Claude Opus 5 dropped last week, and the industry did the thing it always does: benchmark fights, vibe checks, and a thousand posts asking which model you should switch to.
Meanwhile a quieter product category keeps growing: the model router. Middleware that reads your request and sends it to the "right" model. Cheap tasks to cheap models, hard tasks to expensive ones. There are startups whose entire pitch is being the traffic cop between you and the AI.
From a small operator who actually runs this stuff daily: you probably don't need a router. You need a manager.
For the past few months I've run my one-person business on an AI staff. Not metaphorically. A coach model plans my day, audits the system, and makes judgment calls with me. Cheaper models do the mechanical work in batches: file migrations, tag cleanups, research sweeps. Overnight agents patrol the files while I sleep and leave reports for the morning.
So when Opus 5 released, the question was never "should I switch?" It was "what job should this new hire get?" We did what you'd do with any hire: an audition. Two supervised trials on real work, a reviewer checking every output. It passed its first trial by catching a duplicate file and refusing to guess — which told me more than any benchmark. It earned a specific lane: bounded jobs with a clear endpoint, always verified after. It did not earn the keys to everything, because real work showed where it stumbles.
That's routing. No middleware, no config file. A capable model with context about the whole operation decides what runs where, the same way a good manager staffs a team.
(Full disclosure: my current staff all comes from one vendor, the way a small shop might hire through one staffing agency. The org chart doesn't care. The manager pattern is the durable part. The seats are swappable.)
Four rules made it work, and any business can copy them:
Split judgment from mechanics. If a task has a checkable endpoint (did every file get the tag?), a cheap model plus a verification step beats an expensive model working unwatched. Save the strong reasoning for decisions and anything touching your reputation.
Audition, never assume. New model? Two supervised trials on real work before it touches anything that matters. The benchmarks told me almost nothing. The trial where it refused to guess told me everything.
Verify everything unsupervised. The most common complaint about the newest models is claiming finished work that isn't. A verification gate catches that. A router doesn't.
Nothing ships without a human. Every publish, every message, every decision that binds me is mine. The staff drafts. I sign.
This is the adoption gap again, one layer down. The companies disappointed with AI mostly bought capability and skipped the management layer. The gap was never which model. It was whether anyone was running the org chart.
This week my system got audited by its own staff and passed. Next week I start building a task app on top of it. If you want the receipts as they happen, that's what this newsletter is.
One Thing to Do This Week
Pick the most mechanical, checkable chore in your work — something where "done" is verifiable (every row tagged, every file renamed, every link checked). Write three lines before you start: what the task is, what done looks like, and how you'll check it. Hand it to the cheapest AI you have and grade the result against your three lines. That grading step is the whole skill. You just ran your first managed delegation, no router required.
The Implementation Lane is a weekly newsletter about making AI work inside real organizations. Written by Amanda Crawford, an AI Implementation Specialist who builds systems in the gap between configuration and engineering. If someone forwarded this to you, subscribe here.