> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ravenna.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Policy and compliance questions

> Stop answering the same legal question every week: write your positions down once, let the agent quote them, and keep a strict boundary on what it will not answer.

export const Checkbox = ({items = []}) => {
  const [checked, setChecked] = useState(Array(items.length).fill(false));
  const toggle = index => {
    setChecked(prev => {
      const copy = [...prev];
      copy[index] = !copy[index];
      return copy;
    });
  };
  return <div className="custom-checkbox-list">
      {items.map((text, index) => <div key={index} className="custom-checkbox-item flex items-center mb-2">
          <input type="checkbox" checked={checked[index]} onChange={() => toggle(index)} className="mr-2" />
          <span>{text}</span>
        </div>)}
    </div>;
};

Roughly half of what reaches a Legal desk is not a review request, it is a question that has been asked before. "Can I sign this?", "do we need a DPA?", "what is our standard payment term?", "can I share this deck under our NDA?".

Those questions are answerable from documented positions, which makes them the highest-return thing a Legal team can hand to an agent. It is also the setup that needs the tightest guardrails.

<Info icon="list-check">
  **Prerequisites:** a private Legal workspace with a channel and an agent, plus the [contract review](/guides/playbooks/legal/contract-review) playbook, which is where the reusable positions come from.
</Info>

## What good looks like

<Checkbox
  items={[
"The recurring questions get answered in seconds, with the actual rule quoted",
"The agent never gives an opinion, it retrieves a position",
"Anything genuinely novel becomes a ticket rather than a confident guess",
"The list of questions the agent could not answer is a work list, not a failure report"
]}
/>

<Steps>
  <Step title="Write down the positions you already hold">
    You have them. They are in past redlines, in emails, and in a lawyer's head. Getting them into text is the whole project.

    Start with the ten you answer most:

    | Position document                | Answers                                                         |
    | -------------------------------- | --------------------------------------------------------------- |
    | Signature authority              | Who can sign what, at what value                                |
    | Standard commercial terms        | Payment terms, notice periods, term length, auto-renewal stance |
    | Liability and indemnity playbook | Your standard cap, what you will move on, what you will not     |
    | NDA guidance                     | When one is needed, whose template, what counts as confidential |
    | Data processing                  | When a DPA is required, subprocessor stance, transfer mechanism |
    | Security questionnaire answers   | The questions sales gets asked every deal                       |
    | Marketing and claims review      | What needs review before publication                            |
    | Employment basics                | Offer letters, contractor versus employee, references           |
    | Vendor onboarding                | What Legal needs before a vendor is approved                    |
    | Regional differences             | Where your standard answer changes by jurisdiction              |

    One topic per document, with an explicit title. A single "Legal FAQ" page with forty headings retrieves worse than ten focused pages.

    <Callout icon="link" color="#6B7280">Learn more in [structuring knowledge for AI](/guides/best-practices/knowledge-for-ai)</Callout>
  </Step>

  <Step title="Scope the folder to the channel">
    Point the Legal knowledge folder at the Legal channel. Two reasons: the agent stops pulling an IT runbook into a contract question, and you avoid the reverse, where a Legal position leaks into an answer in a public IT channel.
  </Step>

  <Step title="Be blunt in the custom prompt">
    This is the one desk where the boundary belongs in the custom prompt rather than in an attached rule, because it is a global guardrail rather than a conditional behavior.

    <Prompt description="You answer questions about this company's documented legal positions. Quote the relevant position and name the document it came from. You do not give legal advice, interpret a contract clause, assess whether something is enforceable, or state a position that is not written down. If the question is not covered by a position document, or it concerns a dispute, a regulator, or a specific counterparty's language, create a ticket for the Legal team and say that you have done so.">
      You answer questions about this company's documented legal positions. Quote the relevant position and name the document it came from. You do not give legal advice, interpret a contract clause, assess whether something is enforceable, or state a position that is not written down. If the question is not covered by a position document, or it concerns a dispute, a regulator, or a specific counterparty's language, create a ticket for the Legal team and say that you have done so.
    </Prompt>

    Everything conditional, "route privacy questions to the DPO", "tag security questionnaire requests", belongs in an attached rule instead. The custom prompt is for environment and global limits.

    <Callout icon="link" color="#6B7280">Learn more in [writing agent rules](/guides/best-practices/writing-agent-rules)</Callout>
  </Step>

  <Step title="Run it in testing mode for a week">
    Turn the agent on in testing mode and read every response before anyone outside Legal sees one. You are looking for one specific failure: a confident answer that is not in a document.

    Read at least twenty conversations. This is not a formality on a Legal desk, it is the acceptance test.

    <Callout icon="link" color="#6B7280">Learn more about [testing and deploying agents](/documentation/automate/agents/deploy-monitor)</Callout>
  </Step>

  <Step title="Treat the gaps as the roadmap">
    The knowledge gaps report tells you which questions people ask that your positions do not cover, ranked by how often they come up. That is a better prioritized list of documents to write than anything a planning session produces.

    Work the top two each week and the desk gets measurably quieter within a month.

    <Callout icon="link" color="#6B7280">Learn more in [closing knowledge gaps](/guides/how-to/close-knowledge-gaps)</Callout>
  </Step>

  <Step title="Feed the answers back from review work">
    The last step of the [contract review](/guides/playbooks/legal/contract-review) checklist is recording the negotiated position. That is where the next round of position documents comes from, which is what stops this desk from being a one-time knowledge-writing exercise.
  </Step>
</Steps>

## Where teams get this wrong

<AccordionGroup>
  <Accordion title="A soft prompt">
    "Try to be careful" is not a boundary. Name the things the agent will not do.
  </Accordion>

  <Accordion title="One giant FAQ page">
    Retrieval works on documents. Split them.
  </Accordion>

  <Accordion title="Skipping testing mode">
    A wrong legal answer costs more than a wrong IT answer, and it is the one place a week of review is obviously worth it.
  </Accordion>

  <Accordion title="Positions that live in a lawyer's head">
    If it is not written down, the agent cannot quote it, and neither can anyone covering a holiday.
  </Accordion>
</AccordionGroup>

## Next

<CardGroup cols={2}>
  <Card title="Writing agent rules" icon="scroll-text" href="/guides/best-practices/writing-agent-rules" horizontal>
    Prompt versus rules, and why it matters.
  </Card>

  <Card title="Close knowledge gaps" icon="search" href="/guides/how-to/close-knowledge-gaps" horizontal>
    Turn unanswered questions into documents.
  </Card>
</CardGroup>


## Related topics

- [Legal](/guides/playbooks/legal/overview.md)
- [Contract review](/guides/playbooks/legal/contract-review.md)
- [User handbook](/guides/user-onboarding/handbook/simple-edition.md)
- [Workspaces](/documentation/platform/workspaces/overview.md)