What are your biggest issues/ frustrations with using AI for your work??

The Frustration Is Real: Why AI at Work Feels Broken

AI use at work has nearly doubled in two years, yet most employees are flying blind. According to Gallup research, the share of U.S. workers using AI at least a few times a year jumped from 21% to 40% — but only 22% say their organization has communicated a clear AI strategy. Gallup's blunt summary: many employees are "using AI at work without guardrails or guidance."

That guidance gap explains a lot of the frustration. Sentiment data tells the same story: 52% of U.S. workers are more worried than hopeful about AI in the workplace, and only 16% of AI users strongly agree their organization's tools are actually useful for their work.

When you dig into what people actually complain about, the frustrations cluster into five recurring themes:

• Hallucinations and fabricated output — plausible-sounding but false answers, from invented citations to confidently wrong facts.
• Unclear value — tools adopted without a defined business problem, so nobody can tell whether they help.
• Inconsistent outputs — the same prompt producing different answers across sessions and model versions.
• Bland quality — repetitive phrasing, generic research, and an unwanted academic default tone.
• Usage limits and failed execution — paywalled task caps and breakdowns on complex, multi-step work.

The hallucination problem deserves special attention because it is not a bug you can patch. Peer-reviewed research argues hallucinations are "structurally inevitable" — a consequence of how generative models predict patterns rather than verify truth. IBM puts it plainly: they "can be reduced but not fully eliminated." The documented failures are real, too, from a lawyer's fictitious case law to a chatbot that misled a customer with legal consequences.

Inconsistency is the quieter frustration. A peer-reviewed diagnostic study found accuracy ranging from 76.9% to 92.3% across leading models, with response stability varying just as widely — leading its authors to urge evaluating tools on both accuracy and consistency, not just one.

And the organizational layer often makes things worse. MIT CISR researchers find companies starting AI projects without a clear path to value, getting stuck in pilots, and mistaking productivity gains for real business value. As researcher Nick van der Meulen observes, organizations are "applying yesterday's best practices to an inherently different technology."

None of this means AI at work is hopeless — quite the opposite. Users who actually work with AI report far more positive experiences than skeptics expect, and clear communication from leadership makes employees three times as likely to feel prepared. The frustration is real, but so are the fixes. The sections that follow walk through each frustration cluster in turn, with practical ways to mitigate it — including how verification and benchmarking, the approach SkillzBench is building for agent skills, applies to any AI workflow where a wrong output can cascade.

Hallucinations and Inconsistency: The Two Technical Deal-Breakers

You can forgive an AI tool for being slow or bland. What you cannot forgive is being confidently wrong — and that's precisely the failure mode users keep running into.

The technical reality is sobering. Peer-reviewed research from the Harvard Kennedy School concludes hallucinations are "structurally inevitable," stemming from layered vulnerabilities in training data and weak downstream gatekeeping. IBM agrees, stating plainly that hallucinations "can be reduced but not fully eliminated" — because generative AI doesn't know what's true, only what fits the pattern.

The real-world consequences are well documented:

• A US lawyer's ChatGPT-drafted court filings contained fictitious case law, with professional embarrassment to match.
• Air Canada's chatbot misled a customer about bereavement fares — and the airline faced legal consequences for it.
• OpenAI's Whisper fabricated misleading content in medical transcriptions.

Hallucinations are also uniquely hard to catch. As IBM notes, their adherence to real patterns makes them seem authoritative — a fluent fabrication is more dangerous than an obvious error.

The second deal-breaker is inconsistency. A peer-reviewed diagnostic study found accuracy ranged from 76.9% to 92.3% across models, with intra-model consistency (Cohen's kappa) spanning 0.560 to 0.782. Even the same model can give different answers to the same question on different days — and the study's authors caution that such differences "should therefore be interpreted cautiously." The practical takeaway: evaluate tools on both accuracy and stability, not marketing claims.

The stakes rise sharply in multi-agent workflows. IBM documents how a single fabricated metric cascaded through an agentic cloud-cost workflow — an agent deleted a storage bucket under legal hold based on a hallucinated inactivity figure. Downstream agents treat upstream output as verified fact, so one hallucination compounds at every step.

This is why mitigation, not elimination, is the realistic goal. Verification layers, guardrailed prompting, and independent benchmarking matter more than picking the "best" model. SkillzBench is building exactly this kind of scrutiny for agent skills — security scanning and sandbox benchmarking designed to catch risky components before they touch your workflow. Skills, proven — join the SkillzBench waitlist to see which agent skills earn their place before you deploy them.

Mitigation, Not Elimination: Designing Workflows That Catch AI Mistakes

If you cannot stop AI from making things up, the real question becomes how your workflow catches those mistakes before they matter. That reframing matters because hallucinations are not a bug you can patch — researchers argue they are "structurally inevitable" as long as models predict the next likely word rather than verify facts, which is why peer-reviewed analysis treats them as an ongoing technical challenge, not a solvable defect.

IBM's guidance is blunt: hallucinations "can be reduced but not fully eliminated," and they are harder to spot than ordinary errors because plausible-sounding output feels authoritative (IBM). The practical answer is to design verification into the process itself, rather than hoping a different model solves it.

A few mitigations with real evidence behind them:

• Ground the model in your own sources. Retrieval-augmented generation (RAG) feeds the model verified documents instead of letting it improvise from training patterns, dramatically reducing fabricated answers.
• Build guardrails into prompts. Instructions like "if you are unsure, say you do not know" give the model permission to abstain instead of guessing, and structured output formats make errors easier to spot.
• Verify AI-supplied sources. In hands-on testing, the G2 reviewer checked every citation from Deep Research manually — because fabricated references are exactly the failure mode that embarrassed a US lawyer whose ChatGPT-drafted filings contained fictitious case law.
• Test tools on accuracy and consistency. A peer-reviewed study found models varied from 76.9% to 92.3% accuracy and, crucially, differed in how stable their answers were across repeated runs — so evaluate both before adopting.
• Prompt explicitly for tone and style. Practitioners report repetitive phrasing, generic research, and an unwanted academic default tone; explicit style instructions and feeding real transcripts as input fix much of this.

Set expectations accordingly. Pew's survey of 5,273 U.S. workers found users rate AI far more helpful for speed (40%) than for improving work quality (29%) (Pew Research). Treat AI as a fast first draft that a human verifies — not a finished product.

This matters even more in multi-agent workflows, where one hallucination cascades as downstream agents treat it as verified fact. That risk is exactly why SkillzBench is building sandbox benchmarking and security scanning for agent skills — so you can see how a skill actually performs before it touches your workflow.

Fixing the Organizational Gap: From Unguided Pilots to a Clear AI Plan

If your AI projects feel busy but pointless, you're not alone. MIT CISR researchers found that many organizations start AI initiatives without a clear path to value, get stuck in pilots that never scale, and mistake productivity gains for actual business value. Few have turned experimentation into initiatives that move critical business metrics, according to MIT Sloan research.

The root cause is often a guidance gap. Workplace AI use nearly doubled in two years, yet only 22% of employees say their organization has communicated a clear AI plan — while 44% say AI integration has already begun. Gallup puts it bluntly: many employees are "using AI at work without guardrails or guidance" (Gallup).

That gap has a measurable cost in confidence. When leaders clearly communicate an AI plan, employees become 3x as likely to feel very prepared to work with AI and 2.6x as likely to feel comfortable using it. The message lands: "If leaders want greater AI adoption, they will have to help employees find the value," Gallup concludes.

The fix starts with asking better questions before launching anything:

• Which specific business problem does this AI initiative solve?
• How will we know it worked — which metric moves?
• Is this a personal productivity tool or a process-integrated solution?
• What standards and guardrails will employees follow?

MIT CISR researcher Nick van der Meulen describes the pattern: organizations "govern AI like legacy IT" and "mistake productivity shaves for enterprise value." A team saving an hour a week on drafts isn't the same as transforming a workflow that drives revenue or cost outcomes. Barbara Wixom, also of MIT CISR, adds that AI is advanced data science requiring the right capabilities to manage properly.

There's also a familiarity effect worth exploiting. Among employees who actually used AI for customer interactions, 68% reported positive effects — versus just 13% of non-users who expected them. Much of the frustration and skepticism around AI is a familiarity gap, not a technology gap. Getting skeptical employees hands-on, inside a defined plan, converts doubters faster than any memo.

For teams building agentic workflows, the same discipline applies to the components. Published research found roughly 26% of agent skills carry at least one vulnerability, which is why SkillzBench is building a free directory that security-scans, sandbox-benchmarks, and version-pins every skill before you connect it through its free MCP tool.

A clear plan beats a clever pilot. Define the value path first, communicate it relentlessly, and measure against business metrics — not hours saved. And when you're ready to extend agents with skills, join the SkillzBench waitlist to see skills proven before you deploy them.

Your Action Plan: Vetting AI Tools and Agent Skills Before You Trust Them

Vetting AI tools before you trust them starts with a question most teams skip: what business problem are we actually solving? MIT CISR researchers find that organizations fail when they launch AI projects without a clear path to value, get stuck in pilots that never scale, or mistake small productivity gains for real business value. If you can't name the metric you want to move, the tool choice doesn't matter yet.

Once the problem is defined, benchmark candidates on two dimensions: accuracy and consistency. They're not the same thing. In one peer-reviewed study, two models hit 92.3% accuracy while another reached 76.9%, and response stability varied widely between models too. The study authors explicitly warn that these differences — driven by training data, architecture, and versioning — should shape tool selection rather than being assumed away. Test the same prompts repeatedly, across sessions, before you commit.

Then there's the layer most teams never inspect: the agent skills you plug into your workflows. Published research found that roughly 26% of agent skills carry at least one vulnerability, and the stakes are higher in agentic setups. IBM documents how a single hallucination can cascade through multi-agent workflows — downstream agents treat fabricated output as verified fact, with one example of an agentic workflow deleting a cloud storage bucket based on a made-up metric. A hallucinated instruction inside a skill doesn't stay contained; it propagates.

Before you deploy any tool or skill, run this checklist:

• Define the business problem and target metric first — not the technology
• Benchmark accuracy and consistency on your own repeatable tasks, not vendor demos
• Verify any AI-supplied sources or citations before they feed downstream decisions
• Security-scan agent skills and pin them to specific versions so nothing changes silently
• Test skills in a sandbox, with and without them, to measure what they actually add

That last two steps are exactly what SkillzBench is building: every skill in its directory gets pinned to a commit SHA, security-scanned, and benchmarked in a sandbox with the same model — with and without the skill — so you can see the measured lift before anything touches your workflow. A free MCP connector lets you attach only the skills you've chosen to any agent.

None of this eliminates hallucinations — peer-reviewed research argues they're structurally inevitable, and IBM notes they can be reduced but never fully eliminated. The realistic goal is containment: know your tools, prove your skills, and design workflows that assume errors will happen.

Want skills proven before you deploy them? Join the SkillzBench waitlist — the directory, benchmarks, and MCP connector will all be free.