AI Legal Intake Attorney Review: What AI Drafts and What a Lawyer Approves

You want AI legal intake attorney review to work the way a good paralegal hand-off works: something does the first pass, and a lawyer signs off before anything leaves the building. The objection we hear most is that nobody can say where the draft ends and the approval begins. This post answers that for a small New Jersey firm. It lists each intake step, marks it as draft or human-approved, and shows the override log that records every correction. It is general information about software, not legal advice.

It sits under our guide to law firm intake automation, which covers same-day routing in full.

Why small firms hesitate, and the rule we design around

The hesitation shows up in the data. In the U.S. Census Bureau Business Trends and Outlook Survey, covering December 14, 2025 to May 3, 2026, less than 20% of firms with four or fewer employees reported using AI. The same release put 37% of firms with at least 250 employees in the AI-user column. The survey does not say why small firms hold back, and it is not specific to law firms. We will not guess at the reasons for the whole group.

For law firms the reason is easier to see. A lawyer answers for what goes out under the firm's name. The New Jersey Courts preliminary guidelines (January 2024) on lawyers' use of AI say the core ethical duties are unchanged by AI. So the design question is not whether the tool is clever. It is whether a named person approved the output before it was used.

That gives us one rule for every build: an AI draft is a draft until a person approves it, and the system records who and when. Everything below follows from that rule.

Intake step by step: what is draft and what is approved

The table shows the default setup we would scope for a small firm. Your firm may move a line. Where it does, the change should be a decision by a lawyer, written down, not a default in a settings screen.

Intake steps, their status and the human gate

Step	Who or what produces it	Status when produced	Human gate
Form capture	The visitor, through fixed fields	Raw record, stored as submitted	None. The original text is never overwritten.
Acknowledgement	AI draft, or a fixed template with no AI-written text	Draft, unless it is the pre-approved template	Staff or attorney approves each AI draft before it sends. A lawyer approves the template wording once.
Inquiry summary	AI draft from the form text	Draft, internal only	Reviewer compares it to the original and approves or edits.
Matter-type tag	AI suggestion from the firm's own list	Draft	Reviewer confirms or changes the tag before routing fires.
Urgency flag	AI highlights phrases that may signal a deadline	Draft	Reviewer reads the original. A missing flag proves nothing.
Conflict check	Software matches names against firm records	List of possible matches	A person decides. The software never clears a conflict.
Anything substantive: advice, fee quote, engagement letter	A lawyer	Human only	No AI draft is sent.

Two things are worth noticing. The raw form is always kept, so any summary can be checked against what the person actually wrote. And the last row has no AI in it. We would not build a path where a model answers a legal question for a prospect.

The acknowledgement and the summary

Acknowledgement: say less than you think

An acknowledgement confirms receipt. It should say the firm got the message, when a person will look at it, and that nothing in it is legal advice or creates a client relationship. It should not describe the person's case, predict an outcome or name a deadline.

That narrow job is why many firms choose a fixed template. The lawyer approves the wording once, the system fills in the name, and no AI-written sentence is involved. If you want a warmer, more specific note, an AI draft can work, but a person approves each one before it sends. Which option your firm uses is a call for your attorneys and your own ethics counsel.

Avoid a middle path where AI text sends automatically after a timer. A timer is not a reviewer.

Summary: a reading aid, not a record

The summary exists so the reviewing attorney can triage quickly instead of reading a long narrative. It is internal. It is never sent to the prospect and never replaces the original.

In practice the review screen shows the summary beside the original text, with the facts the summary relied on highlighted. The reviewer can approve, edit or reject it. Summaries are where models drop a fact or smooth over a detail, so a side-by-side view matters more than the draft's polish.

The matter-type tag and why routing waits for it

The tag is the step with the most downstream weight. It decides which attorney sees the inquiry first, which intake questions follow and which deadline rules a person should think about. A wrong tag sends a family-law inquiry to a business-litigation queue.

So the model only picks from a closed list the firm defines, such as the practice areas it actually takes. It cannot invent a category. It also returns "unclear" as a valid answer, and "unclear" goes to a triage queue instead of a guess.

Routing does not fire on an unconfirmed tag. A reviewer confirms it with one click. If nobody confirms within the time your firm sets, the inquiry escalates to a named person. It does not auto-route on the AI's suggestion. This keeps a same-day target honest: the clock runs on a human response, not a model's confidence.

We build this kind of closed-list classification and queue logic as part of AI workflow automation with approvals and exception handling. We also scope the reviewer screens, because a gate people skip is not a gate.

The override log: how you prove the review happened

An override log is a table the system writes to every time a person touches an AI draft. It answers three questions: who reviewed, what changed, and how often the draft was wrong. It is also the thing you can show a partner who asks whether anyone is really checking.

A workable log has these fields:

• Intake ID and timestamp for the review action.
• Step: acknowledgement, summary, tag or urgency flag.
• Reviewer: the signed-in user, never a shared login.
• Action: approved as drafted, edited, rejected or re-tagged.
• AI draft and final text, both stored, so the change is visible.
• Reason: a short pick-list (missed fact, wrong tone, wrong category, other) plus an optional note.
• Template or prompt version that produced the draft.

The log is append-only. Nobody edits a past row, including administrators. A correction is a new row.

What to do with the log once you have it

Read it monthly, step by step. We do not have an industry figure for what a healthy override rate looks like, and we will not make one up. Your own numbers are the useful baseline.

Three patterns are worth acting on. If one step is approved unchanged almost every time, ask whether reviewers are reading it or clicking through. If the tag is often changed, tighten the category list or the descriptions the model sees. If one reviewer's approvals take only a couple of seconds, that is a training conversation, not a software fix.

The log also supports your supervision duties. The sample AI policy template in a New Jersey Courts 2026 notice says lawyers and the firm stay responsible for work product that includes AI output. It also says AI content should not go to a client, opposing counsel or the court without lawyer review. A prospect is not yet a client, so how far that wording reaches at intake is a question for your counsel. We apply it as the conservative default.

What we do not know, and what to ask your own counsel

We build software. We do not advise on professional responsibility, and nothing here replaces your firm's own judgment. A few open questions belong with your attorneys or your state bar's ethics line:

• Whether intake text from a prospect counts as confidential information under your state's rules, and what consent or notice the intake form needs.
• Whether the AI vendor stores or trains on what you send it, and what your contract with that vendor says.
• Whether you must tell prospects or clients that AI helped with the summary.

On the first and third, the American Bar Association summary of Formal Opinion 512 (2024) says the opinion recommends informed consent before client confidences go into a generative AI tool. It adds that boilerplate in an engagement letter is not enough. The same summary notes the opinion is not binding. Read the opinion with your counsel, not just a summary of it.

Vendor and hosting costs for AI services sit outside any build. We scope them separately in the proposal and do not resell them.

Building the gates into your own intake system

You can run this in a spreadsheet and a shared inbox for a few weeks to test the rules. The cost shows up in the approval step: a reviewer has to open three places, and the log lives wherever someone remembers to write it.

That is where a purpose-built system earns its place. Our law firm CRM and workflow software page describes the intake-to-matter chain we build around: intake, conflict review, engagement, matters and client updates in one place. A custom CRM can hold the raw form, the draft, the approval and the log against the same record. The firm owns it under the project agreement.

If you want to start smaller, write down your seven steps from the table, mark each as draft or approved, and name who approves it. That one page is most of the design.

Frequently Asked Questions

Does AI legal intake attorney review mean a lawyer reads every inquiry?

Not necessarily a lawyer for every step. A firm can let trained intake staff approve acknowledgements and tags, with an attorney reviewing summaries, urgency flags and anything substantive. What matters is that a named person approves each AI draft before it is used, and the system logs it. Who qualifies as a reviewer is a decision for your attorneys.

Can AI send the first reply to a new legal inquiry automatically?

It can if the message is a fixed template a lawyer already approved, with no AI-written text. We recommend that any AI-drafted reply wait for a person's approval before sending. The acknowledgement should confirm receipt and say it is not legal advice. Whether to automate even a template is a question for your ethics counsel.

What is an override log in a legal intake system?

It is an append-only record of every time a person approves, edits, rejects or re-tags an AI draft. Each row stores the reviewer, the time, the step, the AI version, the final version and a reason. It lets a firm show that review happened and see where the AI is weakest.

Are small firms actually using AI?

Less than larger firms are. The U.S. Census Bureau reported in 2026 that fewer than 20% of firms with four or fewer employees used AI, based on data through May 3, 2026. That is all businesses, not only law firms, so treat it as context, not a legal-industry figure.

Talk through your intake steps with us

If you run operations at a small firm in Medford, Burlington County or the Philadelphia area, bring us your current intake path, and we will scope it as a custom CRM build or a smaller workflow. We will help you mark each step as draft or approved and tell you what we would build. You will also hear what we would leave manual.