Superintelligent Agents for Criminal Defense: Intake and Court Dates
Set intake limits and court-calendar controls before an AI agent touches criminal defense client data. Covers allowed questions, refusals and date approvals.

An AI agent for a criminal defense office should start with a narrow intake role and a court calendar it can read but cannot change on its own. Before any agent handles client messages or matter data, the office needs written limits on what it may ask, what it must refuse, which dates it may propose, and which attorney or coordinator approves each change. Be Human's managed approach treats those limits as the first deliverable, then keeps them under review.
Key takeaways
- Limit the first agent to intake triage and calendar reads, with no legal advice and no edits to court dates.
- Keep prospective client details separate from matter files, and keep privileged material out of general agent memory.
- Treat every court date as an attorney-confirmed fact: the agent proposes, a named person confirms.
- Write refusal and escalation rules before writing any prompt, and test them with staff playing callers.
- Choose one observable success measure, such as urgent messages reaching a named person inside a window the attorneys set.
Why criminal defense intake needs stricter limits than most front offices
A criminal defense office receives some of the most sensitive information it will ever hold in its first message. A caller may describe a charge, a police encounter, a relative in custody, or a court notice they do not understand. Even before an engagement is signed, that exchange can carry confidentiality duties, and the names inside it can create conflicts that someone must check. An agent configured like a generic contact form will eventually store, summarize or answer something that should have gone to an attorney.
The practical answer is to narrow the agent, not to avoid agents altogether. The first role should triage and route, keep a faithful record of what it received, and hand off anything involving legal judgment, custody status or a deadline. The office sets these boundaries, not the software default. A related pattern appears in family law intake and court dates, where the same separation between intake and legal judgment applies.
Which intake questions may an agent ask?
Start with a short list of permitted items, each approved by the managing attorney:
- Full name and the best callback number, plus the caller's preferred language.
- The county or court where the matter is said to be pending, recorded as the caller describes it.
- Whether the person is in custody, released, or unsure, recorded as reported and not verified.
- Any date the caller says they were told to appear, marked as unverified.
- How the caller found the office and whether they have already spoken with another lawyer.
Some topics stay off the script entirely: the facts of the alleged offense in any open channel, the strength of the case, expected sentences, and whether to speak with police or sign anything. When a caller starts describing facts, the agent should acknowledge the message, say that an attorney will review it, move the conversation to a channel the office controls, and stop collecting. The caller's text is data, not instructions. Inbound messages can contain text designed to steer an assistant, which is why the OWASP prompt injection guidance treats input handling as a design concern rather than a wording problem.
Some situations skip the queue and go to a person at once:
- Custody, detention, or a transfer the caller says is happening today.
- A stated hearing, arraignment or deadline within the next few days.
- Any mention of threats, self-harm, or a minor at risk.
- A message from opposing counsel, a court clerk, or another attorney.
- A complaint about the office or about a previous lawyer.
Clinical intake faces a similar problem with different content. The limits described in behavioral health clinic intake show how routing rules can be written before any scheduling logic is added.
How do you keep court dates as attorney-confirmed facts?
A court calendar is a record of obligations, and a wrong entry costs more than a slow reply. The agent may read the office calendar and any docket information the office is permitted to access. It may draft proposed entries and send reminders only for dates a named person has confirmed. It should not create a hearing from a caller's account, move a date, or tell a client that a hearing was cancelled or continued. Continuances and reset notices come from the court record, and the agent should report them as received, not interpret them.
Define these controls before launch:
- Read access to the calendar and docket sources, with the time of the last successful sync shown in every answer.
- Write access limited to a proposal queue, never the live calendar.
- A named confirming attorney or coordinator for each entry, with the approval time recorded.
- A daily reconciliation in which a person compares proposals against the official court record.
- A stale-data rule: if the sync is older than the office allows, the agent says it cannot confirm the date and routes the question to staff.
The OWASP authorization guidance supports the same principle: give each identity only the access its role needs, and check that access at the point of use. Adjacent court-facing work, such as service logs and status updates for process servers, raises the same question of who confirms what.
Illustrative scenario: A three-attorney defense practice lets an agent answer web inquiries outside office hours. One evening a caller writes that a hearing is 'next Tuesday' and asks whether they must attend. The agent does not confirm the date or advise on attendance. It records the county and the caller's own words, flags the message as time-sensitive, routes it to the on-call coordinator, and adds an unconfirmed proposal to the queue. The next morning the coordinator checks the court record, confirms or corrects the date, and passes the matter to the attorney handling that court.
A step-by-step plan for the first engagement
A first engagement should be deliberately slow. A workable sequence looks like this:
1. List every channel that reaches the office, including web forms, phone transcripts, email, text messages and the client portal. 2. Write the intake limits on one page, and have the managing attorney approve each permitted question and each prohibited topic. 3. Inventory every system the agent could read or write, then start with read-only access to the intake inbox and the calendar. 4. Create the proposal queue and name the person who confirms each calendar entry. 5. Build the escalation route for custody, deadlines and threats, and test it with staff playing callers. 6. Choose one observable success criterion, such as urgent messages reaching a named human inside a window the attorneys set, and check it at every review. 7. Review refusals and misroutes each week, and change one control at a time so the cause of any change stays clear.
Each step produces a document or a permission change that someone in the office can inspect. That is the point of the sequence: the office should be able to explain the agent's authority without reading the code.
What should the office log and review each week?
Every agent decision should leave a record: which message arrived, which rule applied, which record the agent read and when, what it proposed, and who approved it. The OpenTelemetry observability primer describes this kind of traces and logs view for complex systems, and the discipline carries over directly. These records may themselves contain privileged content, so they need the office's retention rules and access limits from the first day.
The NIST AI Risk Management Framework gives a useful vocabulary for the review: govern who holds decision rights, map the context in which the agent works, measure its behavior against the written limits, and manage changes as they come. A weekly review can stay short:
- Urgent messages routed, and how long each took to reach a named person.
- Any answer that gave an unsupported date, a legal opinion, or a prohibited topic, with the rule that should have blocked it.
- Calendar proposals rejected at confirmation, and the reason recorded for each.
- Permission changes made since the last review, including any new channel or system.
How does Be Human's managed approach handle this?
Be Human starts with one intake role and defines its memory, tools, permissions and approval boundaries before anything connects to live client data. Its managed operation then covers monitoring of routing decisions, recurring evaluations against the written refusal rules, permission reviews whenever calendar or inbox access changes, and incident response if a proposal reaches the wrong queue. The office keeps authority over its legal judgments and its policies, while Be Human operates the controls that keep the agent inside them. The work is delivered remotely, in English or Portuguese.
If you are weighing where an agent could start safely, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Can an AI agent answer messages from people facing criminal charges? It can handle the first triage step if the office limits it to collecting basic contact details, recording what the caller reports, and routing urgent matters to a person. It should not give legal advice, assess the case, or tell anyone whether to appear. The attorneys decide what the agent may say and review those rules before launch.
Should an agent add court dates directly to the firm calendar? No. Proposed entries can go to a queue, but a named attorney or coordinator should confirm each one against the official court record before it becomes a live obligation. The agent can send reminders only for dates that have already been confirmed.
What should happen if a message contains an instruction aimed at the agent? Treat it as content to record, not as a command. Inbound text can try to change how an assistant behaves, so the agent's allowed actions should be fixed by the office's design rather than by whatever a message asks. Log the attempt so staff can see it and adjust the controls if needed.
How does the office know the controls are working? Check the logs at a regular review: whether urgent messages reached a named person within the window the attorneys set, whether any answer went beyond its limits, and whether calendar proposals were rejected or corrected. Make changes one at a time and record the reason for each, so the effect of every change can be traced.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


