AI Agent for a Bankruptcy Firm: Checklists, Creditor Calls, Deadlines

An AI agent for a bankruptcy firm works best as a drafting and tracking assistant, not a filer. It can keep a client's document checklist current, log each creditor call with its outcome and open follow-up, and surface upcoming deadlines to the responsible attorney. Every draft waits for human review, and every filing stays with a licensed person.

Key takeaways

• Give the agent three jobs: document checklists, creditor call logs and deadline alerts. Keep filing decisions with attorneys.
• Write the approval boundary before the first live file, not after the first error.
• Treat each client file as a record with a status, a named owner and a date of last verification.
• Start with one role and one success criterion that staff can check, such as every open checklist item having an owner and a source.
• Review permissions, call logs and draft quality on a recurring schedule, because client data is sensitive and workflows change.

Why bankruptcy files strain ordinary tools

Bankruptcy files are assembled from many small facts. A client may have pay records, tax returns, account statements, a history of transfers and a list of creditors whose addresses, balances and claims all differ. The information arrives in pieces, by phone, by text and by upload, and each creditor contact can change what the file needs next. A shared spreadsheet records what someone remembered to type. It does not tell you which item is still missing, who promised what on a call, or which date depends on which event.

The cost of an error is high. A missing document can stall a case, and a deadline that lives only in one person's inbox can be missed. That is why the first useful role for an agent in this setting is not drafting a clever letter. It is keeping a reliable record that humans can check quickly. The same lesson applies to intake and deadlines in general, as our guide to AI agents for law firms explains.

Build the document checklist first

A checklist is the simplest place to start because its results can be verified. Begin with a base list for each matter type, written with the attorney who owns that work. Mark each item as required, conditional on a fact the client provided, or waived with a stated reason. Conditional items matter most. A question about a business interest or a recent transfer should add lines to the list only when the answer triggers them, so the client sees a list that fits their situation.

A practical sequence looks like this:

• Write the base checklist for each matter type with the responsible attorney, and version it.
• Tag every item as required, conditional or waived, and record the reason for any waiver.
• Store the source and the person who verified each received document, not only a checked box.
• Let the agent compare uploads against the list and report gaps, duplicates and unreadable files.
• Route every client reminder to a staff queue for approval before it goes out.

Illustrative scenario: A client uploads three months of bank statements when the list asks for a longer period. The agent marks the gap, notes which months are present, and drafts a short request that names the missing statements. A paralegal reviews the draft, edits the wording and sends it. The checklist now shows who asked, when and what came back, which is the record the attorney will need at the next review. Firms that chase documents for other transactions face the same discipline, and our note on AI agents for mortgage brokers covers reminder habits in more detail.

Documents from clients and creditors are untrusted input. A scanned PDF can contain text that reads like an instruction, and an agent that summarizes such files must treat that text as content, not as a command. The OWASP Top 10 for LLM Applications lists prompt injection and insecure output handling as core risks, and anyone configuring the agent should read it before connecting it to case files. For the same reason, client reminders should be built from fields the firm has confirmed, not from free-form text pasted into a prompt.

Log creditor calls so the record survives the call

Creditor calls produce promises, balances that differ from the last statement, a representative's name, a reference number and a next step that someone must remember. Many firms capture this in a brief note that is personal, hard to search and unreadable to anyone else. A useful log entry answers six questions: who called or was called, when, what was said about the account, what was promised or requested, what the client was told, and what follow-up is owed, with an owner and a date.

The agent can draft the entry from a staff note or from a transcript summary, but only where the firm's policy and the law in its jurisdiction allow recording or transcription of that call. Where the source is uncertain, the agent should leave a visible question in the entry rather than fill the gap with a plausible guess. The following decision guide helps staff review each draft:

• If the call changed a balance or a payment arrangement, flag the matter to the attorney the same day.
• If the creditor asked for a document, add it to the checklist with the requester's name and the date of the request.
• If the caller raised a claim or a threat against the client, route the entry to the attorney before anyone replies.
• If a field is blank or contradicts a statement already on file, keep the entry open and ask the staff member who took the call.

This structure turns a call from a private memory into a shared record. It also makes later review easier, because the attorney can see the sequence of contacts without reconstructing it from several inboxes.

Set deadlines the attorney owns

Deadlines in bankruptcy depend on the court, the rules that apply to the case, the type of filing, and events such as notice, service or a scheduled meeting. Some of those inputs come from a judge's order rather than a standard calendar. For that reason, an agent should not calculate a deadline from memory and place it on a calendar. It should read the docket and the firm's rules library, draft a deadline record that cites the rule or order it relied on, and ask an attorney to confirm it.

A workable setup follows five steps:

1. Build a rules library for each court and matter type, reviewed by the attorney who owns that practice. 2. Have the agent draft a deadline record from each docket entry, with the source citation attached. 3. Require attorney confirmation before any deadline enters the official calendar. 4. Set alert intervals the firm chooses, and send each alert to a named owner plus a backup. 5. Review missed or changed deadlines in a recurring meeting, then update the rules library.

Illustrative scenario: A court issues an order that changes a response period. The agent detects the new docket entry, flags that existing deadlines may move, and drafts updated dates with the order cited. The attorney confirms most of them and corrects one that depended on a service date the agent had not seen. The correction goes into the rules library, so the next similar order is handled with the missing input in view. Firms that also manage transcripts will recognize this pattern from our article on court reporting firms.

Draw the line between drafting and filing

The useful boundary is simple. The agent may prepare, summarize and route work, but it does not submit anything to a court, a trustee or a creditor. It has read access to the case file, write access to a drafts folder, and no credentials that would let it file or send on the firm's behalf. Every outbound message or filing draft carries an approval record: who reviewed it, what changed, and when it left the firm. Anything that binds the client, such as a sworn answer or a settlement position, is authored and signed by a licensed person, with the agent's research serving only as support.

This boundary follows the general approach of the NIST AI Risk Management Framework, which treats governance, measurement and human oversight as ongoing functions rather than a one-time setup. The OWASP Authorization Cheat Sheet makes a similar point in software terms: grant each identity only the access it needs, and check that access on every request. Anthropic's guidance on building effective agents also favors simple, well-bounded workflows with human checkpoints over open-ended autonomy. In a law practice, those principles are not optional extras. They are how the firm can explain its process to a supervising attorney, a client or a regulator.

How a managed setup runs this over time

Be Human scopes a first engagement to one useful role, such as the checklist and creditor log for consumer matters. The engagement defines the memory the agent may use, the tools it can call, the points where a human must approve, and a success criterion the firm can observe, such as every open checklist item having a named owner and a source. Managed operations then add monitoring of agent actions, recurring evaluations against sample files, permission reviews whenever staff or connected systems change, incident response, and controlled improvements to templates and rules. The work runs remotely in English or Portuguese, and changes pass through the same approval path the firm uses for its own documents. Take the free Be Human AI diagnostic to find the first task in your firm worth handing to an agent.

Frequently asked questions

Can an AI agent file a bankruptcy petition for a client? No. In this approach the agent prepares internal drafts and summaries, and a licensed attorney reviews, signs and files anything that goes to a court or trustee. The agent holds no filing credentials, so it cannot submit anything on its own, even if a draft contains an error.

Will the agent replace paralegals or intake staff? No. The agent handles repetitive tracking, such as comparing uploads to a checklist or drafting a log entry from a note, so staff can spend more time on client questions that need judgment. Roles shift as the work changes, so the firm should review the division of labor after the first few months rather than assume it.

Do we need to tell clients that an AI agent is involved? That depends on the rules in your jurisdiction, your professional regulator's guidance and your engagement letter. Obligations for automated processing of personal data differ from country to country, so check them with your privacy and professional advisers. This article is operational guidance, not legal advice.

What should we measure during the first month? Choose criteria that staff can observe directly, such as whether every open checklist item has an owner and a source, whether every outbound message has an approval record, and whether every deadline record cites the rule or order it used. Review a sample of files each week with the attorney who owns the matter, and record each correction so the templates and rules library improve in a controlled way.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.