Family Law Financial Disclosure: AI Tools vs Custom Agent

By Jude Lee · October 6, 2026 · Comparison

Family law attorney and paralegal reviewing stacks of bank statements and tax returns in a small firm conference room

The job nobody wants: turning 1,200 pages of PDFs into one schedule

In a contested matrimonial case the work is mechanical and unglamorous. Opposing counsel produces three years of checking and savings statements, two years of credit card statements, W-2s and 1040s, a pay stub run, a 401(k) summary, and a shoebox of scanned receipts. Someone on your team has to turn that into: a completed state financial affidavit or case information statement, a schedule of assets and debts, a month-by-month income and expense picture, a list of transactions worth asking about (large transfers, unexplained cash withdrawals, payments to unfamiliar payees), and a gap list of what’s still missing for a follow-up request.

This is abstraction work with one hard constraint: the numbers have to add up, and each one has to be traceable. Before you price any solution, get your own baseline. Pull your last three contested disclosures, count the actual pages produced, and record the hours your team logged on extraction and schedule-building. Published page counts and vendor averages are worthless here — the volume in a case with a closely held business looks nothing like the volume in a two-W-2 household, and the only number that should drive a buy-or-build decision is yours.

Three ways to do it, compared

Option one: purpose-built family law software plus structured entry. Tools like Family Law Software and Easysoft’s family law modules are built around state-specific forms, support guideline calculations and settlement projections. Their strength is that the output is correct in form — the right line items, the right calculations, the right jurisdiction. Their historical weakness is the front end: a human keys in the numbers. Several vendors in this category now advertise AI-assisted extraction. Check current product documentation rather than assuming, and put these questions to the vendor in writing before you sign:

Extraction accuracy: how is it measured, on which document types, and does the vendor publish or demo results on faxed and photographed statements rather than clean PDFs?
State form coverage: exactly which forms and versions are supported in your jurisdiction, and how quickly are they updated after a rule or form change?
Page-level citation: does every extracted figure carry a link back to the source document and page, visible to the reviewer, or does the UI show only the final number?
Export format: can you get structured output (CSV, JSON, or a direct write into your DMS), or is the data effectively locked in the vendor’s PDF?
Data handling: where do uploaded client financials go, how long are they retained, and are they used for model training?

Option two: a general AI assistant, used ad hoc. Someone drags PDFs into Claude or ChatGPT and asks for a summary of deposits by month. This works surprisingly well for a single account and badly for twelve. There’s no consistent output format, no audit trail, no write-back to the file, and — the real problem — no enforced link between each figure and the page it came from. It’s also where confidentiality risk creeps in if staff are using consumer accounts outside firm governance. If that’s happening at your firm, the fix is a sanctioned, governed path on firm-controlled accounts, not a ban nobody obeys.

Option three: a custom abstraction agent. An AI assistant given a packaged skill (a reusable, versioned set of instructions for this exact job) plus governed access to your document management system through MCP — the open Model Context Protocol that lets an assistant read and write in your systems under permissions you define. The agent reads the production folder for a matter, extracts every transaction, classifies it against the categories your state’s affidavit actually uses, writes a structured schedule back into the matter, and flags what it couldn’t read.

Form-first software + structured entry

Output is jurisdictionally correct by construction. Calculations are deterministic and defensible. Low adoption risk — paralegals already know the form. But entry time scales with page count unless the vendor’s extraction is genuinely good, and transaction-level analysis (dissipation, hidden income patterns) stays manual.

Custom abstraction agent + MCP

Extraction and classification scale cheaply; every figure can be forced to carry a document ID and page cite. But you own the build, the testing, and the failure modes — and you still need the form software or a template for the final filing math.

These are not mutually exclusive, and in our opinion the sensible small-firm pattern is a hybrid: the agent does extraction and tagging, deterministic software or a spreadsheet does the arithmetic, and a human signs.

Why the model should never do the math

Language models are good at reading a scanned statement and saying “this is a $2,400 transfer to an account ending 7781 on March 3.” They are unreliable at summing 900 of those and reconciling the total against the statement’s own ending balance. Build the workflow so the agent emits structured line items and a separate, ordinary piece of code — or your existing software — does every addition, then checks each account’s extracted total against the printed opening and closing balances. When those don’t reconcile, the agent should stop and report, not guess.

Extraction is the AI’s job. Arithmetic is the computer’s job. Attestation is the lawyer’s job. Collapse any two of those and you’ve built the wrong system.

What a good skill definition actually contains

A “skill” here means packaged instructions that make the assistant do this job identically every time. For financial disclosure, the skill should specify: the exact expense categories your state’s form uses (not generic budgeting categories); a required output schema with fields for date, amount, payee, account, source document, and page; a rule that any illegible or ambiguous entry goes to an exceptions list rather than being inferred; a flagging rule set (transfers over a threshold you set, round-number cash withdrawals, new payees in the disclosure window, payments to entities not previously known); and an explicit instruction never to compute totals.

Write those rules once, test them against two closed matters where you already know the right answer, and version them. The skill is the asset — not the model, which will change under you.

Wiring it into the file, not a chat window

The gap between a clever demo and a usable system is almost always the write-back. Via MCP, the agent can pull the production set straight from your DMS or practice management system and deposit the resulting schedule, exception list and follow-up request draft back onto the matter — with firm permissions and logging applied. We’ve covered the plumbing for standing up a custom MCP server over your own matter data.

Scope that access tightly, because this workflow touches a client’s complete banking history. Read-only on the production folder, write access limited to a single output folder, no standing access to unrelated matters, and a log you can actually audit after the fact.

What automates cleanly, and what doesn’t

It’s worth being precise about which sub-tasks move. Extraction automates well: pulling date, amount, payee and balance off a legible statement is exactly what these models do. Categorization automates adequately, provided you supply your state’s form categories and accept that ambiguous payees will need review. De-duplication across overlapping productions automates well because it’s largely a matching problem on date, amount and account — but only if you tell the agent to do it explicitly. Completeness checking against an expected set of statement months automates well and is often the highest-value output: it generates your follow-up request.

What does not automate: deciding which three transactions actually matter to your theory of the case; judging whether the other side’s claimed lifestyle is credible against the spending record; forming a view on dissipation; and anything requiring inference about intent. Those are judgment calls that rest on facts outside the documents, and a model with no access to the client, the history or the opposing party’s demeanor cannot make them.

Where this sits in the stack firms actually run

From what we see in client work, small firms adopting AI tend to end up with three layers: AI features inside the practice management or document platform they already pay for; a general assistant with firm-level governance for drafting and analysis; and, selectively, one or two custom agents built for jobs that are high-volume and specific to the practice. Financial disclosure abstraction is a strong candidate for that third layer in a family law firm — and a poor one for a firm handling four dissolutions a year, where a paralegal and a good template will beat any build on total cost. That’s an observation, not a survey finding; test it against your own volumes.

Where AI genuinely breaks here

The realistic disadvantages, stated plainly: poor-quality scans and faxed statements degrade extraction badly, and the failure is silent; joint accounts produce duplicate transactions across two productions unless you de-duplicate deliberately; categorization drifts on ambiguous payees (is Venmo to a sibling a loan repayment or a transfer?); and no agent can tell you what’s missing from a production it never received.

There’s also a data question to settle before a single client bank statement reaches any model, including on the sanctioned path. Read the vendor’s actual terms: how long are inputs retained, are they used to train or improve models, and is a zero-retention or enterprise tier available and actually enabled on your account? Confirm whether subprocessors are involved and where processing happens. Then consider whether your engagement letter already contemplates third-party processing of client financial records, or whether you need to update it and obtain informed consent — a question to settle with your malpractice carrier and, where the stakes are real, a qualified professional in your jurisdiction.

Supervision isn’t optional either. The ABA’s Formal Opinion 512 addresses competence, confidentiality and supervision duties when lawyers use generative AI, and ABA Model Rule 5.3 covers responsibilities regarding nonlawyer assistance. Read both against your own state bar’s guidance before you deploy anything that touches a client file.

If you need an opinion that will survive cross-examination on dissipation or business valuation, that’s a forensic accountant, not an agent. The agent’s job is to make the accountant’s engagement shorter and cheaper by handing them clean, cited data instead of a box.

Where is your firm losing billable hours?

Get a free automation audit: we map your intake-to-invoice workflow and show you exactly what's worth automating — before you spend a dollar.

Get a free automation audit