I paid Perplexity $100 expecting coordinated AI work. What I bought instead was a clean view of the coordination problem.
I work with four consumer AI systems: ChatGPT and Codex, Claude, Gemini, and Perplexity Computer. Each is capable. Each has a useful lane. Together, they can research, write, build, analyze, and move work forward faster than any one of them alone.
But capability is not coordination.
During one recent run, an AI system recommended three possible additions to my team. Two of them were products already on the team. The system could see the task in front of it, but it could not see the operating structure around the task. That miss was small. The correction was not. I had to stop, explain the roster, re-establish the lanes, and reconstruct context the machines were supposed to help carry.
The four taxes
1. The trust tax
Every small miss teaches me to watch the next move. Monitoring replaces delegation. The promised time savings disappear into supervision.
2. The coordination tax
Research lives in one system, drafts in another, local files in a third, and account actions somewhere else. The handoff still travels through me.
3. The authority tax
Every seat must know what it may carry silently and what requires my approval. Publishing, spending, deletion, access, and commitments made in my name stay protected. Routine execution should not keep bouncing back to me.
4. The identity tax
Each system knows a partial version of the work and a partial version of me. I become the identity layer—reintroducing the purpose, relationships, standards, and decisions that should already be available to the team.
What the market still does not provide
No consumer-facing service currently gives me reliable shared operating memory across these four systems. They can exchange files and prompts. They cannot natively maintain one authoritative understanding of the team, the work, the permissions, and the reason behind the work.
That distinction matters. A shared folder is not a shared brain. A long prompt is not durable governance. An automation is not useful if its exceptions, errors, and handoffs keep landing back on the person whose time it was supposed to protect.
I still have to reconcile and approve every deliverable because no seat carries the complete authoritative context. Until that changes, multi-AI operations remain less like hiring a team and more like managing several gifted contractors who never attend the same meeting.
The real standard
The test is not whether the agents are busy. The test is whether the human stops repeating context, stops catching avoidable language errors, stops receiving unnecessary choices, and stops serving as courier between systems.
Work should keep progressing while the human thinks, speaks, rests, and creates. Disagreements should be preserved. Risks should still be flagged. Protected actions should remain protected. But ordinary operational work should not be transferred back to the person who asked for help.
This is not a complaint about one company. Perplexity made the architecture visible, but the gap belongs to the category. Every vendor is selling intelligence. The next durable advantage will belong to whoever can deliver governed continuity across intelligence.