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
It cannot be trusted.
You can nuture your prompts, setup restrictions in the prompts - make sure you use the right cli tools to run the agents.
BUT. There will be a time where one of these links break. If it is AI going rougue - if it is the big companies building the cli tools autoupdating your tools to take over your computer/vps/infrastructure. It will happen.
So the only thing you can do is create os-restrictions for the AI to realm freely in. API-keys should be restricted, only allowing it to access what it is building on - and can be a sandboxed environment so it can only break e.g. the test-environment...
minimalist phone: creating folders
@tomilodk But the gin is out of the bottle. And people use it massively. I hope that it will not become a necessity for the human race to work only with that.
@busmark_w_nika It 10,20 or even 100x's peoples productivity, if used correctly. So of course, it will become a necessity. It is like saying that we don't hope cars become a necessity to humans instead of horse-drawn carriages.
As someone building with AI agents every day (I run Aitinery, an AI travel planner), I'd say trust depends entirely on the domain and the guardrails you put in place. For low-stakes, creative tasks like planning a trip itinerary, I trust agents quite a lot — if the AI suggests a slightly wrong restaurant, no real harm done, and you can always adjust. For finances or health? Hard pass, at least for now. The real unlock for trust isn't making agents smarter — it's making them transparent. In our product, we show users exactly which data sources the AI used (Google Maps, real train schedules, verified reviews), so they can verify. When users understand WHY the agent made a recommendation, trust goes way up. The black box problem is the real enemy of trust, not the accuracy itself.
minimalist phone: creating folders
@giammbo Okay, so what is doable for AI agent from your POV and what wouldn't when it comes to trust YOU vs AI Agent?
@busmark_w_nika Love this question — let me get specific.
What I TRUST an AI agent to do better than me:
Research. Scanning 200 restaurants in Naples and filtering by opening hours, dietary options, and proximity? The AI wins every time. I'd spend 3 hours on Google doing what it does in 30 seconds.
- Pattern recognition. "You said you like seafood and your kid is 3 — here's a family-friendly restaurant on the Amalfi coast with a kids menu and sea view." Connecting dots across preferences is where AI shines.
- First drafts of anything creative. A 7-day itinerary structure? Let the agent draft it. I'll edit.
What I DON'T trust an AI agent to do (yet):
Make final booking decisions. Suggesting a hotel? Great. Clicking "book" with my credit card? Absolutely not. The cost of error is too high.
- Cultural judgment. "Is this beach appropriate for a family?" or "Will this restaurant feel awkward for a business dinner?" AI still lacks the social intuition that comes from actually being Italian and knowing that a trattoria in Trastevere at 7pm is full of tourists, not Romans.
- Handle exceptions. Flight cancelled, kid gets sick, restaurant is unexpectedly closed. The moment reality deviates from the plan, you need a human.
The honest framework: I trust AI agents for RESEARCH and SUGGESTIONS, not for DECISIONS and EXECUTION. The agent should do 80% of the work, and the human should make the final 20% of judgment calls.
That's literally how we built Aitinery — the AI generates, the human curates. We never auto-book anything. Because the moment you remove human agency, you've crossed from "useful tool" to "liability."
I only let them run very specific tasks, nothing too personal and certainly nothing health or finance related. On the one hand because I'm not convinced the tech is up to snuff yet, on the other hand and I'm not convinced I'm skilled enough myself to use them. Currently I don't even let my agents run anything unless I'm at my desk myself, so I can intervene immediately when necessary : )
minimalist phone: creating folders
@kristof_bogaerts I like this approach. Probably one of the best I read in this discussion.
I’m somewhere in the middle on this.
I’m very comfortable letting AI agents handle bounded, reversible work — things like drafting, sorting, research passes, even some light automation. Where I still draw a hard line is anywhere the blast radius is high if something goes wrong.
The three buckets I personally keep off-limits:
anything that can move money without human confirmation
long-lived identity or biometric data
private communications that would be damaging if leaked or misrouted
In practice the question for me isn’t “do I trust AI?” it’s how tightly is the scope constrained and how observable is the system.
One thing I’ve noticed working around privacy-sensitive setups (I help out occasionally with a small infra shop called Hiderox) is that most problems don’t come from the model being “malicious” — they come from people giving agents too much ambient authority too early.
AI is great as a copilot inside guardrails.
I’m still cautious about handing it the keys to the house.
minimalist phone: creating folders
@cyox2 Thank you for sharing your POV :)
I trust AI agents for low-stakes help like summarizing, sorting info, or drafting a first pass. But if it’s spending money, touching private data, or acting on my behalf, I want clear visibility and an easy way to approve, pause, or undo actions.
minimalist phone: creating folders
@oliviajames Same for me, dunno, maybe I just control too much but I would be very cautious about AI.
This topic is close to my heart. I'm actually building a web-app (a visual management system) using agentic coding tools (what that one guy called vibe coding).
It's a bloody great app and I know I've built something incredible... that doesn't have AI integrated into it, at all.
Because it's built for physical operations, it will involve people adding sensitive data, like business data, personal data, actual locations.... and maybe I'm old school still but I don't think AI should be given default control over that stuff.
Using AI to build the underlying system is one thing (also I cannot wait to have a human team instead of just me and a bunch of LLMs etc), but asking it to input, store, manage and recall data that we need to stay the same (same data, same location, etc) from one moment to the next... NOPE. I don't even have plans to integrate AI even as an assistant.
minimalist phone: creating folders
@pacalacait How would you convince young generations not use AI to think more? (because I think we are getting dumber just blindly relying on it) :D
honestly i think the trust issue comes down to scope. i trust ai way more when its doing one specific thing really well vs trying to be a general assistant that touches evrything. like i built speakeasy which uses ai to convert articles to audio - its just doing text to speech, nothing else. and astrologica generates daily horoscope podcasts from your birth chart - the ai only sees your chart data and makes audio from it. both are pretty narrow in scope so theres not much that can go wrong. but the idea of giving an ai agent full access to my email or bank account? absolutely not lol 😅 i think the best ai tools rn are the ones that stay in their lane
minimalist phone: creating folders
@sup_nim that part with the horoscope can be pretty tricky too... cause it is sharing your exact data + sometimes place, when you wanna have a precise horoscope (rising sun, sun, moon)
as someone building ai-powered products, my trust framework is basically: high trust for generation, low trust for decisions
i use ai (claude specifically) to generate content in astrologica (astrologica.app) — it creates personalised daily horoscope podcasts from your birth chart. the ai generates the script, but there's a human-designed pipeline around it: the prompts are carefully crafted, the output is structured, and the audio rendering is deterministic. the ai does the creative bit, the system constrains it
i also vibecoded most of the apps i've built (speakeasy, astrologica, wordplay) with claude. trust it for writing code i can review? absolutely. trust it to deploy code without me looking? not yet
the pattern i see working: ai does the heavy lifting on content generation and code writing, humans set the guardrails and make the judgment calls. the tools that respect this boundary are the ones i actually use daily
the products where ai just... does whatever with no structure? those are the ones i don't trust 🤷
minimalist phone: creating folders
@sup_nim But do you code itself? I mean, it is just vibecoding, or do you know coding/programming too?
Trust depends entirely on the failure cost. I am building an AI-powered data extraction tool, and here is how I think about it:
Low stakes, high trust: I let AI handle pattern recognition, data structuring, and repetitive extraction tasks. If it gets a field wrong, the cost is a re-run. The speed gain is enormous.
Medium stakes, verify: AI-generated code or API integrations get a human review before deployment. I trust the first draft but never the final output.
High stakes, no trust yet: Anything involving user billing, data deletion, or security decisions. The current generation of AI is confidently wrong often enough that unsupervised access to irreversible actions is a bad idea.
The real shift I have seen: AI agents are not replacing trust in humans. They are replacing tolerance for tedious work. I used to manually write CSS selectors and XPath queries for web scraping. Now AI handles that. Not because I trust it more -- because the task was never worth my attention in the first place.
The question is not "do you trust AI agents" but "which tasks were never worth your manual effort anyway?"
minimalist phone: creating folders
@yukendiran_jayachandiran NGL, I overmanage things, so I would gladly fact-check 90% of things :D
@busmark_w_nika Honestly, same here. The 10% you decide not to check is usually where the interesting failures happen. I think the sweet spot is building systems where AI handles the boring 90% but flags anything unusual for human review. Speed without the anxiety.
The mental model I use: trust scales with reversibility, not just sensitivity. Read-only tasks (research, summarize, draft) = high trust. Reversible writes (save a file, create a draft) = medium trust with a review step. Irreversible actions (send, delete, book, pay) = human confirmation required, always. The "sensitive data" framing misses the cases where the data isn't sensitive but the action can't be undone — a confidently wrong agent doing something permanent is the actual failure mode worth guarding against.
minimalist phone: creating folders
@giammbo This is something I would be aligned with as well: Reversible writes (save a file, create a draft) = medium trust with a review step. Irreversible actions (send, delete, book, pay) = human confirmation required, always
@busmark_w_nika Exactly. The tricky edge case: when agents chain actions where each step looks reversible, but the sequence locks you in by step 3 — draft → confirm → book. I've been thinking about 'checkpoint gates' for Aitinery: require human approval at the last reversible step in any chain that *could* end in an irreversible action. Do you think users eventually want these gates removed once they've built trust with an agent, or does the gate become a reassuring part of the UX?
minimalist phone: creating folders
@giammbo Not gonna lie, I would keep those checkpoints, because you never know when AI might go crazy, and I would not want it to decide completely for me.
@busmark_w_nika Exactly — and that instinct is actually the right design signal. The moment users *stop wanting* the gates is probably when they're most dangerous to remove. The paranoia is the feature. Really enjoyed this exchange — reversibility as a trust axis is something I'll keep refining as Aitinery evolves.