Most people pick one AI model and use it for everything. It's the obvious thing to do, and it's the same instinct that would have you doing every task in your business personally because you're good at your job.
There's a better shape, and it's the one every experienced manager already knows: do the thinking yourself, and hand out the doing.
The manager and the team
Picture a senior person who has been doing this work for fifteen years. Give them a big task (write the quarterly update, review two hundred supplier contracts, research forty competitors) and they'll be excellent at it and completely swamped by it. There's one of them, and the work is wide.
Now give them a team. They break the work into pieces, brief each person, let everyone go at once, and then do the thing only they can do: read what comes back and judge it. Same brain in charge. Vastly more work finished.
That's orchestration.
Why bother
Wall-clock time. Three pieces done at once finish in roughly the time of the slowest one, not the sum of all three. That's the difference between forty minutes and eight.
Cost. The expensive model does the two things that need judgement, planning and reviewing, while the volume of tokens goes through models that cost a fraction as much. You pay premium prices only for the premium thinking.
Attention. The one people underrate. Every task you delegate is a task you're not holding in your head. The point of a manager isn't that they're faster than the team; it's that they're free to look at the whole thing.
Where it goes wrong
Orchestration fails in predictable ways, and they're all versions of bad delegation.
The brief was vague. "Research the competitors" produces five documents with nothing in common. A human junior would come back and ask; a model will guess and present the guess as an answer. Briefs for models need to be more specific than briefs for people, including what the output should look like.
Nobody checked the work. If the orchestrator staples the pieces together, you've built a machine for producing plausible-looking work at speed. The review step is the entire justification for having the expensive model in the loop.
The pieces weren't independent. Splitting work only helps when the parts don't need to talk to each other.
Fans out
- Forty competitor profiles
- Two hundred contracts to check
- Twelve sections of one report
The pieces do not need each other. Run them at once.
Chains
- Find the unpaid invoices, then chase them
- Read the brief, then write to it
- Pull the numbers, then explain them
Step two needs step one’s answer. Run them in order.
It was one small task. Setting up a team to write one email is worse than writing the email. The overhead is real, and below a certain size it swamps the benefit.
Does your task want this shape?
Two questions. Can the work be cut into pieces that don't depend on each other? If yes, it can fan out. Does the whole thing need one consistent judgement at the end? If yes, it needs an orchestrator rather than forty separate chats you merge by hand.
If both are true, you have the shape. Wide work with one judgement at the end is exactly what a manager is for.
When not to use it
Most tasks don't need this shape. A single good model answering a single good question is still the right tool most of the time, and a more elaborate setup is not a better one.
It is worth understanding because the wide, repetitive, sitting-in-a-backlog work is the work that never gets done, not because it's hard, but because nobody has forty spare hours. That's the pile this shape is for.
If you have a pile like that, it's worth an hour of conversation.