Do/have you worked with people who confidently rely on AI to do most of their job??

The Rise of Confident AI Reliance — And Why It's Outpacing Oversight

You know the colleague. The one who pastes the whole brief into ChatGPT, forwards the output untouched, and calls the task done — with total confidence and zero second-guessing. If that made you uneasy six months ago, the data now says you're surrounded by them.

The numbers behind that feeling are stark. According to Gallup's workplace research, the share of U.S. employees using AI at least a few times a year doubled from 21% to 40% in just two years, while daily use doubled from 4% to 8% in a single year. Heavy, habitual reliance is no longer an edge case — it's the fastest-growing pattern in the modern workplace.

The problem is what's not growing alongside it: oversight. The same Gallup data reveals a 14-point governance gap — 44% of employees say AI integration has begun at their organization, but only 22% say leadership has communicated a clear AI plan. Gallup's own conclusion: many employees are using AI at work without guardrails or guidance.

It gets murkier. A Federal Reserve review of workplace AI surveys found that for 29% of workers, management isn't even aware AI is being used. That means nearly a third of AI-assisted work — spreadsheets, reports, analyses — flows into business decisions with no one upstream knowing it was machine-drafted.

The risks of that blind trust aren't theoretical:

• AI models pattern-match rather than reason, and cannot know when they're hallucinating — fluency is not accuracy.
• Microsoft's official guidance warns that "delegating work to AI doesn't transfer accountability", and specifically flags subtle errors like incorrect spreadsheet formulas as requiring human validation.
• Workers themselves rate AI as better at speeding up work (40%) than improving its quality (29%), per Pew Research Center findings.

That spreadsheet example deserves a pause. An AI-generated formula can look impeccable — clean syntax, plausible logic — and still quietly misreference a range or mishandle an edge case. Confident reliance on unreviewed output is exactly how those errors compound across linked workbooks and downstream reports.

The core tension is now clear: reliance is scaling faster than the guardrails. For teams extending that reliance into AI agents, verification matters even more — published research has found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building free sandbox benchmarking and security scanning for exactly this problem, so teams can see what a skill actually adds before trusting it with real work.

The Evidence: When Heavy AI Reliance Helps — And When It Actively Hurts

The most striking finding about heavy AI reliance is that it doesn't affect everyone the same way. In a randomized field experiment with 640 entrepreneurs, high performers boosted profits and revenue by roughly 10–15% with AI access — while low performers saw results decline by about 8%, according to research reported by Harvard Business School. Same tool, opposite outcomes.

The differentiator wasn't access to AI or even the quality of answers received. It was which advice users chose to follow. Rembrand M. Koning, the Harvard researcher behind the study, puts it plainly: you need to ask whether the person using the tool "has enough judgment for tasks that are required." Without that judgment, AI amplifies existing weaknesses instead of compensating for them.

A related study on ad writing sharpens the point. Using AI as a sounding board helped non-experts, but using it as a ghostwriter delivered no benefit — and actually reduced the performance of expert writers. Confident, uncritical delegation is the failure mode; skeptical engagement is the success mode.

The data also suggests AI's benefits skew toward speed rather than quality. In a nationally representative Pew Research survey, 40% of chatbot users said AI was extremely or very helpful for working faster, but only 29% said the same about improving work quality. Fast output isn't the same as good output.

There are structural reasons for skepticism. As Scientific American explains, where a human judges, a model correlates; where a human evaluates, a model predicts. LLMs pattern-match their way to plausible answers, cannot know when they are hallucinating, and cannot check their output against the real world. Fluency is not truth.

This matters acutely in spreadsheet work, where errors hide in plain sight. Microsoft's own guidance warns that when mistakes are "subtle or hidden — like incorrect formulas in a spreadsheet," human-led validation is necessary before output is trusted — and stresses that "delegating work to AI doesn't transfer accountability." An AI-generated formula can look correct and still be wrong.

For teams extending AI reliance into agent workflows, the same verification logic applies to the skills and tools those agents use. SkillzBench is building free sandbox benchmarking that tests agent skills with and without each skill enabled, publishing the Lift in percentage points so you can see what actually helps. You can join the waitlist at skillzbench.com to be first in line when verification goes live.

The pattern across the evidence is consistent: AI rewards people who critique it and punishes people who trust it blindly.

The Spreadsheet Trap: Where Confident AI Errors Hide in Plain Sight

There's a particular kind of spreadsheet danger that doesn't announce itself. An AI-generated formula can look polished, calculate without errors, and still be quietly wrong — and because a single formula often feeds a dozen downstream cells, one mistake cascades through an entire workbook before anyone notices.

Microsoft's own guidance for Copilot and agents makes this risk explicit. "Delegating work to AI doesn't transfer accountability," the company states, and it specifically flags incorrect spreadsheet formulas as the kind of subtle, hidden error that demands human-led validation before output is trusted or reused. The work produced by Copilot or an agent, Microsoft reminds users, is still your work.

This matters because spreadsheets are exactly where confident over-reliance thrives. The output is numeric, which feels authoritative. The formula bar shows something plausible. And unlike a hallucinated fact in a paragraph, a wrong SUMIFS or an off-by-one range reference doesn't look like an error — it looks like a result. Research on how large language models work explains why: where a human judges, a model correlates and predicts, pattern-matching its way to an answer without any ability to check that answer against reality. An LLM cannot even know when it is hallucinating.

Microsoft offers a practical decision lens for deciding what to delegate to AI, built on four criteria:

• Repeatability — is this a task you'll perform once, or over and over where a validated approach pays off?
• Impact — what happens downstream if the output is wrong?
• Error detectability — would a mistake be obvious, or subtle and hidden, like a bad formula?
• Time sensitivity — does a deadline pressure you into skipping verification?

Run a financial model or a reporting workbook through that filter and the answer is uncomfortable: high impact, low error detectability. That combination, per Microsoft's framework, is precisely where human-led validation or manual handling is necessary.

The stakes are rising as AI use spreads. Frequent AI use among white-collar employees has climbed to 27%, up 12 points in a single year, according to Gallup's workplace research — yet only 22% of employees say their organization has communicated a clear AI plan. That gap means many spreadsheets are now being built with AI assistance under no verification standard at all.

The same discipline applies when you extend AI into automated workflows. Published research found roughly 26% of agent skills carry at least one vulnerability, which is why SkillzBench is building free benchmarking that tests every skill in a sandbox — with and without the skill — and publishes the Lift in percentage points. If you're handing more of your spreadsheet work to agents, browse and test free agent skills on SkillzBench and join the waitlist to be first in line when verification goes live.

How to Work (and Manage) Alongside AI-Reliant Coworkers

Working next to someone who lets AI run most of their job doesn't have to mean bracing for their mistakes. The research points to a repeatable playbook — one that turns confident AI reliance into productive, validated collaboration.

Start with a mindset shift: AI output is a draft to critique, not a fact to accept. A Dickinson College / Stevens Institute study found that people working with AI performed better when prompted to critically analyze its output rather than treat it as truth, and a University of Minnesota / UT Dallas study showed AI helped as a sounding board but actually reduced performance when used as a ghostwriter. Skeptical engagement is the success mode.

Before handing off any task, apply Microsoft's delegation framework. The company's official guidance recommends evaluating work on four criteria — repeatability, impact, error detectability, and time sensitivity — before letting an AI agent take the lead. It also warns that when errors could be subtle or hidden, like incorrect formulas in a spreadsheet, human-led validation is necessary before the output is trusted or reused.

• Treat every AI-generated spreadsheet formula as unverified until spot-checked against known values.
• Run a validation checkpoint before AI-assisted work feeds into reports, forecasts, or client deliverables.
• Ask AI-reliant coworkers to explain their reasoning, not just their output — judgment is the skill that separates gain from loss.
• Flag high-impact, low-detectability tasks for manual handling, per Microsoft's criteria.

For managers, closing the governance gap is the highest-leverage move. Gallup found that only 22% of employees say their organization has communicated a clear AI plan, even though 44% say AI integration has already begun — and Federal Reserve researchers note that 29% of workers use AI without management even knowing. Gallup's data shows employees who strongly agree leadership has a clear AI plan are three times as likely to feel very prepared to work with AI.

Training matters more than tool access. A field experiment with 640 entrepreneurs found high performers gained roughly 10–15% in profits with AI access while low performers declined about 8% — the difference was which AI recommendations they chose to follow. If your team is extending AI into agent workflows, published research has found roughly 26% of agent skills carry at least one vulnerability, which is why SkillzBench is building free sandbox benchmarking that tests each skill with and without it, publishing the measured Lift alongside security and quality checks. Skills, proven.

Trusting AI Tools Requires Proof — Extend That Discipline to Agent Skills

If you wouldn't sign off on a spreadsheet formula you didn't check, you shouldn't install an AI agent skill you haven't verified. The same skepticism that protects your work applies — arguably more urgently — to the next layer of AI reliance: agents running entire workflows on downloaded skill packages.

The stakes are real. Published research found that roughly 26% of agent skills carry at least one vulnerability, yet these packages are being installed into Claude Code, Codex, Cursor, Gemini CLI and chat-app agents with little scrutiny. An agent skill is a set of instructions your AI executes on your behalf — and Microsoft's guidance on AI delegation is blunt: "Delegating work to AI doesn't transfer accountability." If an agent produces a subtly wrong formula or a flawed workflow, it's still your work.

The problem is that nobody can currently tell which skills actually help and which are risky. That's the gap SkillzBench is being built to close. As a free, independent directory of AI agent skills, it applies the verification discipline this article has been arguing for:

• Verify — every skill is pinned to a specific version, license-checked, and security-scanned before it earns a place in the directory.
• Benchmark — each skill is tested in a sandbox with the same model, with and without the skill, and the performance Lift is published in percentage points so you can see what it actually adds.
• Equip — a free MCP connector lets you toggle exactly which skills and tools each agent can see, per agent.
• Stay current — upstream changes are tracked, and skills are re-scanned and re-benchmarked when they change.

Trust scores are computed from security, quality, license, maintenance and community signals — and they're kept strictly separate from advertising. Ad revenue never touches scores or listings, which matters because the whole point is evidence over marketing.

This discipline echoes the broader research on AI at work. In a field experiment with 640 entrepreneurs, high performers gained roughly 10–15% in profits and revenue with AI access while low performers declined by about 8% — the difference was judgment about which advice to follow. Verification infrastructure is how teams build that judgment into their tooling rather than hoping each individual has it.

If your team is extending AI reliance into agent workflows, the smart move is to demand proof before trust. SkillzBench is pre-launch and actively building its verification and benchmarking pipeline — join the waitlist at skillzbench.com to be first in line when verified, benchmarked skills go live, before your team's AI reliance outpaces its safeguards.