No. The system prompt explicitly forbids quoting current rates, APR, or estimated payments. Those are licensed activities under state and federal rules. The bot can describe how jumbos generally price relative to conforming (sometimes higher, sometimes lower depending on the lender and product), but every specific rate question is routed to a state-licensed LO with NMLS attached. Your compliance officer can review the exact prompt language before launch.

It explains what documentation typically comes up: two years of vesting history for RSU continuation, K-1 schedules with year-over-year stability for self-employed, and bonus averaging conventions. It does not promise any particular file structure will work because that depends on the lender, the file, and the LO's review. The bot routes equity-comp and self-employed files to the LO most experienced with that income type, if your team page is set up to indicate specialisation.

If you publish the FHFA limits each year on your site, SleekAI reads them. Many brokers maintain a county-level limit page anyway; the bot pulls from that page so it never goes stale on the wrong year. For the rare visitor asking about a county not on the page, the bot defers to the LO rather than guessing. That keeps the bot's framing aligned with what the FHFA actually publishes.

Yes. Run a jumbo purchase bot on /jumbo/purchase, a non-QM bot on /non-qm, and a second-home jumbo bot on /jumbo/second-home, each with its own system prompt and knowledge sources. Multibot and 28+ display conditions handle the routing. Keeping prompts tight per product is the practical way to avoid a generic prompt drifting into territory the bot is not configured for.

It frames non-agency and portfolio products as 'lender-specific programs we may have access to', without naming specific lender partners unless your site already does. It captures the borrower's circumstances and routes to the LO. If your firm runs proprietary or bank portfolio products, the LO surfaces those during the consultation, since pricing and underwriting on portfolio loans is rarely something you want published publicly.

Claude Sonnet and GPT-4o both follow strict 'no rate quotes, no advice' system prompts reliably and produce consistent disclosure framing. Smaller or older models drift into advice-shaped statements when borrowers press, which is the exact failure mode a regulated mortgage site cannot afford. Bring your own key for OpenAI, Anthropic, Google, or OpenRouter and pick based on your firm's risk tolerance and budget.

Yes. Every conversation stores in WordPress with timestamp, page URL, model name, and token count. CSV export is available for state examiner requests, and the data never leaves your hosting except for the calls to your own API key. Most brokers archive transcripts to their compliance system on a recurring webhook so the conversation log lives in the same evidence trail as advertising and email.

The bot captures occupancy intent early because second-home and investment jumbo pricing and reserve expectations differ noticeably from primary. It explains the general framing (occupancy declarations, intended use, rental income treatment if it's an investment) and routes the file to the LO with that flag attached. Misrepresenting occupancy is a serious problem on jumbo files, so the bot frames the question carefully every time rather than assuming.