What I Want My AI to Tell Me Before It Spends Anything
A UX Collective piece on transparent AI cost controls made me realize my own habit of delegating context to agents runs on the same kind of trust users need from products that bill them for AI usage.

I've been getting more deliberate about where my own attention goes. Instead of carrying every open thread myself, I hand more and more of that context-holding over to the agents I work with. Not because I've stopped caring about the details, but because I'm trying not to spin up my own mind on everything that catches my interest. I see opportunities everywhere, and the pull to start all of them at once is real. Handing the context over is how I keep my head from running all day.
That only works if I trust it's actually happening. If an agent is holding a thread for me, I need to know, without asking twice, that it's really being tracked. The moment that trust gets shaky, I'm back to carrying the thread myself, and the whole point of delegating it disappears.
I ran into the same problem from the other side this week, reading a UX Collective piece on designing transparent AI cost controls. It walks through a very practical question: when a product bills AI usage back to its users, in tokens, credits, or raw requests, how do you show that cost without turning every interaction into a math problem? The piece argues for visible limits, cost explained at the point where it's incurred, and units the user can actually reason about, instead of an abstract number that means nothing until the invoice arrives.
Reading it, I recognized my own agents on the other side of the relationship. What I want from the AI I delegate to is exactly what UX Collective is describing for the AI a user pays for: a running, visible account of what's being spent on my behalf, so I don't have to reconstruct it from memory later. Cost, in both cases, is really a stand-in for attention. Mine, when I hand off context. The user's, when they hand off a task they're paying for by the token.
Most products treat that number as something to protect the user from, something to simplify away until it's gone. But the account that gets simplified away is the one nobody trusts. A user who can't see what an AI call is costing them, in real units, at the moment it happens, has only one way left to protect themselves: use the system less. That's the opposite of the confidence a well-designed product is supposed to build.
The quiet trust I put in my agents
Handing context to an agent is a bet that it's keeping better track of a thread than I would, left to my own momentum. I don't check in constantly. I check in when something looks off, and I expect the system to have the full picture ready when I ask.
Where that trust runs out
The bet only pays off if the tracking is actually visible when I need it. An agent that holds context but can't show me what it's holding gives me no way to tell the difference between handled and dropped. That gap is exactly where I'd stop delegating.
What transparent cost design actually solves
UX Collective's argument is the same gap, from the user's side of an AI product. A cost shown in tokens or credits, explained at the point it's incurred, with a visible limit, turns an invisible number into something a user can hold in their head without holding it constantly. It's not a billing feature. It's permission to stop tracking.
What I take from this
Whether it's my own agents or a product built for someone else's users, the design question is the same: what does the person need to see, right now, to stop worrying about what they can't see? If I'm ever building something that bills AI usage back to a user, I'm taking the unit-and-moment approach from this piece directly: show the cost in a unit people can reason about, and show it where the cost actually happens, not buried in a monthly summary.
Was this note useful?
If you would rather try this than read about it, there is a free five day email challenge: Your first automation
One Decision
Every other week, one real decision from my own work. What I built, what I left out, what I handed to a machine and what I deliberately kept. With the reasoning, and what it cost when I got it wrong.
Not a newsletter. One decision at a time, close enough to your own work to be useful. Whether you are building a product or looking for a way to work that doesn't wear you down.
One email every two weeks. Unsubscribe in one click.