Notes on governed operations.
How rules, AI steps and human approval fit together in work that has to be explained later — written against what Miji, the enterprise decision automation platform, runs today.
Latest Thinking
The difference between an AI agent and an AI worker
Agents are impressive in demos. Workers are accountable in production. Understanding the distinction is the first step to running AI labor safely.
What a work trail actually needs to contain
Audit trails in AI systems are not optional. But most implementations capture what happened — not why it happened, who authorized it, or what data was used.
Human-in-the-loop is a feature, not a limitation
The goal is not to remove humans from AI operations. The goal is to put humans exactly where their judgment adds value — and keep them out of the rest.
Deterministic workflows and AI reasoning: when to use each
Not every step in an operation needs AI reasoning. Knowing when to use a structured workflow vs. an AI worker is one of the most important design decisions in AI operations.
For Lenders
All lending postsWhich policy governed this loan? Answering the examiner's question
When a decision is questioned a year later, a log excerpt is a weak answer. The strong answer is the exact policy version, the inputs it saw, and a re-run that produces the same result.
Your verification policy needs version numbers
Credit overlays usually live in a memo, a spreadsheet and three processors' heads. Turning them into a versioned policy is what makes identical files get identical treatment.
Only the files that need judgment: exception queues and approval authority
Reviewing every file is not control — it is a queue. Control is deciding in advance which files need a person, which person, and how many of them have to agree.
What the Fannie Mae and Freddie Mac AI mandates mean for lenders
For the first time, both GSEs require AI governance from every seller — including for AI your vendors provide. Here is what changed, and what you now have to be able to produce.
Where This Is Going
Why AI-native companies will manage goals, not headcount
The assumption that getting work done requires hiring people is breaking down. Here is where Miji is heading — and why the management layer matters more than the model.
ReadUse Case
One manufactured-home file, from the retailer to a decision
The application, documents, credit, the maths and the rate grid, the policy that decided, the underwriter who approved — and what each step leaves on the record.
See it on an operation like yours.
Thirty minutes: the workflow, the rules, the approvals and the record it leaves.