Astra for Law shows how to turn a strong model into an AI product for regulated markets: combine general intelligence with proprietary data, compliance and the workflow customers already use. For anyone working on credit products, that choice says more than any isolated model benchmark.

Today brought one AI launch after another. But one of them made me stop and think more than the others because it connects directly to the product work I do every day in credit.

What Astra for Law shows about products in regulated markets

Whenever I see generic AI become a vertical product in a regulated market, I think about how closely that resembles what we deal with in credit.

This week, OpenAI launched Astra for Law. It is not just another legal chatbot. It combines the Astra model with a database of more than 230 million US decisions and laws, zero data retention and ready-made plugins for systems law firms already use every day, such as Relativity, Clio and iManage.

Access starts out restricted through a partner program, and API integrations are already available with major names in the legal sector, such as Harvey and Legora.

To see the full launch, read the Astra for Law announcement.

The difference between a feature and a vertical product

What catches my attention is not the model itself. It is the product choice.

They did not try to sell “better generic AI.” They packaged a strong LLM with proprietary data, built-in compliance and integration with an existing workflow. That is what separates a feature from a product that can truly survive in a regulated market.

This combination also changes the conversation about AI governance. Governance is not just an approval layer added after the model is ready; it appears in how the product stores data, sets boundaries and connects to the work of the people who use it.

What this teaches credit product teams

It speaks directly to anyone working on the product side of credit, as I do. Receivables, origination data and operating history: that is the real asset. The AI model is the component already available on the shelf.

A product team’s job is to combine those assets with an experience that solves a concrete problem without losing control over sensitive information. An AI product management approach helps treat capability, context, accountability and measurement as product decisions rather than details left for after launch.

The same applies to the level of control. 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.

The bar is rising

The bar is rising quickly for any sector with sensitive data and strong regulatory requirements. Credit, law and healthcare will all face the same question: who can combine general-purpose intelligence with real governance and proprietary data?

I like following this kind of movement because it shows that the game is not only about having the biggest model. It is about where and how you apply it.

In brief

  • A strong model alone does not make a vertical product.
  • Proprietary data and compliance can be central parts of the value proposition.
  • Integrations with tools customers already use reduce the distance between a demo and real operations.
  • In credit and other regulated markets, the advantage lies in combining AI capability with governance and proprietary context.

The rest of the radar

Bonsai 2 27B — Smaller, cheaper models expand what can run at low cost, creating room for new features without breaking the inference budget. Read more

Qwen 3.8 Omni Flash — Another strong multimodal alternative outside the OpenAI/Google/Anthropic axis, putting pressure on prices and expanding supplier options. Read more

Code Scans (Devin) — Agents are moving beyond writing code on command to discovering what needs to change, expanding the scope of engineering-automation products. Read more

Skillsync (YC W26) — Tackles context lock-in between AI assistants, a real problem for anyone orchestrating several agents in the same workflow. Read more

How GLM built its own inference infrastructure — The trade-off between serving your own model and using a third-party provider directly affects the final product’s cost, latency and SLA. Read more

Quantized Qwen 3.8 27B (13.1 GB VRAM) — Quantization lowers the cost barrier to running a frontier model locally, opening room for edge and privacy-sensitive scenarios. Read more

The most important product decision is what you do not build — A reminder to keep scope disciplined when it is easy to pile on “one more AI feature” simply because it is now possible. Read more

How to write with an LLM — Shows what advanced users already expect from a strong human-AI collaboration flow, useful in daily work and as a UX reference. Read more

Towards self-driving codebases — Maps the shift from “AI that writes code on command” to “AI that takes care of code on its own,” useful for positioning a developer-tools roadmap. Read more

That is all for today. Until the next edition.