What actually happens between "ask" and "answer"
After almost every demo, someone asks the same thing: "Okay, but how does it actually work?"
It's a fair question. "AI" covers everything from a chatbot that makes things up with confidence to tools that won't say a word they can't back up. If you're a broker deciding whether to put something in front of your agents, you should know which kind you're looking at.
So here's the plumbing, no jargon, no hand-waving.
The short version: find first, then write
Most general-purpose AI tools work like a very well-read person answering from memory. They've read a lot, they're fluent, and they're usually close. "Usually close" is fine for a dinner recipe. It isn't fine for a contract deadline.
reTEQ works the other way around. Before the AI writes a single word, it finds the specific sources that answer the question; the statute, the line of the AAR form, your brokerage's policy, your designated broker's guidance. Then it writes the answer from those sources only, and shows you where each piece came from.
The technical name is retrieval-augmented generation. The plain-English name is look it up, then answer. The order is the point.
Step 1: The library
Everything reTEQ can say comes from a defined library:
Arizona statutes: Titles 32 and 33, the licensing and property law agents work under.
The Arizona Administrative Code: the Department of Real Estate rules.
The AAR form library: purchase contracts, addenda, disclosures, advisories. Where a form has line numbers, answers point at the exact line.
Your brokerage's own policies: kept separate for each brokerage. One brokerage's floor-time rules never show up in another brokerage's answers.
Designated Broker Q&A: answers written by a real designated broker, in their own words, for the questions agents actually ask.
If something isn't in the library, reTEQ doesn't pretend it is.
Step 2: Finding the right pages
When an agent asks "How many days does the seller have to deliver the SPDS?", reTEQ searches the library two ways at once: by meaning (questions about "disclosure timing" should find text about delivery deadlines even if the words differ) and by exact words ("SPDS," "three days").
That search pulls back about a dozen candidate passages. Then a few rules kick in:
If you name something, you get it. Ask about a specific statute, a specific form, or "line 39 of the purchase contract," and that exact source is pinned to the top. Naming the source is the strongest signal there is.
Your designated broker outranks the generic text. If a Designated Broker Q&A answers this exact question, it moves ahead of the general form pages. That's deliberate: a broker's answer to this question beats a page that merely mentions the same form.
Then the rest gets ranked for how directly it answers the question, and the best five go forward.
Step 3: Writing the answer, with rules
Only then does the AI write. It gets the question, the five sources, and a set of rules it has to follow. A few of them:
Answer from the sources. Quote the form language where it matters, and cite the statute, form and line.
If the sources don't cover it, say so. "I don't have source material on that, check with your broker" is a valid answer. A confident guess is not.
No "typical" commission rates. Ever. Compensation is negotiable and has to be written into the agreement; the AI explains that instead of quoting a number.
No legal advice. Legal interpretation gets referred to an attorney.
When the answer appears, the sources come with it. Click a citation and the actual form opens with the cited lines highlighted. The agent doesn't have to trust the AI, they can check it in one click.
Step 4: Keeping it honest over time
A system like this is only as good as its library and its testing, so we test it the boring way.
We keep a bench of real agent questions for each major form, the purchase contract, the buyer-broker agreement, the seller disclosure statement, with the right answer for each one labeled by a designated broker. Every time we change how retrieval works, or add new Q&A, we re-run the bench and compare.
That testing earns its keep. In the last few weeks it caught a ranking rule that was letting generic form pages crowd out a broker's answer, and it caught two expert answers that disagreed with each other on a disclosure deadline. Both were caught by the bench and fixed the same day — which is exactly why it exists.
Why the order matters
Here's the difference in one example.
Ask a general AI tool when the SPDS is due, and you'll get a fluent answer that may or may not match the current contract.
Ask reTEQ, and the answer is built from the purchase contract's own language, the seller delivers the SPDS within three days after contract acceptance (RRPC lines 154–155), with the contract open beside the answer and the line highlighted.
Same question. One answer you have to trust. One you can verify.
That's the whole trick, and it isn't a trick: find the source first, then answer from it, and show your work. For a brokerage, that's the difference between an AI your agents use and an AI you can actually stand behind.
