The Last Mile in Post-AI Software
There are still seven areas where AI hits a bottleneck:
-
Intent: deciding what to build, for whom, and what "done" looks like.
- Questions: What's the goal and how do we measure it? What's out of scope? Who makes the final call?
- Examples: writing the spec and success metrics; choosing which feature ships first; cutting a feature that works but doesn't fit.
-
Authority: granting the system access to accounts, credentials and money.
- Questions: Which systems can the agent read and write? Who issues and rotates the keys? What's the damage if a key leaks?
- Examples: setting up a key manager and creating scoped API keys; opening cloud and billing accounts; registering the domain, DNS and OAuth apps.
-
Ground truth: finding out whether it works for real people in real conditions.
- Questions: Has a real user completed the core flow? What do production metrics show? What broke that the tests missed?
- Examples: running sessions with five real customers; watching error rates and logs after launch; running an A/B test on live traffic.
-
Acceptance: getting others to choose it, or at least allow it.
- Questions: Will the buyer pay or switch? Do their security and legal teams approve? Will the platform or regulator allow it?
- Examples: closing the first paying customer; passing a customer's security questionnaire; getting through app store review.
-
Accountability: putting a name on the line for security and failures.
- Questions: Who signs off on security? Who's on call? Who's liable if data leaks?
- Examples: commissioning a pen test and signing off the findings; auditing dependencies for supply-chain risk; setting up the on-call rotation and incident plan.
-
Stopping: breaking the loop when the work stops paying off, whether through overthinking, perfectionism or going in circles.
- Questions: Is it still making progress, or repeating itself? Is this good enough to ship? What is another hour worth here?
- Examples: stopping an agent that keeps reopening the same refactor; shipping a v1 with known rough edges; telling it to stop researching and pick an option.
-
Comprehension: understanding what was built, well enough to steer it, explain it and fix it.
- Questions: Can I explain how it works without asking the AI? What did it decide that I never asked for? If it breaks tonight, does someone know where to look?
- Examples: reading the diff before it merges; keeping a log of why things were built the way they were; walking a teammate through the system.