By Sagar Shankaran, Founder of CallSphere
Which after-hours brokerage calls a cheap model can answer, which need the careful one, and which go straight to the managing broker. With costs and rules.
Key takeaways
Which of these two calls should cost you the same amount of money?
Call one, 8:52 p.m.: a man is standing in front of a sign rider on Camden Court and wants to know the list price and whether it has a garage. Call two, 8:54 p.m.: a seller who has been on market 46 days wants to know whether she can cancel her listing agreement and go with the brokerage that just mailed her a flyer.
Today, in most brokerages, both of those go to the same place — voicemail, or a duty agent's cell, or a $6-an-hour answering service reading a script. They are not the same job. One is a lookup. The other is a licensed conversation about an executed contract, and if it goes badly you lose a listing and maybe get a letter from the state real estate commission.
Because until recently there was only one queue. The front desk — the person a lot of offices still call the director of first impressions — answered everything between 9 and 5, and after 5 it went to a rotation. The rotation is the part everyone tolerates and nobody defends. The agent on duty Thursday night is showing property. She sees the missed call at 10:40 p.m. The sign call is dead by then; the seller is already talking to a competitor.
Then a wave of automated answering arrived and made it worse in a specific way: it gave the seller-cancellation call the same three-sentence script as the garage question, because the cheap tool was the only tool. Owners tried it in 2024, heard their after-hours system tell a distressed seller to "visit our website," and turned it off.
Model routing means the fast, cheap system handles the routine contact and the expensive, careful one is only woken up when the question is genuinely hard — and in a brokerage, "hard" has a legal definition, not a technical one.
This stopped being clever and became normal practice this year. Cisco is rolling a personal AI agent out to roughly 90,000 employees and built routing into it deliberately, to keep cost in line with capability. That is a big company solving a big company's bill, but the shape of the answer transfers exactly to a brokerage: most contacts are cheap questions, a minority are expensive ones, and you should not pay expensive prices for cheap questions or cheap prices for expensive ones.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for real estate in your browser — 60 seconds, no signup.
Two other things made it practical for a small office. Frontier AI now costs roughly a tenth of what it did in 2025, so the strong model on the hard call is affordable rather than theoretical. And end-to-end voice answers in about a fifth of a second and can look something up or book something in the middle of the conversation — so the routine branch does not merely deflect, it books the showing.
flowchart LR
A["Sign call, web chat or text after 6pm"] --> B{"Routine lookup or representation question?"}
B -->|Routine| C["Fast model: price, status, square footage, HOA dues"]
B -->|Judgement| D["Strong model: reads the file, drafts a careful answer"]
B -->|Never automate| E["Straight to the managing broker's line"]
C --> F["Showing booked in ShowingTime, lead written to the CRM"]
D --> G["Duty agent gets a briefed handoff with the file summary"]
E --> G
This is the part a vendor cannot answer for you, and it is the only part that matters. Here is the line most brokerages end up drawing, and it is drawn along the boundary of what requires a license.
Routine — the fast lane. List price. Days on market. Square footage and bedroom count as published in the MLS. HOA dues as stated in the listing. Open house times. Whether a property is still available. Booking a showing. "When is my closing." "Which office is closest to me." Confirming that a co-op agent's showing request was received. These are facts already published in your MLS or your calendar. Reading them back is not advice.
Judgement — the careful lane. A buyer who wants to write an offer tonight. Anything about an inspection amendment or a repair credit. An appraisal that came in low. A seller asking about a price reduction. A question about how the buyer's agent is being compensated — a live question at every showing since buyer representation agreements became mandatory in August 2024. Anything about earnest money. Anything where the honest answer starts with "it depends what the contract says."
Never automate — straight to a person. Wire instructions, in any form, in any direction. Full stop; this is where the money gets stolen. Anything that sounds like a complaint about a licensee. And any question that touches a protected class — a caller asking what kind of people live in a neighbourhood, or which schools are "good." That is a steering question, the answer is a Fair Housing problem, and the correct behaviour is a warm handoff to a licensed human who knows how to redirect it to objective sources. Do not let any model, cheap or expensive, freelance on that one.
The designated or managing broker. Not the marketing coordinator who bought the tool, not the vendor's onboarding template, and not the team lead who wants more leads routed to her people. The person whose license is on the wall is the person who signs off on which questions a machine may answer, and that decision belongs in your office policy manual next to the sections on advertising and escrow handling.
Review it monthly for the first quarter, then quarterly. The way to review it is to read transcripts — twenty routine ones and every single escalation — and ask one question of each: was this the right lane? You will find two kinds of mistake. Things sitting in the fast lane that should not be, which is the dangerous kind. And things escalating that did not need to, which is the expensive kind. Fix the first immediately and the second at your leisure.
Illustrative numbers for a three-office brokerage with about 110 agents. Substitute your own call volume from your phone system's report.
| Line item | Volume per month | Cost each | Monthly |
|---|---|---|---|
| Routine contacts on the fast lane | 1,476 | $0.03 | $44 |
| Judgement contacts on the strong model | 306 | $0.38 | $116 |
| Straight-to-broker escalations | 18 | human time | — |
| Total running cost | 1,800 | $160 | |
| If everything ran on the strong model | 1,800 | $0.38 | $684 |
The $524 saving is real but is not the argument. Suppose you convert four additional buyer sides a year out of calls now answered at 8:52 p.m. instead of returned at 10:40. At a $410,000 average sale price and a 2.5% side, that is $10,250 of gross commission each; if the house keeps 25% after the split, four sides is roughly $10,250 a year to the brokerage. Routing is what makes answering all 1,800 cheap enough to bother with.
Still reading? Stop comparing — try CallSphere live.
See the real estate AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
Three honest limits.
The classification is not perfect, and the errors are not symmetric. A seller who opens with "quick question about my listing" and then, ninety seconds in, asks whether she can walk away from the agreement is a hard call that started as a soft one. Insist that any system you buy can change lanes mid-conversation, and test that specific scenario before you sign anything.
Emotion is not a routing category, but it should be an escalation trigger. A buyer whose financing fell through two days before closing is asking a factual question in a voice that needs a person. The cheapest reliable rule: if the caller repeats themselves or raises their voice, hand off.
And the strong model still does not have a license. It can read the executed contract, summarise the deadlines, and tell the duty agent what the seller is upset about before he calls back. It should not tell the seller what her options are for terminating. The value of the careful lane is a better-briefed human, not a replaced one.
Pull last month's call detail and listen to fifty after-hours calls, or read fifty web chat transcripts. Tally them into the three lanes above yourself, before you talk to a vendor. Owners who do this are usually surprised how heavily the fast lane dominates — the sign call asking price and availability is the most common after-hours contact in this business, and the one most likely to go unanswered.
Only if you let it answer from memory. The routine lane should be reading live from your MLS feed and your calendar, and it should say the source out loud: "the MLS shows it active at $429,900 as of this morning." If a vendor cannot show you where each answer came from, that is the thing to walk away from — not the price.
That it does the opposite, and then prove it with the routing rules. Every captured contact should land in the CRM assigned to whoever was on duty or on the listing, with the transcript attached. The agent who was showing property at 8:52 p.m. gets a briefed lead at 9:05 instead of a missed call at 10:40. Put the assignment logic in writing before you turn it on, because you will only have this argument once.
Yes. Your state commission's advertising and supervision rules do not exempt software, and a managing broker's duty to supervise covers what your phone line says to the public. Write down which lane answers what, who reviews transcripts, and how often. Texas and California both had new state AI statutes take effect on 1 January 2026, and several other states have their own — a written policy is the cheap version of that conversation.
If the routine lane is the part you want first, that is the part CallSphere builds: AI voice and chat agents that answer the brokerage line and the website chat at any hour, book showings and appointments, and capture the lead with the transcript attached. The hard calls still belong to your duty agent and your managing broker — the point is that they arrive briefed, at 9:05, instead of as a voicemail nobody hears until Friday.

Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
North Platte, Nebraska family practices serve Union Pacific railroaders on 24-hour schedules. An AI phone agent answers, books, and triages around the clock.
Sixty-one pages of addendum land three days before a DOT letting. How model routing gets an estimator the quantity changes that actually move the bid.
Model routing sends routine product requests to a cheap model and hard ones to a strong one. Here is who draws the line in a B2B software company, and how.
Cheap model for potholes, strong model for water quality, a pager for sewage in a basement. How a public works superintendent writes the triage table.
Luxury buyers in Scottsdale call nights and weekends, and the first team to answer usually wins. How AI voice and chat agents keep Arizona brokerages first.
In Fort Wayne remodeling, the first contractor to respond usually wins the job. How an AI receptionist makes your firm first to answer every homeowner call.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI