“MSA” Means Three Different Things in a Mixed Immigration and Family Practice — General Chatbots Pick the Wrong One

Ask a general chatbot what MSA stands for and watch what happens

In a Texas family law office, MSA is a Mediated Settlement Agreement, and it is irrevocable once signed with the statutory language and the signatures in the right places. In a California family law office two states over, MSA is a Marital Settlement Agreement, which is a completely different animal negotiated over months and merged into a judgment. In the immigration half of the same firm, MSA is not a document at all — MSC is a receipt number prefix from the National Benefits Center, and the paralegal who typed it is asking about a case sitting in Missouri.

A general-purpose chatbot picks one, confidently, and drafts around it. The paralegal gets back three pages of clean prose about the wrong instrument, notices on the second read, and rewrites it. Twenty minutes, forty times a week, is a person.

The 2026 development worth an owner's attention is that vertical AI — models tuned on one trade's own documents, vocabulary and edge cases — became a real product category this year, precisely because general models keep guessing at abbreviations that mean one thing in your office and something else everywhere else. This is not about the model being smarter. It is about it having been shown ten thousand real I-130 packets and real FL-150s instead of the internet's summary of them.

The vocabulary a general model gets wrong in this practice, specifically

Immigration first. AOS means adjustment of status here; a general model trained on the wider web will sometimes read it as an agreement of sale. AP means advance parole, not accounts payable. "Current" is not an adjective about time, it is a status on the monthly Visa Bulletin, and the Bulletin has two separate charts — Final Action Dates and Dates for Filing — with USCIS announcing each month which one adjustment applicants may use. A model that does not know there are two charts will give a confident answer about a priority date being current and be wrong half the year.

Then the codes. On Form I-765 the eligibility category is not decoration: (c)(9) is a pending adjustment, (c)(8) a pending asylum application, (a)(5) an asylee, (c)(3)(B) post-completion OPT. Put the wrong one in the box and the work permit is rejected or issued wrong, and the client's employer runs an I-9 against a document that does not match. Receipt prefixes carry meaning too — EAC, WAC, LIN, SRC, IOE, MSC, YSC each tell you where the file sits and therefore who you call. On an I-94 record, "D/S" in the admit-until field is not a date at all.

Family law is worse, because doctrine is state- and county-specific and the shorthand is named after cases. In California, Epstein credits and Watts charges are opposite sides of the same post-separation ledger, Moore/Marsden apportions a house bought before marriage and paid down during it, a Gavron warning tells a supported spouse to become self-supporting, and Smith/Ostler handles bonus income. Feed those to a general model and you get plausible definitions with the parties reversed. In Texas none of it applies; you are in Chapter 154 net resources and a Standard Possession Order. A model that blends California doctrine into a Texas parenting plan produces a document that looks professional and is unusable.

Names, which is the error nobody talks about

The most common data error in an immigration practice is not a legal one. It is a name. Spanish-speaking clients carry two surnames, paternal then maternal, and a general system routinely files a client under the maternal one, which then does not match the passport, the birth certificate or the police clearance. Vietnamese names come family-name-first. Arabic transliterations vary between the passport and the marriage certificate — one has "Mohamed," the other "Muhammad." Many clients from parts of Indonesia and South Asia have a single legal name, and Form G-28 and the online forms want a first and a last.

USCIS handles some of this with "other names used," and experienced paralegals know to fill it. A general model does not know, so it silently picks a last name. Six weeks later there is a Request for Evidence about an identity discrepancy, and the firm eats the response for free because the client will not accept that it was anyone's fault but yours.

flowchart TD
  A["Paralegal types: prepare the MSA summary for the Ramirez matter"] --> B{"Which vocabulary applies to this matter?"}
  B -->|Family law, Texas| C["Mediated Settlement Agreement, Chapter 154 net resources"]
  B -->|Family law, California| D["Marital Settlement Agreement, FL-150 and guideline support inputs"]
  B -->|Immigration| E["Not an MSA: MSC prefix, National Benefits Center matter"]
  C --> F["Draft pulls Standard Possession Order language and county local rules"]
  D --> F
  E --> G["Draft pulls the NBC case inquiry template and the A-number format check"]
  F --> H["Attorney reviews a draft with the trade terms already correct"]
  G --> H

The branch at the top is the entire point. A general model does not ask that question; it assumes. A tuned one treats the matter type as the first fact and the abbreviation as the second.

What tuning actually fixes, and what it does not

Tuning on a trade's own material fixes four categories of error, and it is worth being precise so you can test for them. Abbreviation collisions, like MSA and AOS and PDD. Form-field conventions — which box takes the A-number, that the I-130A goes with the I-130 for a spouse, that a Florida financial affidavit comes in a short and a long form depending on income. Units and formats: priority dates as a month and year, receipt numbers as three letters and ten digits. And the edge cases a trade takes for granted — that the I-751 is filed in the ninety-day window before a conditional card expires, not after, and that asylum carries a one-year filing deadline with narrow exceptions.

What tuning does not fix is currency. A model tuned on last year's practice will cheerfully produce last year's fee amounts and last year's edition of a form, and USCIS rejects filings on stale form editions. The fee schedule and the form editions have to come from a live, approved reference the firm maintains, not from the model's memory. Anyone selling you a tuned model that "knows the fees" is selling you a rejection notice.

The rework arithmetic

The right measure here is not hours saved on the first draft; it is how often the draft comes back for a vocabulary rewrite. Assumptions, clearly labelled as an illustration for a firm with three paralegals: 45 drafted documents a week across both practice areas, and 22 minutes to fix a draft that used the wrong term or the wrong form convention.

The error rates are illustrative — measure your own by tagging thirty drafts over two weeks. But run the test before you buy anything, because the gap between 38 percent and 11 percent is the difference between staff using the tool and staff quietly abandoning it by month two.

Where a person still has to sit

Vocabulary accuracy is not legal accuracy. A draft that uses "Watts charges" correctly can still be wrong about whether Watts charges apply on these facts, and only an attorney decides that.

Local practice is the second gap. Every family court has habits that appear in no rule book — which judge wants the proposed order lodged three days ahead, which department will not hear an ex parte on a Friday, which clerk rejects a filing for the wrong caption spacing in the state e-filing portal. That knowledge lives with your legal assistant and your court runner, and no amount of tuning imports it.

Third, client-facing language. Tuned models are good at the trade's shorthand and often bad at plain speech, which is the opposite of what a client needs. A client being told about a conditional residence window needs seventh-grade sentences in their own language, not accurate jargon. Keep a human between the correct draft and the frightened person reading it.

Frequently asked questions

Do we have to hand over our old files to get this?

Not necessarily, and you should be careful about it. The useful material is your templates, your form packets with client data stripped, your county checklists and your internal memos. That is enough to correct almost all vocabulary errors. Handing over live client files raises confidentiality questions under Rule 1.6 that you do not need to raise to get most of the benefit.

Which practice area should we tune for first?

Whichever one produces more documents per matter. In most mixed practices that is immigration, because a single family-based petition carries a dozen forms and a supporting-document index, while a family law matter carries fewer documents that are individually more bespoke.

How do we know it is actually tuned and not just a general model with a nice logo?

Test it with the collisions. Ask it what an MSA is without giving it context and see whether it asks which practice area. Give it a client named Maria Fernanda Lopez Garcia and see which surname it puts on the G-28. Ask which Visa Bulletin chart applies this month. Three questions, five minutes, and you will know.

Will this replace a paralegal?

It has not in the firms doing it seriously. What it changes is the mix: fewer hours retyping and reformatting, more hours on client contact, evidence-gathering and the follow-up that actually moves cases. That matters in a field where the difference between a strong packet and a weak one is usually who chased the missing document.

The two-week test

Before you sign anything, build a page of your own trade terms — thirty abbreviations, ten form names, five case-name shorthands your attorneys use — and make every vendor run it in front of you. Then tag two weeks of drafts as "clean" or "reworked" and you have a baseline nobody can argue with.

The same vocabulary problem shows up on the phone, where a caller says "I-751" or spells an A-number and the person or system taking the message has to get it exactly right. CallSphere builds AI voice and chat agents for law firm phone lines and web chat that are set up with your practice's own terms and forms, so an after-hours caller's form number, receipt number and language preference land in the message correctly instead of arriving as "something about a green card."