Vida's Mobi answered over 21,000 broker queries in roughly five months from launch. About 200 a working day, at one lender. Which settles whether brokers will use these things, and moves the argument onto whether the answers are right. Mortgage Solutions, Dec 2025.

The Mills Review landed 6 July. Its read on AI in retail financial services: adoption depends on confidence in explainability, fairness, resilience and accountability, with opaque decision-making named as a core risk. Four conditions, none of which a model gets to self-report.

Our agent refused a "what's the best deal" question correctly, and used the phrase "best deal" inside the refusal. Scored it a fail. The spec now forbids repeating the phrase even to reject it. Golden run: 17 of 18. That was the one.

Veris reads a bank statement, finds the figure, records the line it came from. We've decided the raw file goes after that. Storage limitation says don't keep it, the audit trail says prove it, and keeping the located line does both. Written down before it's built.

A field in our fact find carries a source flag: read from the document, or filled in by the model. The default was "read from the document". So anything the code forgot to stamp claimed provenance it never had. Default's flipped. Unstamped now means unverified.

When Lendmind Tenet can't locate a quote's exact box on the page, it opens the page anyway. Unhighlighted. The alternative was highlighting something close. A box in roughly the right place is worse than no box, because the reader trusts it.

The trade press keeps asking who's accountable when a lender's own chatbot gives a broker the wrong criteria answer. Fair question. Under Consumer Duty someone owns that answer. Hard to own one you can't trace back to the document it came from.

4,144 buy to let products on the market in June 2025. Highest since Moneyfacts started counting in 2011. Every one of them carries criteria. Every bit of criteria is a question a broker might have to ask someone. The product count is a proxy for the phone ringing.

Forward an email, get back a forensic report. No portal, no login, nothing new to learn. Every figure in it is boxed on the page it came from, so you can check the machine instead of taking its word for it. That's Veris.

We used to cite "Lending Criteria, page 12". Which isn't much use if you're the broker who then has to go and read page 12. Every figure now points at the exact box on the page it came from. Number and proof in one view. Should have built it that way the first time.

Worth watching what lenders fund, not what they say about AI. Paragon: a two-year origination platform. OSB: 18 months building for same-day offers. Shawbrook: AVMs inside the AIP. HTB and Pepper: more people on the phones. Every one of those is about getting to yes faster.

Superlative questions break retrieval. "Which product has the highest LTV" has no most-relevant chunk. No chunk knows it is the maximum. So we hard-routed them. Scan the criteria set, order by the field, never retrieve. Obvious in hindsight. Cost us an eval cycle to see.

The most useful answer a criteria tool can give is: this lender does not do that. A clean no, with the line that rules it out, saves days. Most systems are built to always return something. We built Tenet to say when a case is not covered.

Sit with a specialist BDM for an hour and you hear the same criteria questions on repeat: day-rate contractors, recent CCJs, gifted deposits. Tenet answers those from the lender's own criteria, line cited, so a BDM's time goes to the cases that actually need judgement.

A dozen lenders repriced this week. Which means a stack of the rate and product sheets sitting in brokers' inboxes went out of date overnight. The hard part of specialist lending was never finding the answer. It's trusting the version in front of you is still live.

A P60 that does not reconcile with the payslips is one of the most common things Veris flags. Sometimes it is a job change. Sometimes a bonus. Sometimes neither. Cross-document reconciliation catches in seconds what takes an adviser an evening.

Every specialist lender publishes its lending criteria. Yet checking whether an awkward case fits still means ringing a BDM or scrolling a 40-page PDF. Tenet reads that lender's own criteria and answers the case directly, citing the exact line it came from.

Fraud rarely lives in one document. The payslip is clean, the bank statement is clean, the ID is clean. It's when you line them up that the story stops matching: a salary that never lands, an address that doesn't exist, dates that can't both be true.

The fastest way to lose an adviser's trust is an AI that guesses. Tenet reasons only over your published products and criteria, never the open web. When something is not covered, it says so. lendmind.com/#/tenet

A doctored payslip can look perfect to the human eye. It rarely survives a forensic pass: font anomalies, baseline drift, metadata tells. Veris checks every document and gives each a verdict. No portal. Forward an email, get a report. lendmind.com/#/veris