What AI can and cannot do in bankruptcy petition preparation
Short answer
AI reads documents and fills fields. It does not prepare a bankruptcy petition. The current tools extract data from intake packets and supporting documents, and the useful ones tell you which fields to check. The hard part is knowing when an extraction is wrong, which is why field-level agreement across independent reads and a verified read-back of the exported file matter more than which model is underneath. The attorney still signs.
Every vendor in consumer bankruptcy now has an AI paragraph. Most of them describe the same underlying capability, which is a vision-language model reading a document and returning structured values. That capability is real and it is genuinely useful. It is also the easy half of the problem.
The hard half is knowing when the machine got it wrong. A model that reads a handwritten balance as $1,400 instead of $1,900 returns both answers with identical confidence in tone. What distinguishes tools worth buying from tools worth avoiding is not the model. It is whether the product can tell you, field by field, which values to look at again.
Can AI prepare a bankruptcy petition?
No. AI can read the documents a petition is built from and populate fields, and that is the whole of what it does today. Preparing a petition means choosing a chapter, valuing property, selecting and applying exemptions, deciding how each claim is treated, judging whether a transfer needs disclosing and confirming that what the debtor said is complete. Those are judgment calls made by a person who is accountable for them.
The distinction is not pedantry, because it decides what you should be evaluating. If AI is doing extraction, then the questions that matter are what it reads, what it writes, whether the output actually imports, and how it flags uncertainty. Model names, parameter counts and the word “agentic” tell you nothing about any of those.
What parts of petition prep can AI actually do today?
Four things, reliably. It reads handwritten and typed intake packets and returns structured fields. It reads supporting documents such as pay stubs and bank statements. It maps extracted values onto petition data fields. And it writes an importable file for petition software. Everything past that point is either a rules engine, which bankruptcy software has had for thirty years, or a lawyer.
| Task | AI does this today | Notes |
|---|---|---|
| Reading a completed handwritten intake packet | Yes | Vision-language models read cursive far better than the character-matching OCR engines that preceded them. Accuracy still varies by handwriting, scan quality and which field is asked about |
| Reading pay stubs and bank statements | Yes | The most commoditised capability in the category. Several vendors ship it |
| Mapping values to petition fields | Yes | Deterministic once the values are extracted. The failure mode is a wrong value mapped confidently, not a mapping error |
| Producing an importable file for petition software | Yes, where a vendor has built it | The .BCB format used by Best Case has no published specification, which is why almost nobody writes one |
| Deciding a chapter | No | A legal judgment about the client’s circumstances, not a calculation over the intake form |
| Selecting and applying exemptions | No | State-specific, strategy-dependent and the single most consequential judgment in a consumer case |
| Judging whether a disclosure is complete | No | The information that is missing is by definition not on the page the model read |
| Signing and filing | No | The attorney signs, and the debtor declares the schedules true under penalty of perjury |
AI bankruptcy tools compared
Six products are worth naming, and they split cleanly into tools that replace your petition software and tools that feed it. Everything below was checked against each vendor’s own published pages on 21 August 2026, and claims that only appear in a vendor’s marketing are labelled as that vendor’s claim.
| Tool | What it reads | What it writes | Replaces or feeds your petition software | Verification offered | Published price |
|---|---|---|---|---|---|
| Casewell | Handwritten or typed PDF intake packets, read five times independently | A Chapter 7 .BCB file for Best Case covering Schedules A/B, D, E/F, G, I and J. The generated file is read back before delivery to confirm it imports | Feeds. Best Case stays exactly where it is, no migration | Field-level agreement across the five reads, staff confirmation of every field the reads disagreed on, and a read-back of the exported file | Free during the current beta |
| Glade AI | Client documents, described by the vendor as document intelligence | Assembled petitions, and filing across districts | Replaces. It is a filing platform, so adopting it means migrating off Best Case | None published. The site states outcome claims rather than a verification mechanism | Not published. No pricing page |
| Best Case by Stretto | Credit-bureau data through CIN Legal Data Services, which is a structured feed rather than document reading | The creditor schedules, D and E/F | It is the petition software | Not applicable | Cloud $99 per user per month. Desktop from $1,500 in year one |
| NextChapter by Clio | Nothing of the client’s documents. Its “Notices A.I.” reads incoming court notices | Routed notices and scheduled hearings | It is the petition software | Not applicable | $159 to $1,999 a year by tier, plus $500 a year for the MyChapter client portal |
| Jubilee Pro by LegalPRO Systems | Nothing. The client portal accepts uploads, and no AI document reading is described anywhere on the site | Nothing from documents | It is the petition software | Not applicable | $95 per case pay as you go, or $995 to $4,995 a year by filing volume |
| BK Questionnaire by BK Assistant | Uploaded pay stubs, using what the vendor describes as OCR and AI | Means test and Schedule I fields inside its own online questionnaire | Feeds. CSV export, plus a direct Jubilee import | None published | $39.99 to $129.99 a month, with the top tier at $159 for joint debtors |
| CounselPro | Bank statements | Categorised, source-linked transaction data and a forensic report | Feeds. It is an analysis tool, not a filing platform | Source-linked output, so every figure traces back to a statement line | $499 a matter, per the vendor’s own site |
Two things stand out from that table. The first is that only two products read the intake packet itself. Everything else either reads a supporting document or reads nothing at all. The second is that almost nobody publishes a verification mechanism, which is the column a buyer should care about most and the one vendors find hardest to fill in.
Does Best Case already have AI built in?
Not according to Stretto. As of August 2026, neither bestcase.com nor Stretto’s Best Case page mentions AI, machine learning, OCR or automatic extraction from client documents anywhere. What both pages describe is a secure client-intake portal, credit-bureau data import, OneTouch e-filing, court notices through PACER and updated forms. The only quantified saving Stretto publishes is “60 minutes of data entry per case” from the credit-bureau import, which fills the creditor schedules only.
Stretto does have an announced generative AI product, Stretto Conductor, introduced in January 2025. It is built for corporate Chapter 11 creditor communications and precedent research, and the announcement makes no reference to Best Case or to consumer petition preparation. It is a different product for a different practice.
Practitioners in bankruptcy forums have described a paid document-upload feature in Best Case, one calling it too inconsistent to justify the extra cost and reporting that it handled pay stubs more than anything else. Those are unverified secondhand accounts and we have not been able to confirm them against any Stretto documentation, so treat them as rumour rather than as a feature list. If your firm is being sold an AI upload add-on, ask Stretto directly what it reads, what it fills and what it costs, and ask for it in writing.
How do you tell whether an AI read is right?
You cannot tell from a single read, and no confidence score fixes that. A model asked once for the balance on a handwritten line returns one number and no signal about whether the ink was ambiguous. The practical answer is to read the same page several times independently and compare the results field by field. Fields where independent reads agree are very likely correct. Fields where they diverge were genuinely ambiguous, and those are exactly the fields a person should look at.
The comparison has to be per field rather than per document. A packet with ninety-five clean fields and five contested ones is a five-minute review job. The same packet scored as “95% accurate” is an unreviewable document, because the number hides which five fields decide whether the petition is right. Casewell runs five reads by default across different vision models and different prompt framings, votes per field, and puts the disagreements in front of staff with the source page beside them. The mechanism is set out in OCR for handwritten legal intake forms.
The second check is downstream, and it is the one almost nobody does. Extracting a value correctly is worthless if the file you export will not import. Casewell writes the .BCB with the genuine database engine and then reads the generated file back, confirming the schedule codes and creditor rows are there before the firm downloads it. A read-back is a factual test with a yes or no answer, which is a different category of assurance from a percentage on a marketing page.
- Ask what happens when the tool is unsure. A product that never says “check this field” is not more accurate, it is less honest.
- Ask whether agreement is measured per field or per document. Only the first is actionable.
- Ask whether the exported file is tested before delivery. An import that fails at your desk is your problem, not the vendor’s.
- Ask for the failure list from a real case. Run one of your own messy packets during the trial, not the vendor’s clean sample.
- Treat any single accuracy percentage as marketing. Accuracy moves with handwriting, scan quality, form design and the field in question, and one averaged number across a 34-page packet conceals all of that.
What should never be automated in a bankruptcy petition?
Anything that is a judgment rather than a transcription. Exemption selection on Schedule C, chapter choice, property valuation, whether a prepetition transfer is disclosable, how a secured claim is treated, and whether the debtor’s answers are complete are all decisions that require knowing the client and the district. A tool that quietly makes any of them has not saved you work, it has moved a legal decision into a black box and left your name on it.
Completeness deserves its own line, because it is the failure mode automation is structurally worst at. In the Department of Justice and US Trustee debtor audit program for FY2024, 110 of 539 audited cases, or 20%, had at least one material misstatement, with income-related findings in 66% of flagged cases and asset or transfer findings in 43%. No document reader catches an omission, because the omitted fact is not on the page. That is a client conversation, and it is why the written questionnaire reviewed in an interview remains the recommended practice.
This is also why Casewell’s scope is deliberately narrow and stated rather than implied. It covers Chapter 7 in New York districts, across Schedules A/B, D, E/F, G, I and J. It does not do Schedule C exemptions, does not run the means test and does not generate a creditor matrix. Those are places where a plausible wrong answer is worse than a blank, so it leaves them alone.
Does using AI change who is responsible for the petition?
No. Under Federal Rule of Bankruptcy Procedure 9011(b), presenting a petition to the court certifies that, to the best of the attorney’s knowledge and belief “formed after an inquiry reasonable under the circumstances,” the factual contentions “have evidentiary support.” The rule attaches to the person who presents the document. It says nothing about how the data got into the fields, and no vendor can absorb any part of it.
Schedules carry their own exposure. Rule 9011(a) expressly excludes a list, schedule or statement from the attorney-signature requirement, and the debtor declares those true under penalty of perjury instead. That has not protected counsel in practice. In In re Parikh, 508 B.R. 572 (Bankr. E.D.N.Y. 2014), the court imposed Rule 9011 sanctions on debtor’s counsel over schedule omissions and a misstated mortgage payment, and the rule’s twenty-one-day safe harbor does not apply to the filing of a petition.
The practical reading is straightforward. Automation can change how many minutes the inquiry takes. It cannot change the fact that an inquiry is required, or who is answerable when it was not reasonable. Any tool that reduces your review to rubber-stamping has increased your risk while appearing to reduce your workload, which is the worst possible trade against a Chapter 13 fee the court has already capped.
A workflow that respects that looks the same as it always did, with less typing in the middle. Staff review the extracted fields against the source packet, the attorney reviews the imported case in Best Case exactly as they would after manual entry, and the firm files. Casewell’s intake automation workflow is built around the review step rather than around removing it, and the full run-through is on how it works.
Can AI file a bankruptcy petition for me?
No. AI extracts data from documents and fills fields. Choosing a chapter, applying exemptions, valuing property and judging whether a disclosure is complete are legal judgments, and under Rule 9011(b) the attorney presenting the petition certifies its factual contentions after a reasonable inquiry regardless of what software produced the draft.
Is AI accurate enough for bankruptcy schedules?
Accuracy varies by handwriting, scan quality, form design and which field is being read, so a single averaged percentage is not a useful answer. The workable approach is to read each page several times independently and compare field by field, so the tool can tell you which specific fields to check rather than giving you one number for the whole packet.
Does Best Case have AI document reading?
Stretto documents none. As of August 2026 neither bestcase.com nor Stretto’s Best Case page mentions AI, machine learning or OCR of client documents. Stretto’s announced generative AI product, Stretto Conductor, is built for corporate Chapter 11 communications and research, not consumer petition preparation.
What is the difference between Glade AI and Casewell?
Glade AI is a filing platform, so adopting it means migrating off your current petition software. Casewell feeds Best Case a .BCB file and leaves Best Case in place, so the firm keeps its case history and its filing workflow. One replaces, the other supplies.
Will AI make bankruptcy paralegals unnecessary?
No. It removes transcription, which is the least valuable part of the role. Chasing the missing pay stub, spotting the contradiction between Schedule I and the pay records, and knowing that a client who says they own nothing has a car in a relative’s name are all judgment work that no document reader touches.
Do courts or trustees restrict AI use in petition preparation?
No rule addresses the tooling directly. What the rules address is the certification made when the document is presented and the debtor’s declaration under penalty of perjury on the schedules. Some courts have issued standing orders about generative AI in briefs, which is a separate question from data extraction at intake, so check your own district.
What should I ask an AI bankruptcy vendor during a demo?
Ask what it reads, what it writes, whether the exported file is tested before you receive it, and what it does when it is unsure. Then run one of your own difficult packets rather than the vendor’s sample, and look at the list of fields it flagged rather than the fields it filled.
Best Case and Stretto are trademarks of their respective owners. Casewell is an independent product and is not affiliated with, sponsored by, or endorsed by Best Case, Stretto, or any of their affiliates. References to Best Case and Stretto describe compatibility only and are nominative (descriptive) use.
This page is general information for law-firm staff, not legal advice for any particular case.