The line
Paralegal work on the left. Legal work on the right. We only build for the left.
What BatesFlow does
The mechanical work of producing discovery.
The work a paralegal does in a binder room. We just do it faster and with a deterministic record of every decision.
01
Sorting
Group thousands of mixed PDF pages into discrete documents.
02
Classification
Tag each document by type — bank statement, brokerage statement, tax return, deed, etc.
03
Gap analysis
Surface which numbered requests still have nothing responsive.
04
Production mapping
Route each document to the demand request it answers, with human review.
05
Bates stamping
Apply unique sequential numbers across thousands of pages, deterministically.
What BatesFlow doesn't do
Anything that requires a license to practice law.
If a paralegal couldn't do it without a JD signing off, neither can BatesFlow.
01
No legal reasoning
We do not analyze whether a fact pattern supports a legal claim or defense.
02
No citation generation
We never produce case law, statutes, or proposed citations. Full stop.
03
No motion or pleading drafting
We do not write substantive legal text — no motions, briefs, demand letters, or pleadings.
04
No legal advice
No chat interface that answers legal questions. No "should I" recommendations.
05
No professional judgment
Strategy, settlement decisions, and case theory belong to the lawyer.
Why we drew the line here
Lawyers are right to be suspicious of AI.
In 2023, a New York lawyer was sanctioned for filing a motion in Mata v. Avianca that cited six cases ChatGPT had invented. Since then, every state bar that has touched the question has issued a guidance opinion. They aren't unanimous on the details — every citation, every state, but they all converge on the same point: a lawyer who uses AI is responsible for what the AI produces. We took that seriously. So we built a tool that doesn't say anything to a court, doesn't draft anything a lawyer signs, and doesn't claim to know any law. It does paperwork.
Case
Mata v. Avianca, Inc.
22-CV-1461 (S.D.N.Y. June 22, 2023)
"The Court does not believe that lawyers filing fictitious submissions prepared at the direction of ChatGPT, occasioned no harm to the lawful operation of the judicial system."
Rule
ABA Model Rule 1.1, Cmt. 8
Technology competence
"To maintain the requisite knowledge and skill, a lawyer should keep abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology…"
Bar opinion
NY State Bar Op. 1240 (2024)
Generative AI use by attorneys
"A lawyer using a generative AI tool must reasonably understand the tool's limitations and verify its outputs before incorporating them into work product…"
References: California State Bar (2023); Florida Bar Op. 24-1; ABA Model Rules of Professional Conduct, Rule 1.1 Cmt. 8.
For opposing counsel and the bench
Nothing AI-generated reaches a courtroom under your name.
A BatesFlow production looks identical to one a paralegal made: a stamped, indexed XLSX and a Rider DOCX you edited and signed. Opposing counsel never sees AI output, and neither does the judge. The only place the AI's output appears is a classification confidence score in your firm's tenant — for your team's review, before any production is generated.
Our commitments
Five things we will never do.
These aren't roadmap items we haven't gotten to. They are intentional product boundaries. If we ever cross one of these lines, it's a different product with a different name.
01
We will never generate legal text.
No motions, no briefs, no pleadings, no citations. The Rider DOCX template is a structured form that contains your firm's General Objections and numbered Response paragraphs from a Demand. It is not a writing tool.
02
We will never produce a citation.
We do not have access to a corpus of case law and we have no plans to. We do not look up statutes. We do not Shepardize. The lawyer cites; the AI files paper.
03
Every classification surfaces a confidence score and a path to override.
When the model decides a document is a brokerage statement, you see why and you can change it. The model is a draft; the lawyer is the editor.
04
Every classification is logged and auditable.
Every model decision is recorded with its input and output — so a production can be challenged in opposing counsel's office and the firm can show exactly how each call was made.
05
Your client's data trains nothing.
Documents you upload to BatesFlow are processed within your firm's tenant and are not used to train any model — ours, OpenAI's, Anthropic's, or anyone else's.
See the line in practice
The use-case guides show what the work looks like.
Each guide walks through one of the five capabilities above with screenshots, sample output, and the actual scope of what the AI does and does not do.