November 19, 2026 · 6 min read
Which AI Tasks Need a Licensed Human Review: A Risk Ladder for Offices
Not every AI draft needs the same check. This four-rung ladder sorts real estate AI work from internal notes to anything touching a client's money or legal rights, and names who reviews each rung.

An agent can use AI output directly for private work like meeting notes and brainstorms. Public marketing copy needs a read by the person whose name is on it. A message to one client needs a read by the agent who owns that client. Anything touching a client's finances or legal rights needs a licensed person, and a broker or attorney behind them, before it goes out.
That is the whole ladder. Below, each rung gets examples and a named reviewer, ready for the office handbook. If you run a team and want approvals in one place, our page for team leaders shows how we set that up.
Most offices review everything, so agents use AI in secret, or review nothing, so a wrong number reaches a seller. A ladder lets the check match the risk.
The ladder at a glance
Rung 1, private work. Output only you see. Reviewer: you.
Rung 2, public marketing. Output the public sees under your name or the office name. Reviewer: the named agent, then the brand owner on office accounts.
Rung 3, one-to-one client messages. Output a specific client or lead receives. Reviewer: the agent who owns the relationship, before it sends.
Rung 4, money and legal rights. Output that can change what a client pays, signs, owes or gives up. Reviewer: a licensed agent and the supervising broker, with an attorney for contract language.
To place a task, ask: what happens if this draft is wrong and nobody catches it? A bad brainstorm costs nothing. A wrong net sheet costs a seller real money.
This is a management framework, not legal advice. Your state license law and errors and omissions carrier set what only a licensed person may do. Have your broker and attorney read your final version.
Where the ladder comes from
Two Anthropic sources shaped it. The first sets the rule for the top rung. Anthropic's 2026 Usage Policy update, published October 8, 2026, says that when its products are used in ways that can affect someone's health, legal rights, finances, livelihood or access to essential services, it requires "a qualified 'human in the loop'" with the authority to review and, if necessary, change Claude's recommendations. It also requires that the affected individual be told AI was used. The update lists a November 12 effective date.
The second source is a design choice. Anthropic's Claude for Small Business announcement describes 15 ready-to-run agentic workflows and 15 skills, inside tools including QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace and Microsoft 365. Its model is one line: "Claude does the work; you approve before anything sends, posts, or pays." The machine drafts and a person releases. A brokerage can copy that shape with no software at all.
Rung 1: private work, reviewed by you
Examples: a summary of your own call notes, ten subject lines to choose from, an outline you will rewrite, questions for a first seller visit.
The reviewer is whoever asked, and the check is short: does it make sense? Client-data rules still apply, so do not paste offer terms. Our guide to AI for real estate marketing covers the red list. The mistake is letting a draft climb unnoticed. A call summary becomes a text to the client, and once output leaves your screen, it moves up.
Rung 2: public marketing, reviewed by the name on it
Examples: listing descriptions, open house captions, Reels scripts, blog posts, newsletters, bios, and neighborhood posts about places, not people.
This is where agents already use AI most. NAR's 2026 REALTORS Technology Report says the top AI use case is writing listing descriptions, cited by 75 percent of AI users. The report page also says 56 percent of REALTORS use AI to create social media posts and 52 percent to draft emails and follow-up communications.
That volume needs a fast check, not a committee. Use five questions:
Does every figure, date and feature match the live listing page or the file?
Does it sound like the agent, not a template?
Does it describe the home and the process, never who should live there?
Is the office's required disclosure line present?
Does it promise an outcome, price or timeline?
The agent whose name is on the post answers all five. On office accounts, the brand owner (the team leader or marketing lead) takes a second pass for voice. Our free brand drift check shows if that pass is catching anything.
Rung 3: one-to-one client messages, reviewed before they send
Examples: a reply to a website buyer inquiry, a follow-up after a showing, a check-in to a seller, a text confirming an appointment.
The reviewer is the agent who owns that client, and the check happens before sending. This is the Claude for Small Business shape: AI drafts can sit in a drafts folder, but they never send themselves.
A draft may cite a detail the client never shared, promise a call time the agent cannot keep, or steer. Add these checks:
Is every personal detail something this client told us?
Does the reply answer the question the client asked?
Would the agent say this out loud to this person?
If a reply touches price, terms or timing, it is no longer a rung 3 message. Move it up.
Rung 4: money and legal rights, reviewed by licensed people
Examples: a seller net sheet or closing cost estimate, a plain-language explanation of contract clauses or deadlines, a repair request summary, an offer comparison that recommends one, an explanation of commission or fees, and a price opinion.
The reviewer is a licensed agent who reads the source documents, not just the AI summary. Behind the agent sits the supervising broker, and for contract language, an attorney. Anthropic's rule asks for someone with authority to change the recommendation. In an office, that belongs to people with a license and a duty.
Three habits make rung 4 work:
Source first. Open the contract or settlement figures, then read the draft against them. A draft hides what it left out.
Change in writing. The reviewer edits and signs off in the file or ticket. If no one can see who approved it, nobody did.
Tell the client. Where AI helped prepare the work, say so plainly. Anthropic's policy asks for it when AI use can affect someone's finances or legal rights, and it builds trust.
AI can speed up preparation here. It cannot be the last read.
A worked example: one Tuesday at a ten-agent team
An invented team leader sets the ladder up on Monday. On Tuesday one agent uses AI four times.
8:30, a summary of her call notes. Rung 1. She fixes one line and files it.
9:15, three listing captions. Rung 2. She checks price and features against the live page, cuts a phrase describing the buyer instead of the home, and posts.
11:00, a reply to a buyer asking about showings. Rung 3. She removes a time she cannot keep and sends it herself.
At 2:00 she asks for a summary of a repair request response for a seller. Rung 4. It reads clearly but misses a deadline in the contract. The reviewing agent opens the contract first, catches the gap, and the broker initials the file. The seller is told AI helped with the first draft. Our agent enablement sessions teach this routine.
How Urban Marketing Edge handles this
We run rung 2 for our clients. Every post, caption and blog goes through a post approval system before it is scheduled. Agents send requests through a dashboard, not a text to the team leader's phone. The team leader approves work in one place, once a week, and gets a plain-language monthly report. Most spend under an hour a week.
We do not review rung 4. We do not write net sheets, explain contracts or advise on prices, and our work never replaces your licensed agents or attorney. Our plans page shows what each level includes.
Questions brokers ask
Do I need a licensed person to review every AI-written social post?
The agent whose name is on the post is the licensed reviewer, and the five-question check keeps it fast. Add the brand owner's pass on office accounts. Escalate when a post touches price, terms or anything a client could rely on.
What does Anthropic's human in the loop rule require?
Anthropic's 2026 Usage Policy update says that when its products affect someone's health, legal rights, finances, livelihood or access to essential services, a qualified human who can review and change the recommendation must be involved, and the affected person must be told AI was used.
Can an assistant or a new agent be the reviewer?
For rungs 1 to 3, yes, if trained on the checks and the agent who owns the client or license still answers for it. For rung 4, use a licensed agent reading the source documents, with the broker and an attorney behind them.
Does this mean AI is too risky for transaction work?
No. AI can help draft and organize, and a licensed person must verify and change it. Use it for a first checklist or timeline, then have someone confirm every date against the contract.
How do I stop agents from skipping the review?
Make the check small, name the reviewer by role, and record sign-off where the team already works. A two-minute review with a place to record it gets done. One that needs a meeting does not.
Book a 30-minute strategy call
Bring your list of AI tasks to a call. We will place each one on a rung with you, show you how our approval flow handles rung 2, and say plainly what belongs with your broker and attorney instead. No pitch deck, no pressure.
Book a 30-minute strategy call and we will map the simplest approval setup first. To see how the work runs, start with our plans.


