I delegate the building and fixing of internal business tools to AI agents, and I keep a log of it. Across 290 delegated tasks, the recorded failure rate for delegation calls was 7.2%, and only 0.7% had to be re-requested from scratch. (Figures from my own work log of 290 tasks delegated to AI tools; not an independent study.)
In those records, the main source of slowness was queueing. Only 2 of those 290 tasks had ever been delegated in parallel. Everything else waited in line: I waited for one unrelated task to finish before delegating the next. The slowness was not a prompting problem. It was a scheduling problem.
Two questions now decide, in advance, what may run at the same time:
- Do I need to see this task's result before I can even write the request for the other one? If yes, they run in sequence.
- Do both tasks rewrite the same files or the same tables? If yes, they run in sequence — or I split the work areas so they don't collide.
Read-only work — research, re-reading, checking — usually answers "no" to both, so it goes out in parallel. I aim for 3–5 tasks at a time; beyond that, the person reading the results (me) cannot keep up. Flows where each step consumes the previous result ("research → decide → build"), and the final adopt-or-reject judgment, stay sequential.
The same queue can form on the receiving side. I read results in a batch once they have all arrived. Reading and replying to each one as it lands just recreates the queue on the receiving side. And only failed tasks get re-requested — never the whole batch.
After adopting this rule, I revisited a later version of the log containing 328 tasks. In the 13 batches I examined, unrelated tasks could be delegated simultaneously in 10 (76.9%), and total waiting time dropped by about 26 minutes. (The 76.9% is a batch-unit rate — 10 of 13 batches — not a share of tasks. The 26 minutes is the cumulative saving across those batches, where each batch's saving is the sum of its tasks' durations minus the longest single duration: 1,560 seconds in total. The durations include a slight hook startup delay. These numbers are from my own environment and will vary with the work.)
So when things feel slow, I count the waiting time before I doubt the amount of delegation or the AI's quality. In those records, the bottleneck was the waiting between unrelated tasks. It was how much of it I made wait in line.
The two questions above are one part of a larger operation — the role split, the six-field request form, and the acceptance checks. The form itself is free: The Six-Field Request Form. The full operation, including the parallel-delegation test and the complete 8-item acceptance checklist, is in my playbook, Delegating Work to AI ($12; the LAUNCH code takes $3 off until November 23, 2026, at 23:59 JST).
What do you queue up that could run in parallel?
Drafted with AI assistance from my own playbook; all figures are from my own operating log.

