Software is learning to move
The old product waited for clicks. The new one drafts, routes, compares, calls tools, and carries work across steps. The standard must move from usable to auditable, reversible, and legible.
Principles
Keep your eye on who gets to decide.
AI now drafts, routes, compares, and acts. That changes the job of the interface — and the terms on which people hand work over.
We are interested in acceleration with receipts: who sees the action, who can question it, and who owns the consequence when the system gets it wrong.
The questions underneath the noise
The novelty is not that machines can produce output. It is that they can now carry work across boundaries. Context, permission, judgment, and trust become the hard parts.
Work is becoming supervision
More jobs are becoming a loop of directing, checking, correcting, and approving. The person is still there, but the shape of the loop is changing.
Governance is now product behavior
Policy is expressed in defaults, scopes, memory, logs, permissions, and the user's ability to intervene before the system crosses a line.
Verification is the bottleneck
Machines can produce faster than teams can inspect. The hard part is no longer getting output. It is knowing what to trust.
The screen is becoming a checkpoint
The best interface may appear for only a second: a boundary, a consequence, a correction, a final yes before the system gets out of the way.
Weird workflows win
The strongest systems will not be generic assistants for everything. They will know the messy, specific work well enough to make the next human decision easier.
Our editorial lens
These are the tests we bring to an argument, a product claim, or a change in how work gets done. They are not decorative language.
1. Agency Over Automation
Moral stance: People should be able to delegate work without disappearing from the outcome.
Practical rule: Use graduated autonomy: draft, explain, confirm, execute. For high-stakes actions, provide explicit approval and a clear rollback path.
2. Clarity Is Respect
Moral stance: Confusion transfers risk from the system to the user.
Practical rule: Every consequential action must answer three questions: what happened, why it happened, and what can be done next.
3. The Screen Is A Checkpoint
Moral stance: Speed is not a virtue when the blast radius is large. A person should not discover after the fact that software crossed a line for them.
Practical rule: Introduce checkpoints for irreversible actions, external commitments, health decisions, and money movement. Remove friction everywhere else.
4. Privacy Is Architecture, Not Policy
Moral stance: People should not have to trade dignity for utility.
Practical rule: Minimize collection, expire data by default, scope permissions tightly, and make memory, tool access, and data flows visible in plain language.
5. Verification Is The New Interface
Moral stance: When software can produce more than people can inspect, trust cannot run on faith.
Practical rule: Put sources, assumptions, confidence, diffs, tests, approvals, and consequences close to the action itself.
6. Workflows Beat Demos
Moral stance: A system that dazzles in isolation but fails inside real work is not useful. It is theater.
Practical rule: Build for the full loop: context, permissions, handoffs, stale data, exceptions, review, and recovery.
7. Accountability At The Point Of Action
Moral stance: A system that cannot be audited cannot be trusted.
Practical rule: Keep clear traces of decisions, tools used, and user approvals. If agents act inside a workflow, they need owners, boundaries, escalation paths, and a way to be corrected.
8. Design To Strengthen Judgment
Moral stance: A good system does not make the person smaller. It gives them a better view of the terrain.
Practical rule: Surface what changed, what is uncertain, what needs taste, and what deserves a hard no.
The short version
Follow the systems that move work forward — and keep asking where judgment, authority, and trust went.
Read the latest thinking on the blog or reach out at contact@batjko.com.