By Monday you’ve got forty names and no idea which one matters. Some of them toured a listing on Thursday. One opened your email twice, then vanished. One is a relocation specialist from out of state, which in most markets means one thing: they are not buying this year.
That pile is where deals go to die.
Not because buyers are flaky. Because sorting forty names by hand takes an hour, and an hour is long enough for a lead who asked for a showing at 9:15am to be talking to a competitor by 9:20.
AI lead scoring and follow-up automation promise to close that gap. Sometimes they do. More often they move the problem somewhere less visible, where it is harder to notice and easier to blame on the software.
What lead scoring actually automates
A lead score is a number attached to a name, and it is usually built from things you already track without noticing.
Recency matters most. Someone who filled in a contact form at 9am and asked about a three-bedroom in a named school district is not the same person as someone who downloaded a market report in March. Repeat visits to one listing beat a single form fill every time. Email opens are weak signal, because Apple Mail preloads tracking pixels, so the open stopped meaning what it used to mean. Replies matter enormously. A person who answers your automated email with a question is telling you they are still reading.
None of this is intelligence. It is arithmetic on behavior you can already see.
AI scoring doesn’t replace that gut, it formalizes it. That is worth saying plainly, because the vendors will not. The machine is not learning something about your market that you do not know. It is applying your own criteria, consistently, at a pace you cannot sustain by hand on a Monday morning.
This is also why the shift from one-off tools to workflows that run themselves matters more than any single scoring product. The unit of value stopped being the app and became the sequence.
The failure mode: your past mistakes, re-ranked
Trouble starts the moment you let the score rank your pipeline and then stop looking at the names yourself.
A model learns from what you marked as a good lead last year, so it learns your patterns, including the bad ones. The renter who toured three properties and mentioned a flipper’s renovation timeline gets scored as a buyer, because that is what similar-looking people did eighteen months ago, when the market was different and your past deals closed at a rate that no longer exists.
An algorithm trained on your past deals has learned your past mistakes too. A score is a guess wearing a number. Sometimes the guess is right. The failure mode is not a wrong guess. It is a confident guess that nobody checks, delivered to a client who expected a human being.
What to automate, and what never to
Here is the split that survives contact with a real week.
| Automate | Why it is safe |
|---|---|
| First response, sent inside a minute | Buys attention, promises nothing |
| Reminder before a scheduled showing | Cuts no-shows, adds no tone |
| Score recalculation as behavior changes | Pure bookkeeping |
| Alert when a lead replies or goes quiet | Surfaces the work, does not do it |
| The sorted call list you open each morning | Removes the hour of triage |
Never automate these:
- Your reply to a direct question. This is the whole relationship.
- Pricing, valuation, and comparable-sales conversations.
- Negotiation, and anything that follows a competing offer.
- The message after a no-show, before a human checks what happened.
- Bad news of any kind.
The pattern: automate the work that is a function of data, and keep the work that is a function of judgment. That distinction is more useful than any feature comparison, and it is the one most vendor demos blur.
If you want to see the mechanics before buying anything, try building a small tool without writing code. Writing one scoring rule yourself teaches you what the inputs actually do, which no demo will.
The consent rules before you text a lead
This is the part that ends careers, so it is worth being precise.
Under the Telephone Consumer Protection Act, a text message is a call. The FCC reached that conclusion in 2003 and the definition is now codified at 47 CFR 64.1200. Consent follows from that. Prior express written consent is generally required before an autodialer or a prerecorded voice can send a telemarketing text to a wireless or residential number, and the consent must identify the number it authorizes.
One correction, because it is widely misreported. The FCC’s 2023 one-to-one consent rule, which limited consent to a single named seller, was vacated by the Eleventh Circuit in January 2025 and removed from the FCC’s rules in August 2025. Any compliance guide still describing it as current is out of date.
What remains live is revocation, and it is the piece most people get wrong. A recipient can revoke consent by any reasonable means, including one keyword reply. You must honor it as soon as practicable and within ten business days, and you may send at most one confirmation text, which cannot contain marketing. State law is not preempted where it is stricter than the federal floor, and private suits carry statutory damages per violation.
This is general information, not legal advice. Consent requirements for automated lead contact vary by state and by how the lead was obtained. Confirm your position with a qualified attorney before you message leads.
What the “respond in five minutes” evidence actually proves
The industry leans on one statistic. Here is what it says.
In 2007, Dr. James Oldroyd and InsideSales.com published a lead response study built on call data from six companies, covering more than 15,000 leads and 100,000 call attempts. It found the odds of reaching a lead by phone at five minutes rather than thirty were about 100 times higher, and the odds of qualifying them about 21 times higher. That is the origin of the five-minute rule. It is not from Harvard.
Harvard’s paper is separate, and different. Oldroyd, McElheran and Elkington published The Short Life of Online Sales Leads in 2011. Across 1.25 million leads at 42 US companies, firms that responded within an hour were nearly 7 times as likely to qualify a lead as firms an hour later, and more than 60 times as likely as firms waiting 24 hours or more. In the same paper’s audit of 2,241 companies each sent a test lead, 23 percent never responded at all.
Note what qualify means in these studies: a meaningful conversation with a key decision maker. Neither study measured close rates or revenue.
Here is the part that matters if you plan to send a text. Oldroyd’s own explanation for the phone effect was presence. Someone who just submitted a form is sitting at their desk next to the phone, and a missed call is not a delayed call, it is a call that never happened. That is the variable the multiplier measures.
An automated text has no such variable. Delivery is guaranteed, there is no missed call, and reachability does not decay the way physical presence decays. So the hundred-times figure is a phone number, and quoting it to justify an SMS drip is a category error.
Speed still matters, and the direction has held up in later re-analysis. But the honest argument for automated follow-up is coverage and consistency, not a multiplier that does not transfer to your channel.
Where automation costs you the relationship
The bungalow in the suburbs gets a robotic “I noticed you liked this listing!” text about a downtown condo. The buyer who asked a real question gets a template that has no idea what they asked.
An agent who is three hours late can apologize and be forgiven. An agent who replies instantly and identically to two hundred people is something else entirely, and buyers can feel the difference. There is a version of automation that turns you into a debt collector with a friendly opening line.
We have looked at the hidden cost of unattended automation elsewhere, and the pattern holds. Automation is efficient at volume and bad at nuance, and nuance is where relationships are actually made.
AI can buy you time, but it can’t buy you charm.
What to measure in the first ninety days
Do not measure how many texts it sent. Measure these instead:
- Median and 90th-percentile time to first human response
- Contact rate, defined as a real conversation rather than a delivery receipt
- Appointment rate, and show rate separately
- Opt-outs per hundred messages
- Reply rate to automated messages, compared with reply rate to human ones
Establish your own baseline before you switch anything on. Most teams cannot state their current numbers, which is exactly why they cannot tell whether the software helped.
Is it worth it?
It depends on whether your pipeline is lazy or your follow-up is lazy.
If deals die because nobody answered, automation is genuinely the fix. If deals die because the listing was priced wrong, no automation will help, and adding a scoring layer only makes the disappointment arrive sooner.
Non-technical teams are already building their own workflows, so the tooling is no longer the hard part. The hard part is deciding what you refuse to automate.
The software amplifies whatever you already are. An agent with a tight process and a good follow-up habit gets better. An agent with neither gets a faster version of the same problem, and a monthly bill for the privilege.
Sources
- Oldroyd, J. B., McElheran, D., and Elkington, T. (2011). “The Short Life of Online Sales Leads.” Harvard Business Review, 89(3).
- Lead Response Management Study (2007). Dr. James Oldroyd with InsideSales.com, presented at the MarketingSherpa B2B Demand Generation Summit.
- Insurance Marketing Coalition Ltd. v. FCC, No. 24-10277 (11th Cir. Jan. 24, 2025).
- FCC Report and Order conforming its rules to that decision, 89 Fed. Reg. (Aug. 29, 2025).
- 47 U.S.C. 227 and 47 CFR 64.1200.
- FCC Report and Order on caller revocation obligations and one-time confirmation texts (2024).
Independent technology writer focused on artificial intelligence, emerging technologies, and digital innovation. Covers AI applications in sports, productivity, and online business.
