How much do you trust AI agents?
With the advent of clawdbots, it's as if we've all lost our inhibitions and "put our lives completely in their hands."
I'm all for delegating work, but not giving them too much personal/sensitive stuff to handle.
I certainly wouldn't trust something to the extent of providing:
access to personal finances and operations (maybe just setting aside an amount I'm willing to lose)
sensitive health and biometric information (can be easily misused)
confidential communication with key people (secret is secret)
Are there any tasks you wouldn't give AI agents or data you wouldn't allow them to access? What would that be?
Re. finances – Yesterday I read this news: Sapiom raises $15M to help AI agents buy their own tech tools – so this may be a new era when funds will go rather to Agents than to founders.
Replies
The distinction I keep coming back to is read vs. write access, and reversible vs. irreversible actions.
I'm comfortable letting AI read almost anything — it needs context to be useful. What I'm careful about is what it can do with that context. Reading a medical bill is fine. Autonomously disputing a claim on my behalf is a different matter.
Building in the document space, I've landed on a model where AI suggests and humans confirm — every action is a one-click approval, never an autopilot. That's not a limitation, it's actually the right UX: the AI does the cognitive work of reading and understanding, and I stay in control of what happens next.
The trust question isn't really about AI — it's about the design of the human-AI loop. In anything I build, the AI does the cognitive heavy lifting: reading, understanding, extracting meaning. But nothing happens until I say so. Suggestion without action is a very different thing from autonomy.
minimalist phone: creating folders
@henrikpedersen yeah, but in that case we are talking about passive vs active managin the data. I am okay that AI reads something and gives me suggestions, but when actively does things on behalf my name, that's no no
@busmark_w_nika Exactly — passive vs active is a cleaner way to put it. The moment AI acts as you rather than for you, the trust equation changes completely.
human in loop would be better idea, where human reviewing/verifying steps of the agent and let agent do rest of the work
minimalist phone: creating folders
@rajkumar_001 This would be cool! 👆
MultiDrive
recently I wrote an article for CLI, and of course all scripts I checked on my computer. I used Claude for that purposes. But I checked it on my computer, and showed it to my friend, who is a specialist in that field. So I would say I'm using it daily, but I'm checking everything. It still does mistakes.
minimalist phone: creating folders
@tetianai are you also coding?
MultiDrive
@busmark_w_nika I'm trying vibe coding :D So far, I’ve been writing an article that requires some scripts to be created and verified. Overall, I think product managers should also build some prototypes with vibe coding, although I haven’t tried it yet.
minimalist phone: creating folders
@tetianai I am trying Vibecoding too, but the results are not as expected. It looks like it can create a beautiful design, but if you do not have a full understanding of the backend and features, you cannot define the whole concept, so the tool doesn't work properly.
That's why I want to try to code it on my own with the leadership of AI :D
Aura Water: Private Hydration
Great points, Nika. We’re definitely moving from AI as a "search box" to AI as a "manager," and that shift requires a much higher level of digital hygiene.
I agree that finances and biometric data are the biggest red lines right now. Even if the AI model itself is secure, the "agentic" nature means it is interacting with third-party APIs and tools, which creates more links in the chain where a security breach could happen.
Beyond what you mentioned, I’d add unsupervised legal or high-stakes decision-making. While AI is great at drafting, letting an agent execute a contract or file a legal document without a "human-in-the-loop" review is a massive liability.
The news about Sapiom is wild—funding the agents instead of the founders really highlights how much we’re starting to treat these entities as autonomous economic actors. It makes you wonder: if the agent makes a bad "investment" with those funds, who is ultimately accountable?
minimalist phone: creating folders
@anju_locikit yeah, but at the end of the day, someone has to be responsible, and it usually doesn't end up with that agent but with a specific person,
That is indeed a super interesting topic and important one..
I have to say I would start small and let an agent option to execute small tasks only at first and only with clear boundaries and guidelines.
I would first make sure that I have a way to log and track any change of course..
sensitive information I would still keep close to heart..
minimalist phone: creating folders
@nirit_weisbrot_altony So what will be the first things you automate?
I actually just wrote a post on that a few days ago.
From a state of fast, error-prone AI execution, I've arrived at a state of systematic, consistent work within a pre-defined workflow that includes decision making based on proven facts and information. What's beautiful to see is the debate stage between the different personas and the way the conversation unfolds between them and the more pushback is given, whether by myself or another persona, the higher the quality of the outputs soars. I've built a development pipeline where the AI is not allowed to write code, until it earns the right to do so.
I don't. I am building a super nice set up with Claude Code and skills but with the human in the center. I decide what gets shipped, AI recommends.
minimalist phone: creating folders
@norteapp old school approach :D like it
@busmark_w_nika You cant trust AI yet, a lot of hallucinations and confidentially saying the wrong thing.
minimalist phone: creating folders
@norteapp yeah, sometimes, AI is like a pupil who doesn't know the right answer so starts talking nonsensical things just to say something :D
@busmark_w_nika The pain!!!
I don’t have a background in coding, so I really, really wanted to hand over all the keys to the chat agent so he could set everything up himself. But every time, I held myself back because I’d heard scary stories about AI deleting entire projects. Maybe someday I’ll trust it. But for now, I’m setting everything up manually - it’s not that hard with the instructions.
minimalist phone: creating folders
@allurepixel I think it is not certainly perfect. Even when I want something to code, AI will mess it up somehow :D
@busmark_w_nika I'm not a coder, so unfortunately I can't compare them. But I think sooner or later there will be solutions for securely embedding API keys using an AI agent, for example.
I trust AI agents with execution, not with irreversible decisions.
Research, drafts, repetitive implementation, structured workflows? Yes.
Payments, legal commitments, private health data, confidential relationships? Not without hard boundaries.
Building in this space made me believe the winning products won’t be the most autonomous ones.
They’ll be the ones with the best trust model.
minimalist phone: creating folders
@mikita_aliaksandrovich IMO, always think about worst-case scenarios—how much harm or damage the worst possible outcome could cause.
honestly same.
one thing I keep running into isn’t even autonomy risk.
It’s tiny structural failures in agent output. especially JSON that’s almost valid but breaks pipelines.
I ended up building a small fixer for that because I got tired of debugging commas and quotes from API / LLM responses all day.
curious if others here are still manually repairing those or just retrying generations?
minimalist phone: creating folders
@dechefini What is the solution you built?
I think trust is the wrong frame. I trust a calculator too but I still check the output before sending the invoice.
the way I think about it: agents get access to anything where the downside of a mistake is fixable. drafting emails, sourcing candidates, summarizing docs, scheduling. all fine. the blast radius of a bad output is low and you can catch it before it goes anywhere.
what they don't get: anything where a single wrong action is irreversible. moving money, sending legal docs, deleting data, communicating on my behalf without review. not because the AI is bad but because there's no undo button.
the Sapiom thing is interesting but "AI agents buying their own tools" sounds like a solution looking for a problem. if my agent needs a tool, I'll buy it. I don't need my agent to have a credit card.
minimalist phone: creating folders
@curtis_swick don't look only at reversability but also the extent of harm it can cause :D