Honest struggles of bringing AI into everyday work – what I’m seeing and what I’d love to hear from developers

I’m Seohyun, an AX Researcher at Knowverse. My job isn’t to build models or maintain infrastructure; it’s to explore how finished AI services can be woven into the way we actually work – writing reports, synthesising research, preparing meeting notes, and so on. Over the past year I’ve tried to bring a handful of tools (ChatGPT, Claude, Gemini, Perplexity, Notion AI, and a few meeting‑assistant apps) into our team’s daily rhythm. While the promise of time‑saving is real, I’ve run into a few recurring friction points that feel more about process and trust than about the technology itself. I’m sharing them here because I suspect the people who create these tools might see them from a different angle, and I’d value your perspective.

1. Where do we draw the line on trusting AI output?
When I ask a language model to summarise a 30‑page market report, I get a concise bullet list in seconds. That’s tempting to paste straight into a slide deck. Yet I’ve noticed occasional hallucinations – a missing data point, a mis‑attributed trend, or a fabricated quote. The model’s confidence tone makes it easy to overlook those slips. I end up spending extra minutes cross‑checking every claim against the source, which sometimes erases the speed gain. I’m curious: from a builder’s standpoint, how do you design systems that signal uncertainty or encourage users to verify critical facts without adding prohibitive friction? Are there patterns (citations, confidence scores, source highlighting) that have worked well in practice?

2. Getting teammates to actually use the tools.
I’ve introduced a meeting‑assistant that transcribes calls and drafts action items. I use it religiously and find it saves me roughly 20 minutes per meeting. Yet many colleagues still rely on manual notes or ignore the assistant altogether. The reasons I hear range from “I don’t want to change my habit” to “I’m not sure the notes are accurate enough for client‑facing work.” It feels less like a tool problem and more like a cultural inertia issue. What strategies have you seen work when rolling out new AI‑enabled features to teams that are comfortable with the status quo? Is it about training, default settings, or something else?

3. Choosing the right tool for the right task.
Not all AI services are interchangeable for me. For quick factual queries I lean toward Perplexity because of its source citations; for drafting creative copy I prefer Claude’s tone; for dense PDF extraction I find Notion AI’s inline summariser handy. Switching between them means managing multiple accounts, learning different prompt styles, and keeping track of where each piece of output lives. I wonder if there’s a way developers could make the boundaries between services clearer – perhaps through unified APIs, consistent prompt templates, or better interoperability – without forcing a one‑size‑fits‑all model that sacrifices each tool’s strengths.

4. Prompt fatigue and the “just do it myself” moment.
I often spend more time crafting a detailed prompt than I would spend doing the task manually, especially when the output needs to be very specific (e.g., a formatted SWOT analysis with exact headings). After a few iterations I catch myself thinking, “I could have just written this in ten minutes.” That feeling makes me hesitant to rely on AI for tasks that demand precision. How do you think about prompt design to reduce the trial‑and‑error loop? Are there emerging practices (prompt libraries, reusable snippets, guided wizards) that help power users like me get reliable results faster?

5. Privacy and data handling concerns.
When I feed internal documents into a cloud‑based AI, I’m always aware that the data leaves our environment. Even though the providers assure us of non‑retention policies, the uncertainty makes me reluctant to use AI for sensitive strategy work. I’ve started to keep a “safe‑list” of low‑risk tasks (public‑facing research, generic copy) and reserve the rest for offline tools. From your side, what guarantees or technical safeguards do you think would give non‑engineer users more confidence to bring AI into confidential workflows?

These points aren’t criticisms of the technology itself; they’re observations about how AI services intersect with human habits, trust structures, and everyday workflows. I’m eager to hear how you, as the people building these tools, think about bridging the gap between raw capability and practical, sustainable adoption. If you’ve encountered similar feedback from non‑developer users, what adjustments have made the biggest difference?

Thanks for reading – I look forward to your insights.