Deterministic vs. Probabilistic Real Estate AI: Why Math Errors Should Scare You
Most real estate AI tools guess at the numbers and sound confident doing it. Here's how to tell the ones that compute from the ones that improvise — before a wrong figure ends up in front of your client.
The New Buying Criterion: Getting the Money Math Right, Every Time
There's a quiet shift happening in how agents evaluate AI tools. The first wave of questions was about speed and tone — does it answer fast, does it sound human. The question that actually matters is narrower and harder: when this thing produces a number, is it right every single time, or only most of the time?
That distinction is the whole game. A talk track that's 90% as good as a great one is still useful. A net sheet that's right 90% of the time is a liability, because the 10% where it's wrong is the time you quoted it to a seller at the kitchen table. In real estate, the money math is the part where 'usually correct' is indistinguishable from broken.
So before you judge a real estate AI tool on its writing or its interface, judge it on whether its math is computed or generated. Those are not the same thing, and the gap between them is where deals get lost.
Where General-Purpose LLMs Go Wrong on Financial Math
Large language models are probabilistic by design. They predict the next most-likely token, which is exactly what you want for drafting an objection rebuttal and exactly what you don't want for a graduated tax calculation. The model isn't doing arithmetic; it's pattern-matching what an answer of that shape usually looks like. Most of the time the pattern is close. Sometimes it's confidently, fluently wrong.
Washington gives an LLM three specific ways to embarrass you. First, graduated REET: the state rate steps from 1.10% on the portion up to $525,000, to 1.28% from $525,001 to $1,525,000, to 2.75% up to $3,025,000, and 3.00% above that — plus a local share of 0.50% in King, Pierce, and Snohomish (0.25% in counties like Island, Mason, Lewis, and Cowlitz). A model that flat-rates the top bracket across the whole sale price overstates the tax by thousands on a million-dollar deal. Second, net sheets stack a chain of figures — commission, excise, title, escrow, prorations — and a single fumbled line propagates to the bottom number the seller actually reads. Third, business-day deadlines: under RCW 1.16.050, a period of five days or fewer skips weekends and Washington holidays, while a period longer than five days runs on calendar days. An LLM counting straight calendar days through a holiday weekend will hand you the wrong title-review date — and that's the date a missed contingency turns on.
None of these are edge cases. They're the everyday math of the job. And a wrong number in front of a client doesn't just cost you that transaction; it costs you the referral, because the moment a seller catches one bad figure, every number you've ever given them is suspect.
Code, Not Model: The Deterministic Approach
The fix is not a smarter prompt or a bigger model. It's refusing to let the model do the math at all. A deterministic system hands every financial calculation to plain, tested code — the REET brackets live in a function, the net sheet is an arithmetic pipeline, and the NWMLS deadlines run through a Washington business-day calendar that knows which holidays to skip. Same inputs, same output, every time, with no probability involved.
This is the line SENTINEL draws on purpose: the AI handles language — reading your question, pulling the right record, writing the explanation — and code handles arithmetic. REET, net sheets, commission splits, and the Form 35 inspection, Form 22A financing, Form 22T title, and Form 17 disclosure deadline math never touch the language model. That's why the same $1.525M Snohomish sale returns the same excise figure on Monday and on Friday, and why a five-business-day title window lands on the correct date through a long weekend instead of three days early.
The trade-off is honesty. A deterministic tool will tell you it doesn't have the data rather than invent a plausible-looking number to fill the silence. That feels less magical in the moment and is worth far more on a live deal, because the failure mode is 'I don't know' instead of 'here's a confident wrong answer.' You can see the bracket tables and the calendar logic on the methodology page if you want to know exactly how a figure is built.
— Not tax advice. Verify with a WA-licensed CPA. Not legal advice on form deadlines; verify with your broker or a WA-licensed attorney.
Cites-Every-Claim Is a Trust Contract, Not a Tagline
Determinism handles the math. Sourcing handles everything else. The second half of real estate AI accuracy is whether the tool can show its work — whether every property value, market stat, and form deadline comes back attached to where it came from, not lifted from the model's training data.
Treat 'cites every claim' as a contract you can enforce, not a marketing line. A property figure should trace to a county assessor record. A market number should trace to a current data source you can re-pull, not a number the model remembers from some indeterminate point in the past. A deadline should trace to the NWMLS form and the day-counting rule behind it. If a tool can't tell you where a number came from, you have no way to verify it before you repeat it to a client — and repeating an unverifiable number is exactly how a fabrication becomes your problem.
This is also why you should be suspicious of any market figure an AI states as current fact without a date or a link. Markets move week to week in King, Pierce, and Snohomish. A confidently quoted median with no provenance is a guess wearing a suit — always re-pull current data before you repeat a number to a client.
A Checklist to Grade Any AI Tool Before You Trust It on a Deal
Run any real estate AI tool through these six questions before you let it near a client conversation. First, does it compute financial math with code or generate it with the model? Ask the vendor directly — the answer tells you whether REET and net sheets are reliable or merely usually-right. Second, can it cite the source of every property and market number, with something you can re-pull? Third, does it handle graduated REET correctly — feed it a $1.6M sale and check whether it applies the brackets or flat-rates the top tier.
Fourth, does it get Washington business-day math right — give it a five-day title window starting before a holiday weekend and confirm the deadline skips the weekend and the holiday. Fifth, does it say 'I don't know' when it lacks data, or does it always produce an answer? A tool that never admits a gap is a tool that fabricates. Sixth, can you verify a figure independently in under a minute — is the math transparent enough to check, or a black box you're asked to trust?
A simple way to test the first three at once: run the same sale price through the tool and through a trusted standalone REET calculator and see if the excise matches to the dollar. If it doesn't, you've found your answer about whether the math is computed or improvised — and you've found it on a test case instead of on a real deal.
— Not legal advice. Verify with your broker or a WA-licensed attorney.
See the assistant work · no login
Bring Sentinel the real question behind this article.
Start with a prepared Western Washington example, then replace it with the work on your desk. Sentinel keeps the geography, source dates, assumptions, and verification boundaries visible.
Build and inspect the question publicly. Starting a private workspace afterward is optional.
Frequently asked
- Why are LLMs bad at financial math?
- Language models predict the most-likely next token rather than performing arithmetic, so they pattern-match what an answer should look like instead of computing it. That works for prose but fails on multi-step or graduated calculations like Washington's REET brackets or a stacked net sheet, where being close isn't the same as being correct. The reliable fix is to route the math to deterministic code and let the model handle only language.
- What does deterministic AI mean in real estate?
- Deterministic means the same inputs always produce the same output through fixed, tested code rather than a probabilistic model. In practice, financial calculations — REET, net sheets, commission splits, and NWMLS deadline math — run on plain arithmetic and a Washington business-day calendar, not on the language model. The AI is still used for reading your question and writing the explanation, but it never does the math.
- How do I know if a real estate AI tool is accurate?
- Check three things. Ask whether financial math is computed in code or generated by the model. Confirm every property and market number cites a source you can re-pull. And test it: run a graduated REET case and a five-day deadline that crosses a holiday weekend, then verify the answers independently. A tool that admits when it lacks data is far safer than one that always produces a confident answer.
- Can ChatGPT calculate Washington REET correctly?
- Not reliably. Washington REET is graduated — the state rate steps across price bands and a county-specific local share is added on top — and a general-purpose chatbot frequently flat-rates the top bracket or skips the local portion, which can be off by thousands on a high-value sale. Always verify any excise figure with a deterministic tool or a WA-licensed CPA before quoting it to a client. — Not tax advice. Verify with a WA-licensed CPA.
- What is the risk of an AI giving a wrong number to a client?
- The immediate risk is a blown net sheet or a missed contingency deadline, but the lasting damage is to trust. The moment a client catches one fabricated figure, every number you've given them becomes suspect — and that costs the relationship and the referral, not just the transaction. That's why the question that matters is whether the tool computes its math and cites its sources, not how fast or fluent it sounds.