How are people using AI/LLM in their work life??

The Adoption Gap: Everyone's Trying AI, Few Are Living In It

Here's a puzzle worth sitting with: 41% of U.S. workers now report using generative AI for work, yet only 12% touched it daily in the past week. That gap — between "I've tried it" and "I can't work without it" — may be the single biggest opportunity in professional AI right now.

The Federal Reserve's adoption tracking shows 35.2% of workers used GenAI at least once in a recent week, but most engagement stays occasional. Among firms whose employees use AI, 35% spend less than an hour a week on it. That's experimentation, not integration.

Adoption also isn't evenly spread across the economy. It clusters where the work is cognitive and analytical: professional services shows roughly 62% work-related GenAI use and finance around 63%, with finance posting sustained 127% year-over-year growth. If you work in spreadsheets, documents, code, or analysis, your peers are likely already using AI — just not deeply.

One honest caveat before you internalize any single number: adoption estimates range from 18% to 78% depending on methodology. The Census Bureau's firm survey counts 18% of U.S. firms as AI adopters, while the Fed's worker survey finds 41% of individuals using GenAI for work, and a third measure puts 78% of the labor force at AI-adopting firms. Different units of analysis, different answers — all pointing in the same upward direction.

What does "deeper" adoption actually look like? IBM researchers found workers using LLMs for four things: creation, finding and working with information, getting advice, and automation. Today's use is dominated by the first — generating code, improving text. But the same research found workers want exactly what most haven't achieved yet: deeper workflow integration that uses their own context and data.

The headroom is real. If only 12% of workers use AI daily, then roughly 29% have tried it and drifted back to occasional use — often because a chatbot in a separate tab doesn't fit how work actually flows. Closing that gap is less about learning better prompts and more about embedding AI into the tools and steps you already use.

That's also where agent skills come in: instruction packages that extend agents like Claude Code, Cursor, and Gemini CLI into multi-step workflow automation. SkillzBench is building a free directory and benchmarking platform to help you find which skills actually lift your workflow — and a free MCP tool to connect the ones you choose to your agents. The waitlist is open now if you want to move from occasional user to daily practitioner.

• 41% of U.S. workers report work-related GenAI use — but only 12% use it daily
• Professional services (~62%) and finance (~63%) lead all sectors in work-related use
• Adoption estimates span 18% to 78% depending on what's measured: firms, workers, or workers at adopting firms

The Four Things People Actually Do With AI at Work

Strip away the hype, and the answer to "what do people actually do with AI at work?" is surprisingly tidy. Researchers at IBM surveyed knowledge workers across two waves and found that work LLM use falls into just four categories: creation, finding and working with information, getting advice, and automation — and today, creation dominates (IBM Research).

Creation means generating or improving artifacts: writing code, polishing text, drafting documents. The evidence that this is the mainstream use case is measurable at scale. A peer-reviewed study analyzing over 537,000 corporate press releases and 687,000 consumer complaints found that by late 2024, 24% of corporate press releases and roughly 18% of financial complaints showed signs of LLM-assisted writing (Patterns, Elsevier). Adoption surged rapidly after ChatGPT's release, then stabilized by 2024.

The named-company examples reinforce the same pattern. At automotive supplier Valeo, which rolled out Gemini to its entire 100,000-person workforce, more than 35% of code is now generated by AI. At Geotab, which analyzes billions of data points daily from millions of connected vehicles, teams use Gemini in Workspace for research, summarization, status reporting, legal document review, and data filtering — a blend of creation and information work.

The four categories look like this in practice:

• Creation: generating code, drafting and improving text, producing new artifacts and ideas — Valeo's AI-assisted code generation is the flagship example.
• Finding and working with information: summarizing documents, filtering data, reviewing legal agreements — Geotab's day-to-day Gemini use spans exactly these tasks.
• Getting advice: asking for guidance, validation, and new perspectives — IBM found this skews toward personal use, sometimes even on high-risk topics.
• Automation: handing multi-step tasks to AI with minimal intervention — the smallest category today, but the fastest-growing frontier.

Notably, IBM's survey found that workers who used LLMs for actual work rose from 24.5% in summer 2023 to 34.6% by fall 2024, while experimentation-only usage stayed flat at around 27%. People aren't just playing with AI anymore — they're folding it into real deliverables. And workers themselves say they want more: novel insight generation, deeper workflow integration using their own context and data, and task automation.

That last desire points to where this is heading. If most professional AI use today is one-off creation — a draft here, a summary there — the next step is chaining those steps into automated workflows. Agent skills (SKILL.md instruction packages for tools like Claude Code, Cursor, and Gemini CLI) are one way builders make that leap, and SkillzBench is building a free directory where you can browse, verify, and benchmark them before trusting one with your workflow. The waitlist is open now.

From Chatbot to Teammate: The Agentic Shift

The next wave of workplace AI looks less like a chat window and more like a colleague who never clocks out. Google Cloud now catalogs 1,302 real-world gen AI use cases and describes a decisive shift "from passive assistant to active team member" — what it calls the era of the agentic enterprise. Unlike single-task chatbots, these agents can plan, reason, call external tools, retain memory across interactions, and complete multi-step tasks with minimal human intervention.

The money is following the vision. According to IBM's research, 80% of executives are increasing their investment in agentic AI, with spending projected to nearly triple by 2027. Three-quarters of business leaders expect different agents to handle different problems, producing richer solutions than any single tool could.

The early results are striking — with an important caveat. These are vendor-curated success stories, not representative averages, so treat them as illustrative rather than typical:

• Thomson Reuters reports a one-third average reduction in time spent on document review, research, and drafting.
• Toyota's AI platform for factory workers saved more than 10,000 man-hours per year.
• Valeo now generates over 35% of its code with AI, rolling out Gemini to its entire 100,000-person workforce.
• Colombian logistics firm Domina improved real-time data access by 80% and eliminated manual report generation entirely.

For readers in data analysis and BI, the agentic shift has a concrete meaning: agents can detect data issues, generate dashboards for human review, and flag unusual patterns — turning data quality from a periodic audit into continuous monitoring. Organizations are also using natural language interfaces to query decades-old SAP instances and COBOL codebases, letting non-technical staff get answers without filing IT tickets.

So how do you actually extend an agent to handle your workflow? That's where agent skills come in. These are SKILL.md instruction packages that plug into tools like Claude Code, Codex, Cursor, and Gemini CLI, teaching agents new capabilities. It's the practical bridge from "I use a chatbot" to "an agent runs this workflow for me" — the deeper integration that workers themselves say they want, per IBM Research's survey of knowledge workers.

One caution before you install the first skill you find: published research suggests roughly a quarter of these packages carry at least one vulnerability. SkillzBench is building a free, independent directory that security-scans every skill, pins versions, and benchmarks each one in a sandbox so you can see its measured Lift before trusting it with your workflow. Its free MCP connector lets you attach only the skills you choose to any agent. The waitlist is open now — a sensible first step if agentic AI is moving onto your team.

Making AI a Daily Habit: Your Implementation Path

Here's the gap worth closing: 41% of U.S. workers now report using generative AI for work, but only 12% used it daily, according to Federal Reserve data. Most professionals are occasional users with significant untapped headroom. Turning AI from a sometimes-tool into an embedded workflow is a deliberate process, and it follows a repeatable path.

Start with the four categories IBM researchers identified for how knowledge workers actually use LLMs: creation, finding and working with information, getting advice, and automation. Current usage concentrates on creation tasks — generating code, improving text, learning new information — per an IBM Research survey. Pick one recurring task in each category and commit to it daily. If you write reports, draft every first pass with AI. If you analyze data, route every "quick question about the numbers" through a chatbot first. Repetition, not variety, is what builds the habit.

The second step is removing friction between AI and your existing systems. Organizations are building natural-language interfaces on top of legacy infrastructure — 40-year-old SAP instances, mainframes, even COBOL codebases — so non-technical staff can query complex data without waiting on IT. Domina, a Colombian logistics company, improved real-time data access by 80% and eliminated manual report generation entirely. You don't need a full modernization project to start; even connecting a chat interface to one internal data source removes the copy-paste friction that keeps AI use occasional.

The third step is extending agents with skills. Agent skills — instruction packages in SKILL.md format for tools like Claude Code, Codex, Cursor, and Gemini CLI — let you customize what an agent can do, moving from single-task chat toward the multi-step workflow orchestration that Google Cloud's catalog of 1,302 real-world use cases describes as the shift "from passive assistant to active team member." The economics support going deeper: the inference cost of GPT-3.5-level performance dropped more than 280-fold between November 2022 and October 2024, per the Stanford HAI AI Index. Running an agent all day is no longer the cost barrier it once was.

But handing more workflow steps to agents raises the stakes on trust. Published research found roughly 26% of agent skills carry at least one vulnerability, and Stanford HAI notes models still struggle with high-stakes precision tasks. Before installing a skill, verify what it actually does: check its version history, scan for security issues, and benchmark whether it improves outcomes. That verification layer is exactly what SkillzBench is building — security scanning, version pinning to specific commits, and sandbox benchmarking with published Lift scores, all free, with the waitlist open now.

The path, in short:

• Choose one recurring task per use category and apply AI to it every day.
• Connect natural-language interfaces to legacy data systems to remove friction.
• Extend agents with skills once the basics are habitual.
• Verify every skill before trusting it with real workflow steps.

The professionals pulling ahead aren't using more AI — they're integrating it more deeply, one verified habit at a time.