What are the challenges of AI in the workplace?

The Adoption Reality: Soaring Pressure, Stalled Projects

Here's a paradox that should stop every executive planning an AI rollout: 97% of leaders report increased urgency to deploy AI, yet more than 80% of AI projects fail — twice the failure rate of non-AI IT projects. The pressure is real, and so is the carnage behind it.

A RAND study based on 65 interviews with AI practitioners found managers under "enormous pressure to do something—anything—with AI" just to demonstrate they're keeping up. But urgency isn't readiness. Only 14% of organizations report being fully prepared to integrate AI into their operations.

The root causes of failure are stubbornly unglamorous. When RAND researchers dug into why projects collapse, the answers had little to do with the technology itself:

• Miscommunication about the problem to be solved — the single most common cause of AI project failure
• Insufficient or unusable data to train and run systems on
• Technology-first thinking that fixes the tool before defining the user's problem
• Inadequate infrastructure to support deployment at scale
• Applying AI to problems too difficult for even the most advanced models to solve

As RAND's researchers put it, "AI is not a magic wand that can make any challenging problem disappear." They recommend committing to an enduring problem for at least a year — and if a project isn't worth that commitment, it probably isn't worth committing to at all.

Measuring adoption is its own mess. A Federal Reserve review of 16 surveys found firm-level AI adoption estimates ranging from 5% to 40%, with worker-level estimates between 20% and 44%. The spread comes from question phrasing, lookback windows, and weighting methods — not from disagreement about reality. What all surveys do agree on is direction: annualized growth rates run from 38% to 145% depending on the source.

Then there's the quiet failure mode nobody celebrates: de-adoption. The Fed's review documents firms that experiment with AI and then abandon it entirely. Pilots launch with fanfare, stall against unclear goals or weak infrastructure, and quietly get shelved — leaving behind skepticism that makes the next initiative even harder.

For business leaders, the takeaway is uncomfortable but clarifying. The gap between 97% feeling urgency and 14% being ready is where most AI budgets go to die. The organizations that succeed treat AI as a change-management problem first and a technology problem second — defining the problem precisely, securing the data, and committing for the long haul before a single model gets deployed.

Shadow AI and the Hidden Security Blind Spot

Your employees are almost certainly using AI right now — and there's a good chance you don't know about it. That invisible usage, combined with a new generation of autonomous agents holding broad permissions, is quietly becoming one of the biggest security exposures in the modern workplace.

According to a Federal Reserve review of workplace AI surveys, 29% of AI-using workers say their managers don't know they use AI at all. This "shadow AI" means confidential data may be flowing into unvetted tools, with no oversight, no approved-vendor list, and no incident response plan. Ethisphere's governance research notes that most employees don't set out to misuse AI — they simply fail to imagine how a tool could create risk, uploading confidential information or trusting unvalidated output without realizing the implications.

The problem gets sharper with agentic AI. Security researchers at IBM X-Force describe a "lethal trifecta" of agentic risk: deep access to private local data, interaction with untrusted external content, and the ability to communicate outward. An agent that can read your files, browse the web, and send messages is a genuinely attractive target for attackers.

The agent skills ecosystem makes this concrete. A first-of-its-kind large-scale study analyzed over 31,000 skills and found 26.1% contain at least one vulnerability — including prompt injection, data exfiltration (13.3% of skills), and privilege escalation (11.8%). The theory matched reality in the ClawHavoc campaign, when attackers uploaded 1,100+ malicious skills to a community repository disguised as productivity and coding tools, several of which became among the most-downloaded packages before deploying information-stealing malware.

• 26.1% of agent skills carry at least one vulnerability, and skills bundling executable scripts are 2.12x riskier than instruction-only ones
• 5.2% of skills exhibit high-severity patterns the researchers say "strongly suggest malicious intent"
• Vulnerability disclosures are outpacing CVE assignment, leaving AI flaws invisible to standard patch-management tooling
• OWASP has launched the Agentic Skills Top 10, the first comprehensive security framework for agent skills

Traditional defenses miss this entirely. As X-Force puts it, the CVE system "was built for discrete well-defined software flaws, not autonomous systems capable of taking actions, browsing external content and chaining tools." The recommended posture: treat agentic weaknesses as system-level risks, scope permissions tightly, audit agent actions, and assume exploitation may happen before formal disclosure.

For teams equipping agents with skills, vetting before deployment is becoming essential. SkillzBench is being built for exactly this gap — security-scanning, version-pinning, and sandbox-benchmarking agent skills so you can see which ones actually help and which are risky. Skills, proven.

Bias and the Imagination Gap: Unintentional Misuse at Work

Most employees who misuse AI at work never intend to. They paste a client contract into a chatbot to summarize it, trust a confident-sounding answer without checking it, or wire up an autonomous agent over a lunch break — not because they're careless, but because they simply can't picture what could go wrong.

Ethics and compliance researchers call this the imagination gap: employees "do not set out to misuse AI" but fail to imagine how a tool could create risk, according to governance analysis from Ethisphere. The same analysis identifies the most common unintentional failures:

• Uploading confidential company information without realizing the implications
• Relying on inaccurate AI output without validation
• Using enterprise AI licenses for personal purposes
• Deploying agentic AI without proper testing — which "could create massive problems too fast for humans to detect in time"

The scale of the blind spot is measurable. A Federal Reserve review of workplace AI surveys found that for 29% of AI-using workers, management had no idea AI was being used at all. That's a governance vacuum in which imagination-gap mistakes happen entirely out of view.

The risks compound when unvetted tools enter the picture. Research analyzing over 31,000 AI agent skills found that 26.1% contain at least one vulnerability, and skills bundling executable scripts are 2.12x more likely to be vulnerable than instruction-only ones — a supply-chain risk most employees never imagine when installing a "helpful" extension, per the large-scale security study.

Bias adds another layer, especially in HR and hiring, where Ethisphere's recommended governance principles — fairness, transparency, accountability, and human oversight — exist precisely because AI-assisted decisions about people carry outsized consequences when they go wrong. An employee screening résumés with an unvalidated tool isn't malicious; they just haven't pictured the discriminatory outcome.

Here's the critical insight for policy: restriction doesn't work. "An AI policy that focuses only on restriction can miss the point," Ethisphere argues. "The goal is not to scare employees away from AI. The goal is to give them enough clarity to use it well." Because risk profiles differ sharply by function — engineering, sales, HR, procurement — a single all-employee warning can't carry the nuance. Clarity, not fear, drives responsible use. As Ethisphere puts it, good AI policy messaging "closes that imagination gap. It helps employees see the risk before they create it."

For teams equipping AI agents, that clarity extends to tooling choices: platforms like SkillzBench are being built to security-scan, version-pin, and sandbox-test agent skills before they reach your workflows, so employees don't have to imagine vulnerabilities they can't see.

Change Management: Why Static AI Policies Expire in 18 Months

Pull up your company's AI policy and check the date on it. If it was written more than 18 months ago, you're relying on what Ethisphere calls "seriously outdated governance" — a static document trying to manage a technology that reinvents itself quarterly.

The problem isn't negligence; it's structure. AI risk touches legal, privacy, cybersecurity, procurement, HR, compliance, and business teams at once. When ownership is fragmented across all of them, AI policy "can become a document everyone supports in theory and no one manages in practice." Meanwhile, the gap between policy and reality widens: Federal Reserve analysis of workforce surveys found that for 29% of AI-using workers, management wasn't even aware of their AI use.

Shadow AI thrives in that blind spot, and mostly by accident. Employees upload confidential company data without realizing the implications, trust unvalidated AI output, or deploy agentic tools without testing — creating problems "too fast for humans to detect in time." Most don't set out to misuse AI; they simply fail to imagine how a helpful tool could create risk.

Ethisphere's research-backed fix is a three-layer governance model that separates what changes slowly from what changes constantly:

• Ethical AI principles — fairness, transparency, accountability, privacy, security, and human oversight — that endure even as tools change.
• A formal AI development and use policy covering prohibited uses, approved tools, data restrictions, and output validation.
• Dynamic guidance — workflows, approval processes, and vendor review — designed to be updated far more easily than formal policy.

The third layer is where most organizations fail. Because risk profiles differ sharply by function — engineering, sales, HR, and procurement each face distinct exposures — "a single all-employee message cannot carry all of that nuance." Function-specific guidance closes the imagination gap by helping employees "see the risk before they create it," rather than scaring them away from AI entirely.

One practical example: agent skills and extensions are now a real supply chain inside your workplace, and governance rarely covers them. A large-scale security study found 26.1% of AI agent skills contain at least one vulnerability — including data exfiltration and privilege escalation patterns. Dynamic guidance should require vetting before any skill is connected to an agent. That's the gap SkillzBench is being built to address, with security-scanned, version-pinned agent skills and a free MCP connector that lets you control exactly which skills each agent can see.

Static policies expire because AI itself is dynamic. The organizations that manage change well don't write longer documents — they build governance layers that update as fast as the technology does, with clear owners at every layer.

Your Action Plan: Vet, Govern, and Measure AI Before It Fails

The gap between AI ambition and AI readiness is enormous: 97% of leaders feel urgent pressure to deploy AI, yet only 14% of organizations report being fully ready to integrate it. Closing that gap takes deliberate action on four fronts.

First, commit to a clearly defined problem — for at least a year. More than 80% of AI projects fail, twice the failure rate of non-AI IT projects, and the most common root cause is miscommunication about the problem to be solved. RAND's interviews with 65 practitioners concluded that projects should focus on enduring problems, not the technology itself, and that "if an AI project is not worth such a long-term commitment, it most likely is not worth committing to at all" (RAND).

Second, bring shadow AI into the open. For 29% of AI-using workers, management was unaware of their AI use — a governance and security blind spot. Since static policies become outdated within roughly 18 months, governance experts recommend a three-layer approach: ethical principles, a core use policy, and dynamic guidance tailored by function. Engineering, sales, and HR face different risks, so a single all-employee message cannot carry the nuance. The goal isn't restriction — it's giving employees enough clarity to use AI well (Ethisphere).

Third, treat agent skills as a supply-chain risk. A first-of-its-kind study of 31,132 agent skills found that 26.1% contain at least one vulnerability, with skills bundling executable scripts 2.12x more likely to be risky. The real-world ClawHavoc campaign proved the threat: attackers uploaded 1,100+ malicious skills to a community repository, and several became among the most-downloaded packages. Researchers call for capability-based permission systems and mandatory security vetting — meaning every skill should be scanned, version-pinned, and tested before an agent ever sees it.

Fourth, expand monitoring beyond CVE feeds. Vulnerability disclosures are outpacing CVE assignment, leaving many AI weaknesses invisible to standard patch-management tooling. IBM X-Force recommends treating agentic AI weaknesses as system-level risks, strengthening permission scoping and action auditing, and assuming exploitation may occur before formal disclosure (IBM X-Force).

A practical starting point:

• Write down the specific problem each AI initiative solves, and commit to it for at least a year.
• Survey how AI is actually being used in each department, then issue role-specific guidance.
• Require security vetting before any agent skill or extension is installed.
• Give each agent only the skills and tools it needs — nothing more.
• Monitor AI security signals from sources beyond traditional CVE feeds.

This is exactly the gap SkillzBench is being built to close: free security scanning and sandbox benchmarking for agent skills, plus an MCP connector that lets each agent see only the vetted skills you choose. If you want to be first in line when it launches, join the waitlist at skillzbench.com.