AI leaves the demo room

AI is moving into bank fraud teams, muddy job sites and the argument over who should own the model. Monday's issue is about where the useful work now sits.

Singapore's financial regulator is testing AI models that combine cross-bank and public-private data to spot suspicious accounts and transactions. Banks and law enforcement are involved, with findings due by the end of 2026.

The Monetary Authority of Singapore is also widening access. Its Pathfin.ai programme now has more than 300 participants matching smaller financial firms with checked AI tools. Proposed risk guidelines cover the whole AI lifecycle, while its agent framework calls for identity checks, checks before an action is executed and an audit trail.

Founders context: Shared data can catch patterns that one company misses, but nobody wants a clever fraud model quietly blocking real customers. If you run AI across payments, claims or accounts, name who can approve an action, log what the model saw and keep a route for a human to reverse it.

Fergus has added AI assistants to the job software used by more than 26,000 tradespeople across Australia, New Zealand and the UK. The web and mobile tools use existing customer, job, cost, quote and invoice data.

A tradesperson can dictate site notes, turn them into a quote or invoice, ask why a job is losing money, or request a report without moving the data into a separate assistant. Fergus Assistant is available now. A wider Fergus AI rollout is due this week. The launch and claimed time savings come from Fergus, so the proof still needs real usage data.

Founders context: Measure the dead time between finishing a job and sending the invoice. Test whether voice notes can fill the job record and produce a draft invoice before the van leaves the site. Check the margin answer against the books. Faster admin only counts when the numbers are right and payment arrives sooner.

The case for private, fine-tuned AI is running into a moving target. Wharton professor Ethan Mollick argues that keeping a small specialist model ahead of fast-improving general models is expensive and difficult. He sees specialist models as complements to frontier models rather than broad replacements.

A private model can still earn its keep when the data cannot leave your environment, response time matters, or a narrow task needs consistent behaviour. The cost sits beyond the first training run. Budget for refreshes, regular comparisons with general models, reliable serving and people who understand the stack.

Founders context: Price the maintenance before calling the model an asset. Run the same real workflow through your specialist model and two current hosted models every month. Keep the private model for the jobs where privacy, speed or accuracy wins. Route broad reasoning elsewhere.

🧯 OpenAI says it now writes safety cases before some frontier training runs expected to raise capability sharply. It is a new pre-run documentation step alongside the auditor plan covered yesterday.

🔑 A retired agent setup may still have leaked a live key months after its servers were shut down, according to Pieter Levels. He deleted the separate account after a billing alert. The cause is unverified. Shutting down an agent should always include revoking its credentials.

🇰🇷 Artificial Analysis has joined Korea's sovereign-AI programme as an official evaluator. Its index carried 25% of the latest competition score, according to the evaluator.

↩️ Simon Willison released commit-rewriter for cleaning agent-heavy Git history. It creates a backup branch before rewriting commits, which is how tools that edit real records should behave.

🗝️ Vals AI says Claude Fable 5.1 solved a 370-year-old cipher in one long-running test. One result proves little. It is enough reason to retest one neglected document set with expert review.

📈 AI Daily Brief argues for bigger bets from small teams using AI. Saving ten minutes is fine. Shipping work you could not afford last year is the better test.