Today’s radar had everything from an AI that fits in 8 MB to one that is already leading part of its creator’s own R&D. In the middle of it, one story made me stop and think about how we prove that AI can be trusted in sectors like mine.

Whenever someone asks me whether AI can be trusted inside a bank or fintech, my answer is always the same: it depends on how that trust is built.

Auditing enters the development cycle

A story published this week illustrates the point well. Anthropic partnered with Accenture to place “embedded” evaluators inside its own model-development process. These professionals have employee-level access, observe from within how models are trained, test safeguards and check whether security commitments are being met in practice. The investment is in the billions and the partnership has a five-year horizon.

For the full announcement, read the story about the Anthropic and Accenture partnership.

This caught my attention because it is exactly the kind of structure that people working in regulated sectors — credit, payments and sensitive data — know needs to exist, but rarely see treated so explicitly by the companies building the models.

What this changes for products in regulated sectors

In my day-to-day product work, the question that comes up most when discussing AI applied to financial processes is not “does the model work?”. It is “how do I audit this, how do I explain it to the regulator, how do I explain it to the customer?”.

AI governance is not a brake on innovation. It is what allows innovation to reach the market without becoming a problem later. That means making the system’s decisions, owners, evidence and boundaries visible.

What I find most relevant about this move is treating security and compliance as part of the product from the start, rather than as a layer added at the end. It is the opposite of the “launch fast and fix it later” approach we see in so many AI releases today.

Auditing needs to become a product decision

An AI product management approach helps treat capability, context, accountability and measurement as product decisions rather than details left for after launch.

In a regulated workflow, an AI risk matrix helps define which data controls, integrations, evaluations and oversight are needed in each part of the product. More structured evaluation also helps turn “can we trust it?” into observable criteria before and after launch; the AI agent evaluation guide helps organize that reasoning.

I am convinced this is the direction the entire market will take, especially companies that handle money, credit and sensitive data. Anyone building AI products in finance will need to think about auditing and governance from day zero, not as an appendix.

In brief

  • Evaluation and security need to participate in development, with access to evidence from the process.
  • In credit, payments and sensitive-data workflows, trust needs to be explainable to regulators and customers.
  • Compliance is not a later layer: it can be part of the product’s value proposition and operation.
  • Governance from day zero reduces the risk of turning a fast launch into a problem that is difficult to fix.

The rest of the radar

Cactus Needle 3 — Models measured in a few MB enable 100% on-device AI, without API cost or cloud dependence. Read more

Alibaba open-sources a medical diagnosis model — An open model raises the quality bar in regulated niches and pressures proprietary products to justify their value. Read more

Stepfun Step 5 enters the cost-versus-intelligence frontier — Another cheap and competitive option raises the bar for price negotiations with established LLM providers. Read more

Gemini breaks into three companies in a security test — Shows how far autonomous agents are already advancing in offensive tasks and reinforces the need for guardrails before exposing agents to real systems. Read more

Claude Code starts reading AGENTS.md — Signals convergence around an instruction-file standard across agent tools, reducing integration friction. Read more

Stagehand gets faster and cheaper — Token cost and latency are now the main constraints on web-browsing agents, and gains here make production use cases more viable. Read more

Scry tests dynamic pricing for agents — Congestion-based pricing could inspire how AI products charge for heavy infrastructure usage. Read more

A specialized agent makes Mac automation cheaper — Shows that cheaper specialized agents can replace expensive general-purpose models for many automation tasks. Read more

Claude now leads 26% of Anthropic’s R&D — A measure of how quickly the AI product-creation cycle itself is being automated. Read more

It was a full radar day — I separated out what was worth stopping to read.