Contents
The Database Nobody Works
Every agent knows past clients and sphere are the best source of business. Every agent also has a database of a few hundred of them that has not been touched in eighteen months. This is not laziness, and understanding why is the whole design brief.
The feedback loop is broken by design. A cold lead who does not respond gives you information this week. A past client who does not hear from you gives you nothing at all - until four years later, when they use somebody else and you never find out that you lost, or even that there was something to lose. There is no consequence, no notification, and no moment at which anyone in the business is embarrassed. It is the same structural blindness that produces lapsed patients and unfollowed referrals elsewhere in this series, stretched over a much longer timescale.
The horizon defeats intention. Anything with a payoff measured in years loses every scheduling contest against something with a payoff measured in days. An agent with a live buyer and a database is going to work the live buyer, correctly, every single time.
And the obvious automation is the wrong one. The instinctive fix is a monthly newsletter or an automated market update, and it mostly does not work - not because the content is bad but because it is not a relationship. A past client who receives twelve automated market reports a year and no human contact has been marketed at, not stayed in touch with, and they will refer the agent who remembered their kid's name.
So the build has an unusual shape. It is not a messaging system. It is a memory system with a compliance layer. Its job is to surface the right handful of people each week, with enough context that a human can have a genuine conversation, and to make sure you are still permitted to have it.
For where this sits among an agency's other candidates, see AI automation for real estate agencies and AI agents for real estate. The in-transaction side is covered in transaction deadline automation.
The 540-Day Problem
Here is the constraint that should shape the entire strategy, and it is almost entirely absent from the guides currently ranking for this topic.
The FTC's Telemarketing Sales Rule restricts calls to numbers on the National Do Not Call Registry. Under 16 CFR 310.4(b), it is an abusive practice to call a person whose "telephone number is on the 'do-not-call' registry, maintained by the Commission, of persons who do not wish to receive outbound telephone calls," unless the seller "can demonstrate that the seller has an established business relationship with such person, and that person has not stated that he or she does not wish to receive outbound telephone calls."
And 16 CFR 310.2 defines that established business relationship with specific clocks. It arises from:
- "The person's purchase, rental, or lease of the seller's goods or services or a financial transaction between the person and seller, within the 540 days immediately preceding the date of a telemarketing call"
- "The person's inquiry or application regarding a good or service offered by the seller, within the 90 days immediately preceding the date of a telemarketing call"
Now put those numbers against the business. 540 days is under 18 months. A homeowner who buys a house is typically not moving again for years. The exemption expires long before the client becomes a customer again. An agent who closes a sale, does nothing for three years, and then starts calling the database from a list has an eighteen-month-expired relationship and a number that may well be on the registry.
Similarly, the 90-day inquiry window is brutally short against real estate's sales cycle. A lead who enquired four months ago and is buying next spring has already aged out of that basis.
The strategic conclusion is the important part of this article. Because the transactional exemption expires and the buying cycle does not fit inside it, the durable basis for staying in touch is consent you obtained and maintained, not an exemption you are relying on. Which reframes the entire purpose of past client follow-up:
The goal of the first two years is not to generate a transaction. It is to convert a client relationship into an ongoing, consented, welcome one - so that you are still allowed to be in touch in year five, when it actually matters.
That is a completely different brief from "nurture them until they list," and it produces different tactics: ask for permission explicitly while goodwill is highest, make the ongoing contact something people actually want, and treat every opt-out as final. Note also that 310.4(b) requires honouring a person's own request regardless of any exemption - a company-specific do-not-call request overrides everything.
Important limits on what we have told you. This is the FTC's Telemarketing Sales Rule and it concerns telemarketing calls. There is a separate FCC regime under the TCPA with its own consent requirements, particularly for autodialled calls and texts, and it has been through significant litigation recently - we deliberately are not summarising its current state here, because we could not verify it to the standard this article holds itself to and a confidently wrong answer would be worse than none. State laws add further requirements, several of them stricter. Get your actual position from your broker's counsel. The design principle below - build for consent rather than for exemptions - is robust to all of it, which is precisely why we recommend it.
Designing the Database Consent-First
If consent is the asset, the database must be built to record and protect it. Most agent CRMs are built to record contacts, which is not the same thing.
Store consent as data, not as an assumption. For every contact and every channel - call, SMS, email - record what permission exists, when it was given, how it was captured, and the exact wording they agreed to. "They're a past client" is not a consent record. If you cannot produce the evidence, you do not have it.
Ask at the point of maximum goodwill. Closing day, or shortly after, is when a client is happiest and most willing. That is the moment to ask - explicitly and in plain language - whether they would like to hear from you, how often, and by which channel. An agent who does this on every transaction builds a permissioned database as a by-product of doing business, and one who does not is relying on an exemption that expires in 540 days.
Make opt-out trivial and absolute. One click, honoured everywhere immediately, across every channel and every campaign. A contact who unsubscribes from a newsletter and then receives a birthday text has been told their preference does not count. Suppression must be central, not per-campaign. See opt-in and STOP handling.
Track the clocks. If you are relying on an established business relationship for any outbound calling, the system should know the transaction date and show you when that basis lapses - and stop presenting the contact for call-based outreach when it does. Making an expiring legal basis visible in the interface is a small feature that prevents a category of problem entirely.
Record the relationship, not just the record. This is what makes the difference between a database and a memory: the names of the children, the dog, the job, the renovation they were planning, what they were anxious about during the transaction, where they came from. Every one of those makes a future conversation real rather than transactional. Capture it during the transaction when it is known, because nobody will reconstruct it in three years.
On SMS specifically, if you send over a US 10-digit long code you are also inside A2P 10DLC. Twilio describes it as "the standard that United States telecom carriers have put in place to ensure that SMS traffic to US end users through long code phone numbers is verified and consensual," requiring a registered Brand and Campaign, and notes unregistered senders "will receive additional carrier fees for sending unregistered traffic." See our A2P 10DLC guide and the glossary entry.
Surface, Do Not Send
The core mechanic, and the one that distinguishes this from every generic drip campaign.
The automation produces a short weekly list. A human does the contacting. Ten or twelve people, each with a reason and the context attached. That is it. It sounds unambitious and it is the version that actually generates referrals, because what past clients respond to is being remembered by a person, and no sequence simulates that convincingly.
Reasons to surface someone, roughly in order of quality:
- Something happened to them. A life event you know about. This is the best reason and it comes from your notes, not from a rule.
- The purchase anniversary. Genuine, easy, and welcome if it is a real message rather than a card-shaped automation.
- Tenure milestones. Three, five, seven years in the home - the points at which people begin to think about moving, refinancing or renovating. This is where the long horizon becomes actionable.
- Something happened near them. A comparable sale on their street. Useful, specific, and genuinely interesting to a homeowner - but it must be real information about their actual property's context, not a template with an address merged into it.
- Time elapsed. The weakest reason and the necessary backstop: nobody in the database should go longer than a set period without a human touch.
Rank and cap. The list should be short enough that it always gets done. A weekly list of forty names is a list that gets skipped in week three, and a skipped list is worse than none because it teaches the agent the system does not work. Twelve is better than forty. This is the same capacity-driven sizing argument as the gym attendance build in this series - tune the output to what a human will actually do, not to what the data could support.
Where AI genuinely helps here: summarising everything you know about a contact into three lines before a call, so the agent walks in informed rather than scrolling notes. Drafting a personal message from real recorded details for a human to edit. Extracting facts from conversation notes into structured fields so the memory compounds instead of rotting. All low-risk, all high-value.
Where it does not help: writing the relationship. A message generated from nothing but a template and a first name reads exactly like what it is. If the system has no real context for a contact, that is a signal to have a conversation and collect some, not to send something anyway. See what AI agents can do reliably.
A Worked Example, With the Arithmetic Shown
Illustrative. We report no client results, and this is the build in this series with the longest feedback loop - which is exactly why the arithmetic matters more than usual.
Baseline. An agent has 340 past clients and sphere contacts. Suppose the current state is honest: perhaps 60 receive any meaningful contact in a year, and the database produces 4 transactions a year from repeat business and referrals combined. Average commission per transaction: $9,500 - your figure.
Current annual value from the database: 4 x $9,500 = $38,000.
The intervention. A surfaced weekly list of 12 contacts, worked by the agent. Annualised: 12 x 50 = 600 contacts a year, which across 340 people is roughly 1.8 touches each - a modest and achievable cadence, and the point is that it is consistent rather than frequent.
Time cost, which is real and must be counted. If each contact takes an average of 6 minutes including the call, the note and the occasional follow-up: 600 x 6 = 3,600 minutes, or 60 hours a year, a bit over an hour a week. That is the actual investment, and it is the reason this fails without a short list.
Modelling the return. Rather than assert a conversion rate we cannot verify, invert it. Suppose the build costs $2,500 plus $40/month - your quotes; see automation build cost. First-year cost: $2,500 + $480 = $2,980. Add the agent's 60 hours valued at, say, $60/hour of opportunity cost: 60 x $60 = $3,600. Total first-year cost: $6,580.
Break-even: $6,580 / $9,500 = 0.69 transactions. The system needs to produce fewer than one additional transaction a year to pay for itself, including the agent's time.
Against a database of 340 people who already know and presumably liked you, one additional transaction a year is a low bar. It is the clearest case in this entire series - and note that the largest cost component is the agent's own hours, not the software, which is why the design priority is a list short enough to actually work rather than a cleverer system.
Two honest caveats. First, the payoff is lagged. Year one may produce very little, because you are rebuilding relationships that went cold and re-establishing permission. Judging this build on twelve months will usually kill something that was working. Commit to three years or do not start. Second, the counterfactual is unknowable here. Unlike the gym example in this series, you cannot cleanly hold out a control group of your own past clients over a multi-year horizon and you probably should not try. Accept that this is a build justified by mechanism rather than by measurement, and be honest about that rather than manufacturing attribution.
Our ROI of AI agents piece covers defensible cases, payback period the framing, and AI automation cost the budget ranges.
Running It Over Years
Build order. One, clean the database and record what consent actually exists per channel - this will be sobering and it is the foundation. Two, capture relationship context for recent clients while it is still recoverable. Three, build the surfacing rules and the weekly list. Four, add the consent-expiry visibility. Five, add drafting assistance. Six, and only then, consider any automated sending, and keep it to things people would miss if they stopped.
Ask for permission on every transaction from now on. This is a process change, not a software feature, and it is worth more than the rest of the list combined. Within two years it changes what you are legally and practically able to do.
What to measure. Percentage of the database contacted in the last 12 months - the single number that best predicts whether this is working. Consent coverage by channel, which should be rising. List completion rate each week, your leading indicator of abandonment. And transactions attributed to the database, understanding that attribution here is soft and lagged.
Failure modes:
- Drifting into bulk sending. The gradual failure. It starts as a surfaced list, someone notices you could just email all 340, and within a year it is a newsletter nobody reads and the referrals stop. The whole value was that a human made contact.
- Contacting people who opted out, because suppression was per-campaign rather than central. Unrecoverable with that individual.
- Relying on an exemption that lapsed. The 540-day clock runs whether or not your CRM knows about it. Make it visible.
- Stale context used as if current. Referencing a job someone left three years ago is worse than referencing nothing. Timestamp facts and let old ones age out of the summary.
- The list that grows until it is ignored. Cap it and enforce the cap.
- The silent stop. The scheduled job fails, no list appears, and its absence is indistinguishable from a quiet week - and unlike an unanswered lead, nothing else will remind you. Alert on zero output. See observability and error handling.
- Abandonment in month four, which is the most likely outcome of all. The mitigation is not software - it is a short list, a fixed slot in the week, and a decision made in advance to judge it in year three.
If your constraint turns out to be inbound you are not capturing rather than a database you are not working, that is a different and cheaper build - see what an AI receptionist costs a real estate agency and missed-call text back. General nurture mechanics are in AI agents for lead nurturing.
For a second opinion on which of these to build first, the free automation audit works through it, and automation services explains how we scope. On the general boundary, see when not to use an AI agent.
FAQ
- How long can I rely on an established business relationship to call a past client?
- Under the FTC's Telemarketing Sales Rule, 16 CFR 310.2 defines the established business relationship by a purchase, rental, lease or financial transaction "within the 540 days immediately preceding the date of a telemarketing call," or an inquiry or application "within the 90 days immediately preceding" one. Since 540 days is under 18 months and homeowners typically move on a much longer cycle, that basis usually expires well before a past client is ready to transact again. Confirm your position with your broker's counsel.
- What does that mean for how I should think about past client follow-up?
- It means the goal of the early years is not to produce a listing - it is to convert a transactional relationship into a consented, welcome one, so that you are still permitted to be in touch in year five when it matters. Ask for permission explicitly at closing when goodwill is highest, record exactly what was agreed and when, and make the ongoing contact something people would actually miss. That basis does not expire on a clock the way an exemption does.
- Does an exemption override someone asking me not to call?
- No. 16 CFR 310.4(b) conditions the established business relationship exemption on the person not having "stated that he or she does not wish to receive outbound telephone calls." A company-specific do-not-call request stands regardless of any relationship. Practically, this means suppression must be central and absolute across every channel and campaign - someone who unsubscribes from your newsletter and then gets a birthday text has been shown their preference does not count.
- Is this article a complete account of the rules?
- No, deliberately. What we have quoted is the FTC's Telemarketing Sales Rule, which concerns telemarketing calls. There is a separate FCC regime under the TCPA with its own consent requirements for autodialled calls and texts, which has been through significant recent litigation - we are not summarising its current state because we could not verify it to the standard this article holds. State laws add more, some stricter. Get your position from your broker's counsel. The design principle of building for consent rather than exemptions is robust regardless.
- Should the system send the messages automatically?
- Mostly no. The mechanic that generates referrals is a person making contact, and no sequence convincingly simulates being remembered. Have the automation produce a short weekly list of ten or twelve people, each with a reason and the context attached, and have a human do the contacting. The characteristic slow failure of these systems is drifting into bulk sending because it is easier - at which point it becomes a newsletter and the referrals stop.
- How long should the weekly list be?
- Short enough that it always gets done, which for most agents means around a dozen. A list of forty names gets skipped in the third week, and a skipped list is worse than no list because it teaches the agent that the system does not work. Tune the output to what a human will realistically do rather than to what the data could support - the constraint is attention, not detection.
- What should trigger a contact?
- In descending order of quality: a life event you actually know about, from your own notes; the purchase anniversary; tenure milestones at three, five and seven years when people start thinking about moving, refinancing or renovating; a genuinely relevant comparable sale near their home; and finally elapsed time as a backstop so nobody goes longer than a set period without a human touch. The first is by far the best and it comes from recorded context, not from a rule.
- How do I know if it is working?
- Imperfectly, and you should be honest about that. Track the percentage of your database contacted in the last twelve months, consent coverage by channel, and weekly list completion rate - those are leading indicators you control. Transaction attribution is lagged and soft, and unlike a short-cycle campaign you cannot cleanly hold out a control group of your own past clients over several years. This is a build justified by mechanism rather than measurement, so commit to three years or do not start.
If this resonated
Module 02 is the long version of this post.
Eval set, refusal pattern, kill switch. Build it in a weekend.
