A consumer bankruptcy firm can use AI to extract figures from pay stubs and bank statements into a source-linked worksheet, draft schedule fields and chase missing documents. The lawyer must verify every amount, creditor, asset, income period and legal classification against the records and the client's answers. Nothing should be signed or filed until the lawyer and client review the completed forms.
Document extraction is the narrow useful job. A model can misread a pay period, mistake a transfer for income or treat a bank balance as the full value of an asset. It cannot decide exemptions, ownership interests, chapter choice or whether a creditor was omitted. Those decisions require legal analysis and a conversation with the client, supported by a traceable record.
Start intake with a document map, not a blank upload box
Set out the document categories the firm actually needs for its chosen matter type and current forms: identity and household details, income records, bank and other account statements, asset records, creditor notices, secured-loan statements, regular expense evidence and relevant prior filings. The lawyer decides the period and exact list for each case. A single “upload everything” request creates a disorganised pile and often misses the one record needed to resolve a schedule item. Group requests by question, with a clear status for received, readable, reviewed, missing and not applicable.
Use a secure client portal or approved document route. Pay and bank records contain sensitive information, sometimes about household members and third parties. Review the AI vendor's retention, access and training terms before uploading them. ABA Model Rule 1.6 bars a lawyer from revealing information relating to the representation without the client's informed consent, and that duty does not disappear because the firm uses an extraction service. Keep matter access limited to the team and avoid pasting raw statements into a general chat without the firm's confidentiality review.
An illustrative checklist entry might say: “Income: two pay records received for the current job; earlier job mentioned on intake form; earlier records missing; reviewer [name]; request sent 9 October.” The status should be “missing”, not “complete”, even though the upload folder contains pay stubs. A second entry might say: “Bank: statements for account ending 4421 received; monthly gap between June and August; client asked for July.” Specific gaps are easier for a client to fix than a broad request for more paperwork.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowLabel each file before extracting a number
Give files matter ID, document type, owner, date range and version. Check page count and readability. A photograph of the first page of a bank statement is not a full statement. A pay stub may show net pay prominently while the schedule or related calculation requires a different figure. Optical character recognition can confuse a decimal point or transpose a digit; the reviewer must see the original image beside extracted values. Flag handwritten changes, cropped pages and conflicting versions instead of choosing the neatest number.
Build a source ledger with columns for field, extracted value, document, page, date range, confidence, reviewer and final value. For each pay record, preserve gross pay, deductions and net pay separately. For each bank statement, preserve opening balance, closing balance, date range and account identifier. When AI cannot read a field, it should return “unclear” and the page reference, not a guessed figure. An extraction tool that produces confident numbers without source links is a poor fit for this task.
Picture a blurry pay stub where gross pay reads as either $2,180 or $2,780. A model may choose the higher number because the monthly total then looks plausible. The correct action is to mark the field unclear, ask for a clearer copy or verify it against another source. Entering a guessed figure into a schedule and later fixing it during signature review wastes work and risks carrying the error into several derived totals.
Reconcile income across pay records and deposits
Income extraction needs a period table, not one average copied from the most recent stub. Record employer, pay date, pay period, gross amount, deductions and net amount for each record. Look for changes in hours, bonuses, overtime or a second job. Compare deposits to the bank account, but do not assume every deposit is wage income. Transfers between the client's own accounts, refunds, loans and reimbursements can look similar in a transaction export. Ask the client what an unexplained deposit represents and retain the answer.
An illustrative sequence has four net deposits of $1,640, $1,590, $1,715 and $1,625, plus a $600 transfer from another account owned by the client. The four deposits total $6,570; the transfer should be investigated and linked to its originating account, not added as wages by default. The pay stubs supply the gross and deduction details. A simple sum of bank inflows would produce $7,170 and overstate this narrow wage picture. The lawyer decides how the facts map to the applicable forms and time periods.
Keep household income questions separate from document extraction. A missing household member's contribution, a new job or an irregular payment may not appear in the supplied statements. An intake interview should ask about these possibilities. If records disagree, do not silently select the number that seems most favourable. Create a discrepancy note: what differs, by how much, what source is missing and who will resolve it. The financial tracing tutorial covers more complex movement of funds; routine schedule preparation still needs a basic transfer check.
Build asset and debt lists from multiple sources
Statements can reveal an account, but not every asset. Ask structured questions about vehicles, property, household goods, business interests, claims, insurance interests and other categories reflected in the current official forms. Compare client answers with documents and prior intake notes. Do not have AI infer an item's value from a merchant transaction or estimate a vehicle from an old loan payment. The lawyer selects the appropriate schedule field and determines how an interest should be described and valued under the applicable rules.
For debts, combine creditor letters, account statements, collection notices, credit information the firm is authorised to use and the client's list. De-duplicate carefully. Two notices from the same debt collector may concern one obligation, while a card issuer and a collector may refer to the same balance at different stages. Keep original account numbers restricted; use a partial identifier in the working table. Record creditor, address source, amount date, secured or unsecured question, dispute status and reviewer. A model may suggest duplicates, but the lawyer decides the final creditor list.
For example, a client uploads a card statement showing $4,800 due and a later collection letter asking for $5,050. AI might create two debts of $9,850 or simply replace the earlier amount without explaining the change. The reviewer should link the records, confirm whether they describe one obligation and ask about interest or fees. The final amount and creditor treatment are legal and factual decisions. Keep both documents attached to the row so the lawyer can see why the draft changed.
Turn verified rows into draft schedule fields
The official forms index lists the individual-debtor schedules as A/B (property), C (property claimed as exempt), D (creditors who hold claims secured by property), E/F (creditors who have unsecured claims), G (executory contracts and unexpired leases), H (codebtors), I (income) and J (expenses), followed by the Declaration About an Individual Debtor's Schedules. Use the current forms and instructions for the matter rather than a model's memory of form names. A controlled template can map verified ledger fields into draft form fields and produce a list of unmapped items. It should never silently drop a category because no source record was uploaded.
Ask AI to create a draft mapping report before it creates prose. For each proposed field, show the source and the reason for the match. A useful mapping line (illustrative) reads: “Account ending 4421: closing balance $920 on statement dated 30 September; candidate property entry; lawyer to verify relevant valuation date and ownership.” That is a useful draft. “Cash assets $920, complete” overstates what the statement proves. It ignores other accounts, cash and the legally relevant date.
Do not let the extraction tool choose exemptions or legal categories. A secured claim requires more than a lender name in a bank payment; an expense category may need context beyond a recurring debit. The lawyer should review the form's wording and the client's circumstances. The two-lawyer document workflow offers a broader example of separating source collection from legal judgement, but bankruptcy schedules have a distinct completeness burden and signature step.
Chase one missing item at a time
Use an exception queue to send short, specific requests. Each request needs the document type, period, safe upload route, owner and reminder date. Avoid ordinary email attachments when the firm has a portal. Do not tell a client the filing is ready while important evidence remains missing. A reminder should explain why the request matters in plain language without stating legal conclusions or exposing account details in the subject line.
An illustrative message reads: “Hello [first name], we have your statements for the account ending 4421 for June and August. Please upload the complete July statement through the secure portal, including every page. If you cannot access it, reply to tell us and the team will discuss another way to verify the information. We will review the draft schedules after the missing period is resolved.” That is more actionable than “Please send all outstanding documents immediately.”
Make partial uploads a separate status. If the client sends only a transaction screenshot, staff should say which full statement is still missing. If a pay record is illegible, request a clearer version and retain the original for traceability. Do not allow the model to mark an item received merely because a file with a similar name exists. Use a staff-confirmed status change before a reminder stops. The law-firm status update tutorial helps with general client communication, while this checklist should stay tied to exact evidence gaps.
Run a matter through totals and discrepancy checks
Consider an illustrative consumer matter with 14 uploaded files: four pay stubs, four bank statements, three creditor letters, a vehicle loan statement, a lease and an intake questionnaire. Extraction produces 68 candidate fields. Staff verify 52, mark nine unclear and identify seven missing source items. They spend 45 minutes labelling files, 70 minutes comparing extracted values with pages and 25 minutes contacting the client. The lawyer spends 60 minutes on classifications and the draft schedules. These are example effort figures, not a claim of typical savings.
In this illustration, the bank export shows deposits totalling $7,170, while verified net pay is $6,570. The $600 difference is a transfer from the client's other account. The debt table initially totals $28,400, but a duplicated collector letter adds $5,050, so the reviewed total is $23,350 for those listed debts. The asset worksheet has a $920 account balance from a statement but still lacks the client's confirmation of the relevant date and any other accounts. Each adjustment is documented in the source ledger before the form draft is refreshed.
The seven missing items do not all have the same importance. One missing page may clarify a small bank fee; another may establish an entire account or income source. The lawyer prioritises requests according to the case and deadlines, not simply by the number of empty checklist cells. If a critical record cannot be obtained, record what alternative evidence was reviewed and the lawyer's decision. AI can display unresolved questions; it cannot deem the risk acceptable.
Conduct a client-and-lawyer read-through before signature
Print or display the final draft schedules beside the source ledger. The lawyer should ask the client to confirm completeness in ordinary language: are there any other accounts, assets, creditors, household income sources or recent changes? Compare totals across forms and worksheets. Check names, addresses, account suffixes, date ranges and whether the same debt or asset appears twice. The client should see the actual final version to be signed, not an earlier extraction summary. The official Declaration About an Individual Debtor's Schedules makes the signature stage consequential.
Have the lawyer sign off on a version number and date. If a client later sends another statement, reopen the affected rows, revise totals and repeat the final comparison. A “reviewed” tick on a prior version does not carry over automatically. Before filing, check that every field in the official forms can be traced to a source, client statement or documented lawyer judgement. Where a value is estimated or disputed, the lawyer must decide how to state it under current law and procedure.
Keep a correction log after filing as well. If the team discovers an omission, the lawyer decides whether an amendment or other action is required. Do not let an AI assistant draft a corrective filing from a short chat message without reviewing the original forms and evidence. For a small firm, the most valuable measurement is the percentage of candidate fields corrected before signature, the number of missing documents discovered at final review and the time spent resolving each discrepancy. Those measures show whether the intake process is becoming more reliable.
Test the process with awkward documents first
A pilot should include more than clean digital PDFs. Select one photographed pay stub, a bank statement with a missing page, a creditor letter with a balance different from the latest statement, and an account transfer that appears in two banks. Use fictional or properly de-identified material for the first test. The operator should know the expected answer and score the tool on whether it flags uncertainty. An extraction that stops for an unreadable field is safer than one that produces a polished but guessed worksheet.
Build an error table with columns for document, expected value, extracted value, error type and correction time. In a 20-field test (illustrative), perhaps 16 values are correct, two are blank with a clear warning and two are wrong. The two wrong values deserve more attention than the four unresolved ones, because they can pass into forms unnoticed. Check whether the mistakes came from image quality, a field label, duplicate pages or a prompt that asked for an average when the form needed a dated amount. Fix the process, then test a different set of documents.
Give staff a way to stop a draft from moving forward. Mark a matter “hold for lawyer” when an asset is disputed, a creditor may be duplicated, a document is missing across the relevant period or the client changes an answer during the final read-through. The workflow should not auto-clear that hold because a new file arrives. A person needs to inspect the file and update the ledger. This is a practical safeguard against a reminder system mistaking “uploaded” for “resolved”.
Check the firm's own time accounting in the pilot. Measure file labelling, extraction review, client follow-up, lawyer classification and final read-through separately. If the first three stages get faster but the lawyer spends longer resolving opaque AI fields, the process has shifted work rather than improved it. A useful target is a higher share of schedule fields with a source reference and fewer surprises at signature, not a fixed number of minutes per case. Keep case complexity in the comparison; a simple matter and a disputed-property matter should not be averaged without context.
Retain the trace without retaining unnecessary copies
Decide where original records, extracted tables, prompts and final forms live under the firm's retention policy. The matter file needs enough information to reconstruct why a figure was used, but staff should not create uncontrolled duplicate archives in personal drives or chat histories. When a third-party extraction tool is used, confirm its access controls and deletion options and document who can retrieve the data. If the terms are unclear, use a local or already-approved route until the firm has resolved that question.
Here is how copies multiply, as an illustration: an assistant downloads four bank statements, uploads them to a separate AI account and then copies the resulting table into the matter system. The table may be useful, but the statements now exist in three places. A better controlled route defines one approved upload location, links each extracted row to the source there, and removes temporary working copies according to policy. The lawyer should be able to audit a figure without asking staff to retrieve it from a private chat.
Questions before an AI-assisted schedule is filed
Can bank statements alone establish every debt and asset?
No. Statements reveal transactions and balances for the accounts supplied, but may omit cash, property, secured balances, disputed claims and accounts held elsewhere. Ask the client about each category, compare creditor notices and other records, and mark missing evidence. A lawyer decides what belongs on each schedule.
May an AI system populate the official forms automatically?
A controlled system may transfer verified fields into a draft if staff can trace every value to a source and review the result. The lawyer must check completeness, legal classification, exemptions, totals and the final version with the client before signature or filing. Do not allow extraction confidence to stand in for legal approval.
How should the firm chase a missing document without revealing sensitive details?
Send a short, reviewed request through the firm's approved secure channel, naming the document and relevant period without repeating private financial facts in ordinary email. Track the request, due date and response. If the client sends a partial or unreadable file, explain exactly what is still needed rather than marking the checklist complete.
Further reads
- Best AI for Solo Lawyers: A Stack That Fits One Person — Choose a small-firm tool stack with realistic controls.
- AI App for Lawyers: Score Accuracy, Privacy and Cost in a Short Trial — Score a pilot on accuracy, privacy and review effort.
- Can Lawyers Use AI Note Takers in Client Meetings? — Capture intake conversations with consent and review.
- Free ChatGPT for Lawyers: Where the Free Plan Stops Being Safe — Understand the limits of a free assistant for client work.
- AI and Attorney Fees: Can You Bill Clients for AI-Assisted Work? — Record review and correction time accurately.
- AI Loan Processing: Catch Missing Documents Before Underwriting — A small-brokerage routine: classify each file's documents, test them against a loan-type checklist, run stale-date formulas and have a processor sign off.
- AI for Family Law Attorneys: Summarise Financial Disclosure Safely — A step-by-step method for turning bank statements into a categorised schedule with AI, and the spreadsheet checks that prove the totals.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: the courts' official Bankruptcy Forms index (B 106A/B to B 106J and the B 106 Declaration About an Individual Debtor's Schedules); ABA Model Rule 1.6 and its comment on confidentiality. Confirm current forms, local filing requirements and client facts with the responsible lawyer.