Every contractor I have worked with has the same folder of dead money, and most of them do not think of it that way. It is the pile of estimates that went out, got a polite "let me think about it," and were never touched again. The homeowner did not say no. The job did not go to a competitor, at least not immediately. The quote just sat there while the business moved on to the next call, because chasing old estimates feels awkward and there is always a newer, warmer lead to deal with.
As Gideon Wafula, AI Automation Engineer, I have spent the past year building revenue automations for local businesses, and I keep coming back to this one because the economics are so lopsided. The lead was already paid for. The site visit already happened. The price is already written down. The only thing missing between the business and the money is follow-up, and follow-up is exactly the kind of repetitive, rule-based, easy-to-forget work that automation does better than people. This post is a teardown of the estimate follow-up system: why the unsold pile exists, what the sequence looks like, how it differs from speed-to-lead, and what it costs to run.
The pattern shows up in industry surveys year after year. Field-service benchmark reports consistently find that the typical home services company converts only a fraction of the estimates it sends, and that the gap between average operators and top performers is not marketing spend or pricing. It is follow-up discipline. Contractor coaching data suggests businesses that follow up consistently on open estimates close on the order of twenty to thirty percent more jobs than those that send a quote and wait. In a 2025 contractor survey covered by the trade press, a large share of high-revenue companies said that working their unsold estimates accounted for a meaningful slice of total income, low double digits as a percentage, which is remarkable for work that requires no new leads at all.
The reason people do not follow up is not laziness. A busy HVAC or plumbing operation quotes several jobs a week, and each open quote needs a touch at a different time: two days after sending, then a week, then two weeks. Nobody holds that schedule in their head. The CRM technically has reminders, but reminders that require a human to act get snoozed. And there is a psychological tax too: calling someone who did not respond feels like pestering, so the follow-up that does happen is usually one email, sent once, then silence.
Meanwhile the homeowner is not sitting there rejecting you. They collected three bids, got busy, and the whole project drifted. When a friendly text arrives ten days later asking whether they had any questions about the quote, a surprising number of them reply within minutes. The job was never lost. It was just unattended.
The system I build is deliberately boring, which regular readers will recognize as a theme. It watches wherever quotes live, whether that is Jobber, ServiceTitan, Housecall Pro, QuickBooks, or a spreadsheet, and when a quote passes 48 hours without a response, the sequence starts.
A typical build runs five touches over about three weeks, alternating channel and intent:
Two rules make or break the whole thing. First, every reply stops the sequence instantly and alerts a human. Nothing destroys trust faster than a customer answering a question and then receiving an automated nudge the next day. Second, every message is written to sound like the owner typed it on their phone. No corporate templates, no "Dear Valued Customer." The customer should never suspect a robot, and with a well-built sequence they never do, because a human really does take over the moment there is a conversation to have.
The schedule above needs no AI at all, and I tell clients that plainly. Where a language model earns its keep is at the edges. When a customer replies with a question, the model drafts an answer from the quote details and the business's pricing rules, and the owner approves it from their phone in one tap instead of composing from scratch in a parking lot. The model also classifies replies, so "we went with someone else" closes the record with a reason code, "can you do it for less" flags a negotiation for a human call, and "we're waiting on insurance" schedules a check-in for next month rather than next week. That classification step is what turns a dumb drip into something that behaves like a competent office manager.
Readers of this blog will notice this is the third automation I have covered that lives on the same timeline of a customer relationship, so it is worth placing them side by side. Speed-to-lead operates in the first five minutes, when a new inquiry arrives and the fastest responder usually wins the job. Database reactivation operates months or years later, waking up past customers who already trust the business. Estimate follow-up sits exactly between them, in the two-to-three-week window after a price has been named, and it is arguably the highest-intent window of the three: this is a person who asked for a number and received one. They are not a cold lead or a lapsed customer. They are a decision waiting to be nudged.
The three systems also compound. A business running all of them has automated the entire commercial spine: no inquiry goes unanswered, no quote goes unchased, and no past customer is forgotten. That full pipeline is what I increasingly get asked to build as a single project rather than three separate ones.
The running costs are small. The automation layer, whether self-hosted n8n or a platform subscription, plus SMS fees and a modest amount of model usage typically lands well under a hundred dollars a month for a single-location business, with SMS volume being the main variable. Setup is the real cost, and in the current market a scoped build like this from an independent automation engineer generally runs from several hundred to a few thousand dollars depending on how messy the quote data is and how many systems need connecting. Agencies package the same thing inside monthly retainers, commonly in the range of five hundred to two thousand dollars a month for high-ticket trades, which is worth knowing when you compare offers.
Against that, run the math on your own numbers rather than trusting anyone's case study. Count the estimates you sent last month that got no response. Multiply by your average ticket and your normal close rate. Field-service platform data suggests automated sequences recover a real fraction of initially silent estimates, with most recovered jobs closing on the second or third touch, so even a conservative assumption of one extra closed job a month pays for the entire system many times over for most trades. For an HVAC or roofing business with four-figure average tickets, the payback period is usually measured in weeks. Do not fabricate precision here; the honest claim is simply that some percentage of your silent quotes are recoverable, that percentage is not zero, and right now you are recovering none of them.
If you want to test the thesis before automating it, do this manually for two weeks. Pull every estimate from the last ninety days that never got an answer and send each one a single, personal text asking if they had questions about the quote. Track what comes back. Nearly every contractor who runs this experiment books at least one job from the graveyard, and that result is what tells you the automated version is worth building, because the automation simply does the same thing forever without anyone remembering to do it.
When you are ready to systematize it, the build order matters: get the quote data flowing first, then the stop-on-reply logic, then the schedule, and only then the AI layer. If you want to see how this fits into the broader set of revenue automations local businesses are actually buying this year, I keep a running overview on my services page, and the manual-first principle applies to every system on it.
Gideon Wafula builds custom AI automation systems, n8n, WhatsApp, Voice AI, and more.
See Services →