What companies use agentic AI?

The Adoption Illusion: 79% Say Yes, But Few Go Deep

Ask ten executives whether their company uses agentic AI, and most will say yes. But dig one level deeper and the confident answers start to wobble: using an agent is not the same as being transformed by one, and the numbers prove it.

In a PwC survey of 308 US executives, 79% said AI agents are already being adopted in their companies. Yet independent analysis of the enterprise market found that only around 2% of agentic AI projects reach at-scale deployment — a spread one researcher called "the defining statistic of this market." The same PwC data shows 68% of respondents report that half or fewer of their employees actually interact with agents in everyday work, and 18% say their companies aren't using agents at all.

So what explains the gap? PwC's own interpretation is blunt: reports of full adoption often reflect excitement about what agentic capabilities could enable, not evidence of widespread transformation. In practice, much of what counts as "adoption" is employees using agentic features built into existing enterprise apps to speed up routine tasks — capabilities that MIT Sloan's analysis notes are being distributed at scale by Microsoft, Salesforce, Google, and IBM, embedded directly into the software companies already run.

That distinction matters for anyone evaluating where their organization actually stands:

• Only 35% of adopters describe their rollout as broad, and just 17% say agents are fully adopted across almost all workflows and functions.
• Fewer than half of executives are fundamentally rethinking operating models (45%) or redesigning processes around AI agents (42%).
• The deepest adopters — the 17% — are the exception, not the rule, despite 73% believing agents will give them a competitive advantage within 12 months.

The practical takeaway: when a vendor or case study claims "agentic AI adoption," ask what's actually happening behind the number. Is it a feature toggle in a CRM, a handful of pilot projects, or a redesigned workflow where agents handle real operational load? Most of the time it's the first two.

This is also why the tooling layer matters more than the headline adoption figures. As companies move from embedded features to purpose-built agents, the real work becomes choosing which capabilities an agent should have and verifying they're safe to use. SkillzBench is building a free, independent directory of AI agent skills — version-pinned, security-scanned, and benchmarked so teams can see which skills actually lift performance before wiring them into an agent. For readers navigating the gap between adoption claims and real deployment, that verification layer is where shallow adoption starts becoming deep impact.

Who's Actually Using It: Leading Industries and Named Companies

Seventy-nine percent of executives say AI agents are already on the job at their companies, according to PwC's survey of 308 US business leaders — but the honest answer to "who's actually using agentic AI?" is more nuanced than a simple company list.

Three sectors consistently lead in both adoption and spend: financial services (BFSI), technology, and healthcare, according to independent market analysis. IBM separately names healthcare, finance, HR, and software development as the verticals adopting agentic workflows. But the industry label isn't what matters — it's the operating conditions behind it.

As one analysis puts it, the real signal is whether a company has structured workflows, repeatable decisions, clear compliance boundaries, and enough transaction volume — not which vertical it happens to sit in. That's why BFSI leads: "compliance forces data discipline," and the same constraint that once slowed cloud adoption in banking now accelerates agent ROI. Regulated firms move slower to production, but their wins are more durable because governance work happens first.

The verifiable named examples are fewer than listicles suggest — most research anonymizes adopters as "a leading hospitality company" or "a major retail company." The companies that are named include:

• JPMorgan Chase, Walmart, and Microsoft — cited in MIT Sloan's coverage alongside Salesforce, Google, and IBM, with use cases spanning fraud detection, loan approvals, personal shopping, and customer service
• Wyndham — deploying AI agents to improve franchise operations and customer service, per PwC's research
• Riyadh Air — building "the world's first AI-native airline" with IBM, and Comparus — a conversational banking assistant on watsonx.ai, per IBM's case studies

By function, adoption clusters where the volume is: customer service leads at 57%, followed by sales and marketing at 54% and IT and cybersecurity at 53%. If your own workflows share those operating conditions — repeatable decisions, clear boundaries, high volume — agent skills are the practical way to test agentic use cases before committing. SkillzBench is building a free, independent directory where every skill gets security-scanned and benchmarked, so teams can match use cases to capabilities with evidence rather than hype.

Where Agents Earn Their Keep: Top Use Cases and Measured Value

Agentic AI is delivering measurable results where companies have learned to match the technology to the right business problem. The function-level adoption data tells a clear story about where autonomous systems are pulling their weight today.

Customer service leads the pack at 57% adoption, followed closely by sales and marketing (54%) and IT and cybersecurity (53%), according to PwC's 2025 AI Agent Survey. These front-line and back-office functions share something critical: high transaction volume, structured workflows, and clear decision boundaries. That combination is what makes them fertile ground for agents that can act without constant human oversight.

The reported outcomes are striking:

• 66% of adopters report increased productivity
• 57% report cost savings
• 55% report faster decision-making
• 54% report improved customer experience

Healthcare offers one of the most concrete examples. Clinical documentation agents reduce documentation time by 30–42%, saving clinicians up to 66 minutes per day, per Keyhole Software's market analysis. That time goes back into patient care, not paperwork.

Multi-agent orchestration is emerging as the next frontier. Rather than single agents handling isolated tasks, companies are deploying systems of agents that coordinate across functions like finance, customer service, and R&D. PwC describes this as building an "operating system" for agentic organizations, one that orchestrates multiple agents from multiple vendors across complex business processes.

Yet trust remains a significant constraint. Only 20% of executives trust agents with financial transactions, while 38% trust them with data analysis and 35% with performance improvement, per the same PwC survey. This trust gradient shapes which use cases scale first and which stay in pilot mode.

The gap between adoption and transformation is real. Gartner projects over 40% of agentic AI projects will be canceled by end of 2027 due to escalating costs, unclear value, or inadequate risk controls, according to Keyhole's analysis. The bottleneck is rarely model capability. It's integration, governance, and operational fit.

For teams exploring which agentic use cases fit their industry, the practical next step is understanding which agent skills can be trusted to execute those workflows. SkillzBench is building a free, independent directory where every AI agent skill is version-pinned, security-scanned, and benchmarked with a published Lift, plus a free MCP connector that lets any agent see only the skills you choose. It's a way to move from use-case inspiration to operational reality without flying blind on quality.

Why 40% of Projects Will Fail — and What Winners Do Differently

The agentic AI boom has a failure rate that should give every executive pause. According to Gartner, over 40% of agentic AI projects will be canceled by the end of 2027, killed by escalating costs, unclear business value, or inadequate risk controls. That projection cuts against the 79% adoption headline that dominates vendor decks.

The disconnect is structural. MIT Sloan research on a cancer patient adverse-event detection agent found that 80% of the implementation work was consumed by data engineering, stakeholder alignment, governance, and workflow integration — not model capability. The bottleneck is rarely the AI. It is everything around it.

Winners treat governance as a prerequisite, not an afterthought. Keyhole Software's analysis notes that regulated clients "move slower to production, but their wins are more durable because the governance work was done first." In banking, compliance forces data discipline — the same constraint that slowed cloud adoption now accelerates agent ROI once projects ship.

The pattern repeats across high-function deployments:

• Structured workflows with repeatable decisions and clear compliance boundaries
• High transaction volume that justifies integration overhead
• Phased rollout that starts narrow and expands after governance proves out
• Cross-agent orchestration rather than isolated pilots in single functions

PwC's survey found that top-ranked barriers were cybersecurity and cost (34% each) — what the firm calls "mostly... safe excuses." The real challenges ranked lowest: connecting agents across applications (19%), organizational change (17%), and employee adoption (14%). Companies comfortable naming budget concerns often avoid the harder conversation about redesigning work itself.

Kate Kellogg, a professor at MIT Sloan, frames the trap precisely: "Just because an agentic AI model reclaims 20% of someone's time, that doesn't mean it's a 20% labor-cost savings." Time reclaimed without process redesign becomes absorbed by other tasks. Value requires re-engineering the workflow, not just inserting an agent into it.

For teams exploring which agentic use cases fit their industry, the practical next step is testing skills in controlled environments before committing to production builds. SkillzBench is building a free, independent directory where agent skills are version-pinned, security-scanned, and benchmarked with published Lift scores, plus a free MCP connector that lets any agent see only the skills you choose — a way to validate capabilities before they touch regulated workflows.

From Reading to Doing: How to Match Use Cases to Your Business

The companies that succeed with agentic AI share a pattern that has nothing to do with industry labels. As one analysis of the enterprise agentic AI market puts it, the real question is whether a business has structured workflows, repeatable decisions, clear compliance boundaries, and enough transaction volume — the "operating conditions" that make agents pay off source. Before you pick a vendor or a platform, audit your own operations against those four conditions.

Start with one high-volume workflow. The most common entry points are customer service (57% of adopters), sales and marketing (54%), and IT and cybersecurity (53%), according to PwC's survey of US executives. Choose a process that is repetitive, rules-based, and measured today — if you cannot currently quantify how long it takes, you will never know whether an agent improved it.

Then pressure-test the agent before it touches production. Gartner projects that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear value, or inadequate risk controls source, and in one real deployment, 80% of the work turned out to be data engineering, governance, and workflow integration rather than model tuning source. Vetting what you plug into your agent is where most teams skip steps — and where risk creeps in.

A practical vetting checklist looks like this:

• Verify provenance: know exactly which version of a skill or tool your agent is running, pinned to a specific commit.
• Scan for security issues before deployment — published research found roughly 26% of agent skills carry at least one vulnerability.
• Benchmark impact: test the agent with and without each skill under identical conditions, and demand a measurable Lift in percentage points.
• Limit scope: your agent should only see the skills and tools you explicitly choose, nothing more.

That last point matters because trust, not technology, is the binding constraint. PwC found that executives' trust in agents drops sharply by task — from 38% for data analysis down to just 20% for financial transactions source. Scope control is how you earn that trust incrementally.

This is precisely the gap SkillzBench is being built to close. It is a free, independent directory of AI agent skills where every skill is version-pinned, security-scanned, and benchmarked with a published Lift — paired with a free MCP connector that lets any agent see only the skills you choose. SkillzBench is pre-launch; you can join the waitlist now to be first in line when the directory and connector go live.