Where AI Work Stops and Attorney Decisions Begin

AI tools are good at turning a messy production into something that looks finished. They can group emails, label them and summarize them in a clean table. That neatness is the risk.

Suppose an AI tool sorts the emails in a contract dispute and labels 40 of them "privileged." The spreadsheet looks complete, and a busy reviewer may treat it as done. But what if the tool assumed that anyone with "Legal" in their email signature is a lawyer? That assumption is a weak premise: a starting point that looks reasonable but may be wrong. A neat format hides it. Nothing in the table tells you the label rests on a guess.

This course draws one line and holds it throughout:

• AI prepares. It organizes material and suggests reasons.
• A named attorney decides. Every privilege determination and every discovery-response decision belongs to a specific attorney.

In this lesson you sort review tasks onto the right side of that line. You also learn to spot AI wording that crosses it, and you write a task boundary list for one of your own matters. A task boundary list is a short written statement of what AI may and may not do on that matter.

On the preparation side, AI can do a lot of useful work:

• Sort documents into review categories, such as "possible attorney communication" or "business only."
• Summarize what a document or thread is about.
• Extract metadata. Metadata is information about a document rather than its content, such as author, recipients and date.
• Group similar documents, such as all replies in one email thread or near-duplicate drafts of the same memo.
• Suggest a reason an item might belong in a category.

Here is an example. A VP of Sales emails in-house counsel Dana Ruiz about a termination clause. The AI pulls the author, recipients and date. It summarizes the email as "asks counsel how to read the termination clause." It groups the email with six replies. Then it suggests a reason: "may request legal advice from in-house counsel."

All of that saves a reviewer from starting cold. None of it settles anything. Each output is preparation for review: material an attorney uses to make a decision, not the decision itself.

Some calls are decisions, and they belong to an attorney:

• Whether a document is privileged or work product. Attorney-client privilege protects certain confidential communications made for legal advice. Work product protects certain materials prepared in anticipation of litigation. Your jurisdiction's rules govern both, and applying them is the attorney's call.
• Whether to withhold, redact or produce a document. Redact means to black out the protected part and produce the rest.
• Whether to assert an objection to a discovery request.

This course does not teach AI to make these calls or to recommend them. An AI tool can suggest a reason an item belongs in a review category, such as "sent to outside counsel, asks about indemnity." It should not produce "withhold this" or "object as overbroad" as if that were the answer.

Here is a quick test you can use. If the output tells you what the document is or what to do with it, it has crossed the line. If it tells you what might be relevant to your decision, it is preparation.

A quiet decision is an AI output that reads as if a decision were already made, even though no attorney made it. Watch for three common forms:

1. A determination stated as fact. "Document HB-0412 is privileged." The word is does the damage.
2. A log entry with no review status. A privilege log row listing a basis such as "attorney-client communication," with nothing showing whether anyone reviewed it.
3. An objection that reads as final. "Defendant objects to Request No. 7 as overbroad and unduly burdensome." It is ready to serve, and nothing marks it as a draft.

All three look normal. A privilege log and a set of responses are supposed to contain sentences like these, which is why they slip through. The fix is to rewrite each one as a suggestion with an open status. For example: "Suggested reason, UNREVIEWED: sent to in-house counsel; appears to request advice on the termination clause."

The rewritten version says the same facts and still helps the reviewer. The difference is that nobody can mistake it for a call that was already made.

Use one word for every open item in the workflow: UNREVIEWED. Put it on sort categories, suggested reasons, privilege log entries, draft responses and draft objections.

Why one word? Picture what happens without it. One draft says "pending," a log uses "TBD," a response says "draft – check," and someone else writes "needs review." Before anything goes out, you have to remember every variation and search for each one, and you will probably miss one.

With a single tag, one search for UNREVIEWED across the log, the responses and the review notes finds every item no attorney has decided yet. If the search returns anything, the work is not ready.

Here are some rules for the tag:

• AI outputs start as UNREVIEWED. Your instructions to the tool should say so.
• The tag comes off only when a named attorney decides. Formatting the item, checking it or moving it to another sheet does not remove the tag.
• Write it the same way every time, in capitals, so the search is reliable.

Later lessons show how to record the attorney's decision when the tag comes off.

"Decision-maker: the legal team" sounds responsible, but it leaves no one accountable. When a log entry is questioned months later, "the team" cannot explain why a document was withheld.

Each matter needs a named decision-maker: one specific attorney recorded against each privilege determination and response decision. You might write, for example, "Decisions: M. Okafor." If your practice uses a backup attorney for absences, name that person too.

Paralegals keep a central role. A paralegal may:

• run and check the AI sort,
• prepare log entries and draft responses,
• route flagged items to the attorney,
• track which items are still UNREVIEWED.

What a paralegal does not do is make the call. When the decision is recorded, it is recorded against the attorney who made it. This protects the paralegal as much as the client. Nobody should end up as the de facto decision-maker just because they were the last person to touch the spreadsheet.

A task boundary list fits on one page and covers four things for one matter:

1. Tasks AI may do. For example: sort, summarize, extract metadata, group similar documents, suggest reasons.
2. Tasks AI may not do. For example: determine privilege or work product, decide to withhold, redact or produce, finalize any objection or response.
3. The attorney who decides. A name, not a team.
4. Where drafts stop for review. For example: every log entry, every objection and every withheld or redacted document stops at UNREVIEWED until the attorney decides.

It helps to head the list with the matter name and jurisdiction, so anyone using it knows which rules apply.

Here is an example for Harlow v. Brightline Supply:

AI may: sort the production, summarize threads, extract author, recipients and date, group duplicates, suggest reasons tagged UNREVIEWED. AI may not: decide privilege or work product, decide what to withhold, redact or produce, finalize objections. Decides: M. Okafor. Stops: all log entries and draft objections stop at UNREVIEWED for M. Okafor.

The list does not slow the work down. It reminds everyone that a tidy draft is not a reviewed one.