most signals on most vendors will read not addressed at first, and that is a finding worth publishing rather than a gap worth hiding. This market sells to a profession governed by written duties, and the number of vendors who have published a term, a retention period, a subprocessor list or a corpus source that a buyer can check is smaller than the number who describe themselves as secure. The index records the difference and dates it, so a vendor that publishes tomorrow can be redated the day it does.

If a record here is wrong or out of date, send the source through the contact form and it will be reviewed. Grading standards, bands and limits are on the methodology page.

Contact

Correct a record, or ask how something was graded

Every grade and every signal on this index is drawn from public sources and dated. If a record is wrong, out of date, or missing an artifact the index did not locate, send the source and it will be reviewed and the record redated. Vendors are welcome to submit documentation. Nothing on this index is for sale, including a listing, a placement, or a grade.

AI Legal Index

The AI Legal Index is an independent index that tracks changes to AI vendors in legal. It holds 430 vendors across 9 categories, each graded on the same 15 capability axes and recorded against 12 legal signals, from public sources. No vendor pays for inclusion, placement, or rating.

Index Status
Last index update
October 14, 2026
The AI Legal Index is an editorial reference. It is not a regulatory body, not a law firm, and nothing published here is legal advice or a recommendation to retain or avoid a vendor. Records are verified against published sources, bar guidance and public court records. Where a record reads not addressed, the material was not located in public sources on the date shown. See the Methodology page for evaluation standards and limitations.
© 2026 AI Legal Index
3801 N Capital of Texas Hwy, Ste E240 · Austin, TX 78746