How much do you trust AI agents?

I use AI daily for work, but I draw the line at anything that requires actual accountability. If an agent messes up my schedule, it's annoying; if it messes up a bank transfer or a government application, it's a disaster. Until AI can be legally held responsible for its mistakes, my wallet and my ID stay offline.

I work on agent tooling and the thing nobody talks about is that trust is a spectrum, not a binary. Right now most people either let the agent do whatever or don't use it at all.

What changed my mind was building replay/trace infrastructure — when you can go back and see exactly what an agent did step by step, you stop worrying about whether to trust it. You just check. Same way you'd review a junior dev's PR, not because you don't trust them, but because that's how you build confidence over time.

The actually scary failure mode isn't "agent accesses my bank account." It's 50 small reasonable-looking decisions that compound into something you didn't want.

Great thread, Nika—trust in increasingly autonomous AI agents is one of the defining questions of the next few years.

My personal boundaries are quite similar to yours:

Hard no-go zones (no access, no exceptions):

• Full control over personal banking, crypto wallets, or payment methods (only ever a capped “burner” amount I’m willing to lose)

• Direct access to health records, genetic data, continuous biometrics, or medical history

• Any end-to-end encrypted or highly confidential communication (family, legal, therapy, C-level business secrets)

• Actions with real-world legal, financial, or physical consequences (signing contracts, posting publicly on my behalf, controlling smart home/security devices)

What I do delegate today (with monitoring):

• Read-only analysis of finances (categorization, forecasting, anomaly detection)

• Email triage, drafting low-stakes replies, meeting prep & follow-ups

• Research, summarization, task triaging, calendar suggestions

• Code generation/review in isolated environments

The Sapiom news is a fascinating (and slightly dystopian) signal—if agents start managing their own budgets and tooling autonomously, we’re moving toward agent-to-agent economies where human oversight becomes even more critical. That could unlock insane productivity… or create entirely new classes of misalignment risk.

Where do you personally draw the line between “helpful proactive assistant” and “too autonomous to feel safe”? Curious about your take as a minimalist-tool builder. 💭

Upvoted—excellent conversation starter. 🚀

As a data engineer, I think about this differently: it's not just about trusting the AI — it's about where your data goes when you use it. Most AI tools require uploading your files to their servers, and that's where the real trust question lies. I'd rather use tools that process data locally whenever possible. For repetitive data tasks (like reformatting spreadsheets), you don't even need AI — you need well-built deterministic tools. Not everything needs to be "smart" to be useful.

Hi Nika,
I also limit AI involvement in activities not related to personal sensitive areas.
The main risk in my opinion is AI use in countrie's attack/defense systems - mainly those with nuclear weapons. How far AI can interfere in this area - today and tomorrow?

I am less worried about the first barrier (agents handling sensitive data) however I am way more worried about the stochasticity. Agents ARE NOT RELIABLE. And I say this after I've had patents on solving for agent reliability. Even if you end up building a Vault for the agents world, the usage of the said vault is so stochastic that I don't want to trust agents yet. In a multi-step workflow, this compounds catastrophically.

Like with most people, I'd trust AI for most things, except for the things you listed (e.g. personal finances). At least not without approval processes built in (e.g. get a push notification to approve paying a bill).

I can barely control myself on Black Friday, and now I'm supposed to let a robot handle it? No thanks, the card stays with me. It can pick references for me, but spending money? I've got that covered just fine on my own 🤣

I don't mind it accessing productivity-related services (Notion, Google Docs, etc). I have been careful with giving it the ability to email/contact people, however I don't mind it reading my emails.

I'm also not comfortable giving it system level access to my Mac, and obviously that's a common concern which is why everyone's buying their own hardware.

Great question! As someone building AI products, I think about this constantly.

The trust issue isn't just about the AI itself - it's about the entire data pipeline. Where does your data go? Who has access? Can it be deleted?

I draw hard lines around:

- Financial accounts (read-only at most, never write access)

- Personal communications (drafts only, never auto-send)

- Health data (absolutely off-limits)

The bigger concern for me is the "black box" problem. Most AI agents don't show you their reasoning process. You see the input and output, but not what happened in between. That's scary when dealing with sensitive tasks.

I think the future is "trust but verify" - agents that show their work, have audit trails, and give you granular control over what they can access. Until then, treating them like talented interns (not managers) seems like the right approach.