Today’s radar brought the usual: a model launch, a delayed launch, and another AI feature rolled back over privacy. But the item that held my attention most was not about parameters or benchmarks. It was about how to organize work when part of it is done by an agent.
There is a question that comes up whenever I discuss automation on a credit roadmap: does this become a standalone AI project, or does it become part of the normal way of working? In practice, most people choose the first answer. I think that is where the mistake lies.
Forrester published an article this week proposing what it calls a “cognitive operating model.” The core idea is easy to explain and hard to apply: instead of treating an AI agent as a “digital employee” that gets its own seat on the org chart, treat each cognitive skill as the unit of work. Reading a contract, validating data, producing a summary, deciding a next step. Each of these skills can be performed by a person, by an agent, or by both in sequence, depending on the complexity and risk involved.
This connects directly with what I see every day in structured credit and receivables products. There is a large number of cognitive tasks spread across the operation: checking the collateral for an invoice, preparing a calculation memo, reviewing a covenant, summarizing a portfolio for a committee. Today, much of this is still treated as either manual work or a separate automation project, with its own scope, sprint, and dedicated team.
Forrester’s model suggests a different, more useful product question: for each of these skills, who should perform it now, a human or an agent, and what changes if that allocation is dynamic rather than fixed? It shifts the work from “writing an automation PRD” to “designing the governance of who does what, and when that can change.”
It is not about replacing people. It is about stopping the treatment of automation as a separate chapter of the product and starting to design the entire flow, with the right skill in the right place, whether human or not.
I like this kind of provocation because it takes AI out of the realm of the “special project” and puts it where it really creates value: inside the routine, resolving the right bottleneck, without needing an announcement to do it.
I have included Forrester’s full analysis for anyone who wants to go deeper: Read more
The rest of the radar
OpenAI releases GPT-5.6 to the public after federal testing — shows that frontier models face government scrutiny before launch, making availability an operational risk. Read more
Google delays Gemini 3.5 Pro until July 17 — the competition among major labs is shifting from chat benchmarks to efficiency in long-running agent tasks and cost per token. Read more
Mistral launches Robostral Navigate, robotic navigation with a single camera — lowers the hardware barrier to building autonomous-navigation products in logistics and robotics. Read more
Meta disables its AI image feature days after launch — shows the cost of releasing AI features with opt-out rather than opt-in for user data. Read more
Migrating a production agent to GPT-5.6: 2.2x faster, 27% cheaper — exposes the real pitfalls of changing models in production before comparing the full cost-benefit trade-off. Read more
Claude Code sends 4.7x more tokens than OpenCode before reading the prompt — shows that the choice of harness, not only the model, directly affects the cost and speed of AI coding tools. Read more
What xAI’s Grok Build CLI actually sends to xAI — a warning about privacy and security due diligence before adopting third-party AI CLIs across engineering teams. Read more
“Harness engineering”: the agent is the model plus everything you build around it — points to where to invest engineering effort in agent products: the scaffolding, not only model selection. Read more
“Six months to live” for open models, analyst warns — signals a real regulatory risk for products that depend on open-weight models. Read more
That is what I set aside today. More tomorrow.