Open for review This is version 1.0, shared with invited reviewers. Comments that remove a capability are as useful as comments that add one. Submit feedback on this document.

Functional agenda

Brief Help Question Answering: Functional Agenda

This is the list of what a system needs to be able to do when someone asks a legal question and needs an answer, an action plan, deadlines, forms, and a referral.

Version
1.0
Status
Draft for review
Updated
August 2026
Cite an item as
BHQ-4.2 (brief-help-qa v1.0)
Back to the Standards library
Submit feedback

Start here

  • What this is

    A checklist of what a legal question answering system needs to be able to do, for the public, for navigators, and for junior staff.

  • What to do with it

    Use it to scope a build, to review a chatbot already running on your site, or to write the requirements a vendor has to answer.

  • What it does not do

    It endorses no product and sets no passing scores. The content the system retrieves from has its own data standard.

  • Where to start

    Press Required and critical only. The critical items are drawn from published audits of live systems, and they are where harm has been found.

The full explanation of this document type, including how it is versioned and published, sits on About functional agendas.

Capability list

Filter the list to the levels you care about, or search it. This agenda has no scope selector, so every capability applies to every deployment.

What this workflow covers, and the principles behind the task areas

A person has a legal question. They type it or speak it in their own words. The system returns an answer grounded in content the deploying organization stewards, along with the steps to take next, the deadlines that apply, the forms involved, and the human services that can help.

Three audiences ask these questions, and they need different things from the same system.

AudienceWhat changes for them
A member of the publicPlain language, an action plan they can follow alone, and an honest refusal when the matter is beyond self-help.
A navigator or community advocateReferral accuracy, eligibility detail, and a clear account of what to tell the person they are helping.
A junior staff member or volunteerCitations to authority, the reasoning behind the answer, and a statement of what to verify before relying on it.

The system serves all three by varying the presentation and the level of citation. The underlying answer stays the same.

Where this workflow ends and another begins

This workflowA different workflow
Starts from a question and ends with an answer and a planVoice AI intake starts from a caller and ends with an eligibility decision and a case record
Recommends the right form and explains itDocument assembly fills the form out
Names the service that can helpReferral routing scores and ranks services against capacity
Explains what a document means when askedDocument diagnosis takes an uploaded document as its starting input

A single deployed product often crosses these lines. The agendas stay separate so that conformance can be tested one workflow at a time.

The five principles behind the task areas

PrincipleWhat it requires
GroundedEvery substantive statement comes from content the deploying organization stewards and can correct. The system does not answer from the model's general knowledge.
ActionableThe person leaves knowing what to do, by when, and with what. An answer that is correct and leaves the person no better able to act has failed.
BoundedThe system knows what it does not cover, says so plainly, and sends the person somewhere real.
SupervisedA person can see what the system produced, review a sample of it, and correct it through content or configuration rather than through a rebuild.
AccountablePerformance is measured against a fixed set of questions, the measurement is repeated after every change, and the record can be shared.
On the tension between grounded and actionable

Research with 46 community members and 21 legal experts found that actionability matters more to users than precision. A vague answer that is technically correct scores worse than a specific answer that is slightly imprecise. That finding sits uncomfortably beside a grounding requirement, and this agenda resolves it in one direction. The system should be specific about what the content supports and silent about what it does not, rather than hedging everything.

Step one

Levels
Search the capabilities

How to read the levels, and how to cite a capability
LevelKeywordWhat it requires
RequiredMUST The system does this before it serves the public. A deployment that misses a required capability has a failure mode the field has already documented.
CriticalMUST A required capability where failure harms a person directly. These are the items to test hardest and to review by hand.
ExtendedSHOULD There may be good reasons to skip one of these, and the reasons should be written down. Most are what a mature deployment adds after its first year.
AspirationalMAY Few systems do these yet. They mark the direction of travel, and they are the capabilities most likely to change in the next version.

The keyword column follows BCP 14, the convention used in internet specifications, where the capitalized words carry the requirement level and ordinary prose does not. The keywords appear in the CSV and JSON downloads so that they can be pasted into a solicitation or a grant condition.

Identifiers

Each capability carries an identifier in the form BHQ-<task area>.<capability>, for example BHQ-4.2. Identifiers do not move or get reused inside a version. A retired capability keeps its number and is marked retired rather than deleted. Because numbers can change between versions, cite an item with the version attached: BHQ-4.2 (brief-help-qa v1.0). This follows the identifier rule used by the OWASP Application Security Verification Standard.

Scopes

This agenda has no scopes. Question answering does not yet divide into shapes of deployment that the field has agreed on, so every capability applies to every deployment. Tell us if your deployment needs a narrower selection, and which capabilities you would put in it.

Formats

This page is the reference view. The CSV and JSON downloads carry whatever the filters are currently showing, so a scope and level selection becomes a working checklist. The print view produces a PDF of the same selection. Every download records the version, the status, and the scope it came from.

Task areas and capability counts

Jump to a task area

    Where this comes from, and how to comment

    The capability list draws on published user and expert research, three public audit reports of live legal help systems, and practices that deploying teams have described in conference presentations and technical webinars. The evidence base is set out in full further down this page.

    This is a cross-cutting workflow rather than one tied to a single legal issue. The companion conformance standard, which sets the thresholds and the review gates, is in development.

    Draft for review. This is a working draft. The capability list and the levels are both open to challenge, and a comment that removes a capability is as useful as one that adds one. Send corrections and additions to legaldesignlab@law.stanford.edu.

    Evidence base

    Where the evidence comes from

    User and expert research

    The quality and safety criteria behind task areas two, three, seven, and eight come from Hagan (2024), Measuring What Matters: Developing Human-Centered Legal Question and Answer Quality Standards through Multi-Stakeholder Research, presented at JURIX 2024. The empirical base is 46 community member interviews and 21 expert evaluations.

    Five findings from that research diverge from how legal AI answers are usually judged, and they are the reason several capabilities here are worded the way they are.

    1. Actionability matters more to users than precision. A vague but accurate answer helps less than a specific answer that is slightly imprecise.
    2. Some accurate information is dangerous without context. Stating a right without describing the risk of exercising it is a safety failure rather than a neutral act.
    3. Citations do not raise perceived quality for public users. They are neutral, and citation density is not a proxy for a good answer.
    4. Over-caution is a failure mode. Repeated advice to consult a lawyer, where the person could act, makes the answer worse.
    5. Procedural errors are the most harmful kind. Wrong deadlines, wrong courts, wrong forms, and wrong filing steps do more damage than admitted uncertainty.

    Published audits of live systems

    Three public audit reports from a university law and technology center inform task areas five, seven, and nine. Two examined live public-facing systems, and one examined the cost of scoping a new build for a statewide legal help site. Their published findings include rules from the wrong jurisdiction presented as local law, absent confidentiality disclosure, indefinite session retention, a defect that allowed access to another user's data, high abandonment from an embedded entry point, and a distribution of answer quality scored on a five-point scale running from helpful and actionable down to harmful.

    Publicly documented production deployments

    Several capabilities in task area nine reflect practices that deploying teams have described publicly in conference presentations and technical webinars. Those practices include maintaining content in one place and syncing rather than duplicating it, holding a fixed control question set and re-running it after prompt changes, reviewing thousands of real conversations by lawyers, correcting failures by changing content or the system prompt rather than the model, moving to automated risk flagging as volume grows, and scrubbing personal information before logging.

    The rest of the package

    Other documents for this workflow

    A functional agenda is one of the documents the Commons produces for each workflow. These are the others for brief help question answering.