Approval on everything: safe, or just a slower to do list?
Solo on growth at a small team, so I want to borrow the brains of this group before I overthink it more.
We build MyInc, a product that runs the revenue side for people operating lean. Not busywork, the actual money work. It ships the content, brings in leads, follows them up before they go cold, and moves the stuff that turns interest into sales, mostly while you are not working. I run growth on it, so I live with how much rope we give it.
Here is the call I pushed hardest for, and the one I now keep second guessing. We built it so nothing goes public or spends money without a human yes. It felt obviously right. You do not let a tool post to your audience or charge a card on its own.
Then people started using it and split in half. One half said that approval step was the only reason they trusted it enough to connect their real accounts. The other half said the opposite. If I have to approve every post and every follow up, you just handed me a slower to do list, which is the exact thing they came to us to escape.
Both are right, which is what makes it hard. The control that earns trust is the same control that recreates the bottleneck.
So, genuine gut check. If something was actually bringing in customers and sales for you, how much would you let it run without asking you first?
I will go first. I would let it run content and outreach on its own, but keep approval on anything that spends or charges. I hold that loosely though. Where do you land?
P.S. If you want to poke at it while you answer, it is at https://myinc.app/?utm_source=indiehackers . IH folks get 2 weeks free, no card, with code MYINC-IHFORUM. But honestly I mostly want the gut check.
The confidently wrong case never raises its hand, so I wouldn't build the line out of confidence. It can't know it's wrong, but it does know whether an action is undoable and how often it's already hit that person, so fragile lead becomes a counter: no more than two touches per contact in 14 days, enforced in code.
I'd cut it by reversibility instead of category: a bad post can be deleted and a charge refunded, but a pushy follow-up to a lead who was almost ready is gone for good. So my line would be approve the rules once, tone, discount floor, who never to contact, then let it run inside them and only ask when something falls outside. That turns approval into an exception queue instead of a slower to-do list. The fuzzy part is outreach, I don't know how you'd encode 'this lead is fragile' into a rule, which is maybe why your users split in half.
Exception queue over gating every action, that is exactly it. Fragile lead is where I am stuck too, best I have is reading it as signals (recency, reply speed, quiet after a price question) not a hard rule. You clearly think about this sharper than most, any chance you would poke at it and tell me where it breaks? Free code is up in my post, no card. Would love your read.
Deal. I'll grab the code and run it like it's mine, probably over the next few days, then reply here with wherever it breaks or surprises me. Fair warning: I'm the paranoid kind of user who reads what an agent is allowed to do before admiring what it does, so expect most of my notes to be about the rope, not the features.
The risk I'd watch for is treating approval as a product setting instead of a trust outcome.
People don't really want approval. They want confidence. If the product can earn that confidence over time, autonomy can increase naturally instead of forcing every user into the same level of control.
Yeah, that reframe genuinely reorganized it for me. The concrete version of your point I keep landing on: it earns rope the boring way, does a bunch you keep approving without edits, then asks to stop checking on those. Feels like the track record basically is the trust. Enough for you, or still want a hand on it?
Interesting.
Your last sentence is the part I'd want to explore.
I think it quietly commits the product to a direction that isn't obvious at first glance, and I'd be reluctant to give you a one-paragraph answer because I think it would be misleading.
If you'd like to unpack it properly, happy to continue over email.
I wonder if the tradeoff is not really approval versus automation.
For me, it is more about confidence versus uncertainty.
If the system is highly confident and the cost of being wrong is low, asking for approval every time can become unnecessary friction.
But when confidence drops or the potential impact of a mistake is high, that feels like the right moment to bring a human back into the loop.
So instead of asking, "What always needs approval?", I would probably ask, "Under what conditions should the system stop trusting itself and ask me?"
Flipping it from what needs approval to when should it escalate is a way better question. The catch: it only works if the system actually knows when it is unsure. The scary case is not low confidence, it is being confidently wrong, it does not ask because it thinks it has this one. So it all rides on how well it knows what it does not know. So would you rather set those escalate triggers yourself, or let it figure them out over time?
I would probably start with explicit triggers.
Not because they are always right, but because they are visible, predictable, and easy to challenge when they fail.
If the system learns everything on its own from day one, it becomes much harder to understand why it escalated one case but not another.
Over time, I would absolutely let it learn from experience. I just think trust grows faster when people can first observe and refine the boundaries instead of treating them as a black box.