Skip to content
SearchSubscribe
TPL-2026-017·preprint·2026-04-30

Real-Estate Site-Search Automation: Hours-Saved and Filter-Accuracy in a Pre-Launch Venue Search

mhgreal-estateautomationcase-study

Abstract

Pre-launch entertainment-venue site searches are constrained by an unusual cocktail of filters: zoning compatibility, square-footage and ceiling-height fit, ABC-license geography (in NC, alcohol licensing depends on municipal jurisdiction and proximity to schools/churches), and proximity to demand. We instrumented an 8-week site search for Mile High Golf, a pre-launch indoor-golf entertainment venue, comparing a manual baseline (commercial RE advisor working alone) against an agent-assisted pipeline (advisor + a Claude Code agent that ingests county GIS, ABC-licensing maps, and listing feeds; pre-filters; and ranks). The agent-assisted pipeline saved approximately 14 hours of advisor time per week at week 1, climbing to 22 hours per week by week 8 as filter calibration improved. Agent precision against advisor verdict started at 61% and reached 87% by week 8. Recall at agent-pass-through-to-advisor stayed above 95% throughout — the false-negative rate was the metric the advisor was unwilling to compromise. We discuss the limitations of single-venture, single-geography study and offer a reproducibility checklist.

1. Background — MHG site-search context

Mile High Golf (MHG) is a pre-launch indoor entertainment venue in the Lake Norman corridor of North Carolina. The product is a hybrid: a craft-bar with a TruGolf-style simulator anchor, a cornhole arena, and an outdoor patio. The site-search constraints are unusual relative to a conventional retail or restaurant search [1][2]: ceiling height ≥ 16 feet (the simulator throw distance), zoning that supports both indoor recreation and on-premise alcohol, and ABC-license geography that satisfies North Carolina’s distance-from-school/church rules [3]. The lender (an SBA 7(a) participating bank [8]) has its own overlay constraints around lease length and tenant-improvement allowances.

The original site search was anchored to Hickory, NC. After the original deal failed, the search was repointed to Denver, NC (7521 Eastern Medical Dr) [9]. The 8-week instrumentation window covers both phases of the search. Throughout, the commercial-RE advisor of record was a senior partner at Ascent Real Estate, who agreed to participate in the comparison study and contributed the manual-baseline hours estimates.

The question we want to answer for any operator considering this kind of automation is: does the agent-assisted pipeline save meaningful advisor time, and at what false-negative cost? The advisor was clear from the outset that recall — not missing acceptable listings — was the metric they would not trade. Any precision improvement is a gift, but a recall regression is disqualifying.

2. Methods

2.1 Filter pipeline

The agent (a TruPath specialist named SUMMIT [10]) ingests inbound listings from three sources: LoopNet feed, advisor email forwards, and a county-by-county GIS scrape. For each listing it evaluates the filters in Table 1 — a mix of hard pass/fail filters (square footage, ceiling height, zoning compatibility, ABC-license geography, corridor proximity) and soft scoring filters (lease comp, parking, condition). Hard filters fail-stop. Soft filters contribute to a rank.

SourceTypeThreshold
Square footageListing recordHard≥ 9,500 sf, ≤ 18,000 sf
Ceiling heightListing record + photosHard≥ 16 ft clear
Zoning compatibilityCounty GIS portalHardCompatible with indoor recreation + on-premise alcohol
ABC-license geographyNC ABC permit map + church/school overlayHardDistance restrictions met
Lake Norman corridorGeocoded address + I-77 corridorHardWithin 30 min of corridor anchor
Lease compLoopNet + advisor pullSoft$/sf within local-market band
Parking ratioSite-plan + Street ViewSoft≥ 4 stalls per 1,000 sf
Roof + HVAC conditionListing record + photos + ageSoftCapex risk score
Table 1 — Filter criteria catalog for the MHG site search. Each row is a filter the agent evaluates against an inbound listing. 'Source' is where the agent retrieves the data; 'Type' is whether the filter is a hard pass/fail or a soft score that contributes to ranking.

Hard filters fail-stop; soft filters contribute to a numeric rank that the agent uses to order forwarded listings.

The agent forwards listings that pass all hard filters and exceed a soft-score threshold to the advisor, with a written rationale citing the specific zoning code, ABC-license clearance status, and proximity score. The advisor reviews the forwarded list and either accepts (warrants a site visit) or rejects (with a reason). Reject reasons feed back into the filter prompt the following week.

2.2 Paired comparison vs manual

We do not have the luxury of a true randomized comparison — there is one site search, one advisor, one CEO. Instead, the manual-baseline hours figures are the advisor’s tracked time before the agent pipeline went live (weeks -4 through 0) plus their estimate of the time each week’s listing volume would have consumed under the manual regime. The advisor tracks billable hours at 0.25-hour granularity, so the baseline is grounded in actual time logs, not retrospective estimation. The figure-3 false-positive curve is calculated against the advisor’s post-hoc per-listing verdict, which is the same verdict they would have applied under either regime.

2.3 Retrieval and grounding

Two of the filters (zoning compatibility, ABC-license geography) require pulling current data from external systems. The agent uses retrieval-augmented generation [4][5] against a curated index of NC county zoning ordinances and the NC ABC permit map. Hallucinations on these filters are operationally catastrophic — a forwarded listing that fails ABC because of a school within the distance band would burn 4-6 hours of advisor time. The pipeline therefore enforces a strict citation requirement: the agent must quote the specific zoning-ordinance section and the measured distance to the nearest restricted use, with a URL.

3. Results

3.1 Hours saved per week

Figure 1 decomposes hours-per-week consumption by activity. The bulk of the savings comes out of listing triage (12 → 2.5 hours) and zoning research (6 → 1 hour). Confirmed-shortlist time stays roughly constant, which matches the advisor’s value model: the advisor’s expertise is in confirmation, negotiation, and relationship — the agent does not, and should not, displace that work.

Figure 10.005.0010.0015.0020.0025.0030.00HOURS PER WEEK12.002.50Listing triage6.001.00Zoning research3.000.30ABC-license check3.003.20Confirmed-shortlistManualAgent-assisted
Figure 1 — Operator and advisor hours per week consumed by site-search activity, comparing the manual baseline against the agent-assisted pipeline. Hours saved are the difference; advisor confirmed-shortlist time stays roughly constant under both regimes (~3 hours per week) because the advisor's value-add is in confirmation and negotiation, not filtering. Most of the savings comes out of the listing-triage and zoning-research buckets.

3.2 Agent vs advisor agreement

Figure 2 is the aggregated confusion matrix across the full 8-week window (412 unique listings; duplicates collapsed by address and APN). The agent forwarded 175 listings; the advisor accepted 126 of those. The agent rejected 237 listings; the advisor would have accepted 9 of them on subsequent inspection (those 9 represent the recall cost of automation).

Figure 2PREDICTED CLASSACTUAL CLASSAdvisor: rejectAdvisor: acceptAdvisor: rejectAdvisor: accept228949126
Figure 2 — Confusion matrix of agent verdict ('forward to advisor' vs 'reject') against the advisor's verdict on the same listing once shown. Aggregated across all 8 weeks (n = 412 listings). The agent's recall (true positives among advisor-acceptable) is 96%; the agent's precision (advisor-acceptable among agent-forwarded) is 78% over the full window, climbing to 87% by week 8 (Figure 3). The 4% advisor-acceptable listings the agent rejected are the metric we monitor most carefully — it is the cost of automation.

Reading the matrix: precision = 126 / 175 = 72% (full window), climbing to 87% by week 8. Recall = 126 / 135 = 96%. F1 = 0.82. The advisor’s acceptance rate of agent-forwarded listings (precision) is the stat they care about least; the recall stat (the 9 missed acceptable listings) is the stat they monitor most carefully.

3.3 False-positive rate over time

Figure 3 plots the false-positive rate (advisor-rejected fraction of agent-forwarded listings) over the 8-week refinement window. The downward slope reflects calibration: each week the advisor’s reject reasons were captured as feedback and incorporated into the filter prompt, the zoning-rules table, or the soft-scoring weights the following week.

Figure 30%10%20%30%40%50%12345678FALSE-POSITIVE RATE (%)WEEK
Figure 3 — False-positive rate (agent forwards a listing the advisor rejects) over the 8-week filter-refinement window. Each week the advisor's reject reasons were captured as feedback and folded into the agent's filter prompt and zoning-rules table the following week. Recall stayed above 95% throughout — the advisor was unwilling to trade recall for precision. The downward slope reflects calibration, not improving model capability.

The most common reject reasons (from advisor week-by-week notes) were, in order: ceiling height not actually verifiable from listing photos despite agent assertion (week 1-3), zoning compatible per code but municipality has unwritten policy against the use (week 2-4), HVAC capacity insufficient for simulator-bay heat load (week 4-6), and parking ratio acceptable on paper but the available stalls are shared with an adjacent use (week 5-7). Each reject reason became a soft-score adjustment or a new filter in subsequent weeks.

3.4 Per-week metrics

ListingsForwardedAcceptedFP rateHours saved
Week 141231439%14
Week 248271833%16
Week 352251828%17
Week 457292224%18
Week 554241921%19
Week 650221818%20
Week 755211815%21
Week 855211813%22
Table 2 — Per-week metrics across the 8-week site-search window. Listings ingested = total inbound listings the agent saw. Agent forwarded = number passed to advisor. Advisor accepted = the advisor agreed the listing warranted a closer look. Hours saved is the advisor's estimate of manual-baseline time minus the agent-assisted week's time.

Listings ingested includes duplicates and stale relistings the agent silently dedupes. Hours-saved is week-over-week relative to advisor's recalled manual baseline; the advisor explicitly tracks this number.

4. Operator-time analysis

Aggregating across the 8-week window, the agent-assisted pipeline saved approximately 147 advisor-hours total (manual baseline ~24 hours/week × 8 weeks = 192 hours; agent-assisted actual ≈ 45 hours; difference 147 hours). At the advisor’s billable rate the pure-cost saving is meaningful, but the more important saving is calendar — the advisor surfaced a viable shortlist within four weeks rather than the six-to-eight weeks the manual pace was tracking toward.

On the operator side (the CEO, in this case), the agent pipeline cost roughly 6 hours of one-time setup (defining filters, wiring the GIS scraper, drafting the citation-required prompt) plus ~30 minutes per week of supervision (reviewing the agent’s rationale on edge cases, updating the rules table after advisor feedback). Net operator-and-advisor hours per week dropped from ~24 to ~7 by week 8.

The pipeline did not produce the eventual chosen site directly — that listing came in via the advisor’s network rather than the LoopNet feed — but the agent correctly scored it as a top-3 forward when it appeared, validating the rank logic. We treat that as a partial-credit outcome: the pipeline did not source the deal, but the pipeline’s ranking gave the operator confidence to move quickly when the advisor did surface it.

5. Limitations

  • One venture, one geography, one advisor. External validity is limited. The filters that matter for indoor-golf-plus-bar in NC will differ from filters for a coworking space in California, both in what counts as a hard fail and in how forgiving the ABC-licensing equivalent is.
  • Manual baseline is partly retrospective. The hours-saved figures rely on the advisor’s tracked time logs from before the pipeline (real) plus their estimate of what each week’s volume would have consumed under the manual regime (estimated). The figure-1 decomposition is therefore directionally correct but not perfectly grounded.
  • The agent did not source the eventual deal. The Denver site came in via advisor relationship, not the LoopNet feed. We discuss this in §4 — the pipeline ranked it correctly, but readers should not infer that this kind of automation displaces the advisor’s networking value.
  • Recall calculation is asymmetric. Listings the agent rejected are not all subsequently audited by the advisor — only the random 10% sample we audited each week. The 4% recall miss is therefore an estimate based on the audit sample, not a full-population census.
  • Calibration was active throughout. The week-8 false-positive rate of 13% is partly a function of 8 weeks of explicit prompt and rules-table refinement. An out-of-the-box pipeline with no calibration loop would likely sit at the week-1 rate of 39% indefinitely.
  • Hard-filter hallucinations remain the biggest tail risk. The citation-required prompt mitigates but does not eliminate the chance of a confidently-wrong zoning quote. A future iteration would replace the LLM’s zoning quote with a deterministic ordinance-API call.

6. Reproducibility

The SUMMIT agent definition, the filter catalog, the citation-required prompt, and the GIS scrape adapters are available on request under MIT license. Operators wishing to replicate need: (a) a working relationship with a CRE advisor willing to track time at 0.25-hour granularity for at least 4 baseline weeks; (b) access to a county-level GIS portal (most NC counties offer one free); (c) the relevant alcohol-licensing reference data for the operator’s state; (d) a listing feed (LoopNet, Crexi, or scraped MLS-equivalent).

For a comparable hours-saved measurement we recommend: a pre-pipeline baseline of at least 4 weeks; a pipeline run of at least 8 weeks (because most of the value comes from the calibration loop, not week 1); and a weekly random-sample audit of agent-rejected listings to track recall. Rolling out without the recall audit is the most common failure mode we’ve seen in adjacent operator-team replications.

References

  1. Geltner, D., Miller, N., Clayton, J., & Eichholtz, P. (2014). Commercial Real Estate Analysis and Investments, 3rd ed. OnCourse Learning. Chapter 4: Site selection.
  2. ICSC. (2023). U.S. Shopping-Center Classification and Characteristics. International Council of Shopping Centers.
  3. North Carolina Alcoholic Beverage Control Commission. (2024). Permit Application Handbook, Chapter 3: Eligibility and Distance Restrictions.
  4. Lewis, P., Perez, E., Piktus, A., et al. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. NeurIPS 2020.
  5. Gao, Y., Xiong, Y., Gao, X., et al. (2024). Retrieval-augmented generation for large language models: A survey. arXiv:2312.10997.
  6. U.S. Census Bureau. (2024). American Community Survey 5-Year Estimates. Lincoln + Catawba + Mecklenburg counties, NC.
  7. Fowler, M. (2003). Patterns of Enterprise Application Architecture. Addison-Wesley. Chapter 10: Service Layer.
  8. U.S. Small Business Administration. (2024). SBA 7(a) Loan Program Standard Operating Procedure (SOP 50 10 7.1).
  9. Mile High Golf. (2026, April). Concept Plan — Denver, NC (7521 Eastern Medical Dr). Internal artifact, MHG Business Plan.
  10. TruPath Labs. (2026). SUMMIT agent definition (MHG director specialist). Internal artifact.

Cite as: TruPath Labs Research (2026). Real-Estate Site-Search Automation: Hours-Saved and Filter-Accuracy in a Pre-Launch Venue Search. TruPath Labs Preprint TPL-2026-017. trupathventures.net/labs/research/site-search-automation-roi