If you’re about to risk five figures on a tax-deed bid partly because a tool flagged the parcel, you need to know which part of that tool is a dumb rule, which part is a model, and which part is still you. There is a lot of loose talk about “AI-powered” real estate tools right now, most of it unhelpful and some of it dishonest. This post is an attempt to be precise about what is actually doing the work inside LotBrief, because the honest version is more useful than the marketing version.
The short answer: the part that shrinks a 900-parcel list down to a shortlist is not AI at all. It’s six deterministic rules. The part where AI genuinely earns its place is the research and the plain-English reasoning — the work a human analyst would otherwise do by hand. And the part that decides whether to bid is still you.
The real problem
A California county publishes a tax-defaulted property auction list as a PDF, usually 90 days before the sale, usually 900 to 1,000 parcels long. The Riverside County TC-223 sale that closed on April 28, 2026 was 946 parcels in the version we ingested.
Most of those parcels are not worth a serious buyer’s attention. The trouble is that “most” is not “all,” and the cost of getting it wrong is asymmetric. Bid on a paper lot and you’ve tied up money in something un-developable that takes a decade to resell. Bid into a bulk speculator’s failed desert-lot cluster and you’ve bought someone else’s bad thesis at retail. Miss an HOA-dues trap or a commercial parcel with an environmental history and a $50,000 bid becomes a much larger loss. The list is mostly junk, but the junk is interleaved with the few real opportunities, and a beginner reading 946 rows in sequence will miss at least one trap on at least a few rows.
So the job splits naturally into three layers: throw out the obvious junk, research the survivors, and decide. Different layers want different tools.
What deterministic rules do well: the mechanical reduction
The first layer is pure mechanical filtering, and it should be boring, fast, and reproducible. LotBrief applies six rule-based filters — string matches, numeric thresholds, and blocklist lookups. No model, no inference, no “AI.” Every dropped parcel records exactly one reason, so the arithmetic is auditable.
Here is what actually happened on Riverside’s 946-parcel TC-223 sale: six rules dropped 771 parcels and kept 175. The full per-filter math — which filter dropped what, and why — is in the Riverside retrospective.
The honest punchline: one filter — “does this parcel have a street address?” — does 687 of the 771 drops, roughly 89% of the entire reduction. It is the least sophisticated rule imaginable, a single field check, and it is by far the most valuable one.
Anyone selling you a neural network to find paper lots is selling you a neural network to read a field that already says SITUS ADDRESS: NONE.
That is the whole point of keeping this layer deterministic. The rules are transparent: you can read all six in plain English on the methodology page, you can see exactly why any given parcel was dropped, and you can re-run the list next year and get the same answer for the same input. A model would make this layer slower, more expensive, and harder to audit, in exchange for nothing — these are not judgment calls, they are field checks and threshold comparisons.
The 175 survivors had minimum bids ranging from $10,106 to $195,409, with a median of $25,082 and an average of $35,985. That’s a tractable set a human can actually work through.
One caveat I want to state plainly, because it matters for trust: the AVM-ratio filter (which would drop parcels whose opening bid is already too high a fraction of estimated value) shows 0 drops because it had almost no data to act on. A filter that needs a property valuation can only fire on parcels that have one, and at the time of this sale we had run valuation enrichment on only a small test sample, not the full surviving set. So “0” here does not mean every survivor passed a value check — it means the check mostly had nothing to evaluate yet. Running enrichment across all 175 survivors is the next step, not a finished result, and I’d rather tell you that than imply a clean valuation pass that didn’t happen.
What AI genuinely adds: the analyst layer
If the rules handle the mechanical reduction, what’s left for AI? The part that used to require a human analyst with a browser open and an afternoon to spare.
Due-diligence research. Before you can filter a list, somebody has to find and validate the list, and that turns out to be the messy, human-time-intensive part of this whole exercise. Is the sale even still on, or was it postponed? Where does the official parcel list actually live this year — and did the county quietly migrate its auction from one platform to another between sales, leaving last year’s link dead? What’s the redemption deadline, the deposit requirement, the exact sale window? Are there zoning, special-assessment, or environmental records that color a given parcel before you ever pull a title report? This is reading, cross-referencing, and reconciling dozens of inconsistent sources — county treasurer pages, auction platforms, assessor GIS, calendars that contradict each other. It’s exactly the kind of unglamorous synthesis a language model does well and a human finds tedious. AI here is a research assistant that reads everything and surfaces what it found, with sources, so you can check it.
Per-parcel plain-English reasoning. The deterministic layer can tell you a parcel survived and attach risk flags — owner is an LLC, TRA suffix suggests a special-assessment overlay, address pattern looks like a condo unit. But a stack of flags is not an explanation. The genuinely useful AI step is turning those raw signals into a short, readable “here’s why this parcel is — or isn’t — worth a closer look, and here’s the specific thing to check next.” That is assistive reasoning: it organizes and explains the signals the rules produced. It does not generate the signals, and it does not reach a verdict.
In both cases the model is doing analyst work, not oracle work. It reads, it cross-checks, it explains. Every claim it surfaces should be traceable to a source you can open yourself.
Where AI does not belong
It is just as important to be clear about the boundaries.
AI does not make the buy/no-buy call. The output is “worth a closer look” or “probably skip, here’s why” — never “bid.” The decision is a judgment about your capital, your risk tolerance, and your local knowledge, and a model has none of those.
AI does not replace a title search or a title policy. No amount of model reasoning tells you what liens survive the sale on a specific parcel. In some cases — for example, when a tax sale divested a junior federal tax lien and the IRS received the required advance notice — the United States can hold a post-sale right to redeem the property (a 120-day federal floor under 26 U.S.C. § 7425(d)(1)); in other cases a federal lien simply survives the sale untouched. A tax sale extinguishes HOA back-dues liens, but the recorded CC&Rs carry through — so the buyer owes dues going forward — and California’s special-assessment districts — Mello-Roos community-facilities taxes, bonded assessments — can survive the sale and keep attaching to the land, to the extent the sale proceeds don’t clear them. Which of these applies to a specific parcel is exactly what you confirm with a title professional, not a chatbot.
AI does not guarantee data accuracy. A third-party automated valuation is a model someone else trained; it is an estimate, frequently wrong on atypical properties, and we don’t own it or vouch for it. Treat every enriched number as a starting hypothesis to verify, not a fact.
And AI does not predict what will sell or what the final hammer price will be. We don’t claim it does. Anyone who does is guessing with extra confidence.
Putting it together
The division of labor that actually holds up is unglamorous. Deterministic rules do the mechanical reduction — 946 to 175 — fast, transparent, and reproducible, with one trivial field check doing nearly 90% of the work. AI does the analyst layer: the research that finds and validates the sale and the context around each parcel, and the plain-English reasoning that explains the signals. The human does the part that is irreducibly human — the decision.
If a tool blurs those lines, be suspicious. The honest framing is the smaller claim, and the smaller claim is the true one: we surface signals and explain them, you do your real diligence, and you decide.
If you want to see exactly what the analyst layer produces — the per-parcel write-ups, the sources, the things we couldn’t confirm — the free sample brief is the artifact itself, complete and free. For each upcoming California sale we publish the same work in full, free, before bidding opens. Leave your email and the newsletter sends one message when the next county list drops — nothing else.
This is a case study and an explainer, not legal, tax, or investment advice. We are not a licensed real estate broker, agent, or attorney in any jurisdiction, and nothing here is a recommendation to bid on any property. Automated valuations and research summaries are estimates that can be wrong; verify everything — title, liens, zoning, occupancy, and environmental history — with the appropriate professionals before bidding. We surface signals. The decision is yours.