The state of AI employees in 2026 comes down to one gap. Most large companies now say they are building or deploying AI agents, yet only a small share have given those agents a defined job, clear authority, and the governance to own an outcome the way an employee does today. Adoption is broad, but accountability towards that outcome has not kept up.
We pulled the most recent primary research from McKinsey, KPMG, Deloitte, PwC, Gartner, Microsoft, and Databricks, read the methodology behind each number, and mapped it onto a single question: how far has enterprise AI moved from helping people do their existing work to owning work on its own freeing up their bandwidth? Every figure below links to its source, with the survey date and sample so you can judge it yourself.
An AI employee is a persistent, job-scoped software agent that takes in work, applies company context and policy, acts across business systems, and stays accountable for the result like a human employee. It manages an ongoing process rather than answering one prompt, and it routes decisions outside its authority or confidence to a named person with the evidence attached.
The label is not an industry standard, and vendors use it loosely. Our complete guide to AI employees covers the definition in depth, and the digital employee glossary entry covers the older term. In the context of this report, a useful question we evaluated is more closer to everyday operational reality: does the agent own a recurring job end to end, with its own identity, permissions, approval gates, and job-level measures?
No major survey asks companies how many AI employees they run. The research measures proxies: AI use in general, agent experimentation, agents at scale, multi-agent systems, and agent governance. We arrange those proxies into what we call the AI employee maturity ladder and report each rung separately, because blending them is how inflated adoption headlines get written.
The rungs come from different surveys with different populations: McKinsey and Deloitte survey global respondents, while KPMG surveys US companies with at least $1 billion in revenue. The ladder shows direction and relative scale. It is not a single funnel from one sample.
Rung | What it measures | Best current data point | Source |
|---|---|---|---|
AI in regular use | Any AI in at least one function | Nearly 9 in 10 organizations | McKinsey, 2026 |
Agents in play | Building, developing, or deploying agents | 62% of US companies with $1B+ revenue | KPMG, Q3 2026 |
Agents at scale | Scaling agents in one or more functions | 40% of large organizations, 22% of smaller ones | McKinsey, 2026 |
Multi-agent systems | Developing or running coordinated agents | 25% of US companies with $1B+ revenue | KPMG, Q3 2026 |
Governed autonomy | A mature model for agent governance | 21% of companies planning agentic AI | Deloitte, 2026 |
Read top to bottom, the table shows the shape of the market. Each rung is narrower than the one above it, and the drop is steepest at the bottom when it comes to governance. The AI employee sits at the bottom most of this tier, where they run work at scale under explicit governance.
The most consistent pattern across these sources is a mismatch between how many companies use agents and how many have decided what an agent is allowed to own. KPMG's 62% building or deploying agents sits next to Deloitte's 21% with mature agent governance. PwC's 79% adoption sits next to its finding that most companies still have half or fewer employees working with agents at all.
Our reading is that much of today's "agent adoption" consists of agentic features inside existing applications, which PwC also observes. Those features help people work faster, but a person still owns each task. The move to AI employees starts when a company assigns a recurring job, a non-human identity, and an escalation route to the software, and then measures the job instead of the model.
Microsoft's 2026 Work Trend Index surveyed 20,000 AI users in 10 countries and found that organizational factors such as culture, manager support, and talent practices explain more than twice as much of reported AI impact as individual mindset and behavior, 67% versus 32% (Microsoft Work Trend Index 2026). Only 26% of AI users say their leadership is clearly aligned on AI.
McKinsey points the same way. Nearly three-quarters of its AI high performers report fundamentally redesigning workflows because of AI, compared with about one-quarter of other respondents. High performers are also more than three times as likely to be scaling agents in most business functions.
For AI employees, this means a company that bolts an agent onto an unchanged process gets a faster step. A company that redesigns the job around what the agent does and what the human approves gets an owned outcome.
39% of McKinsey respondents expect AI to reduce their organization's total employment in the coming year, up from 32% a year earlier. But only 14% of organizations using AI report that it actually contributed to a workforce decline over the past year, less than half the share that expected one.
PwC data complicates the replacement story further. In PwC's survey, 48% of executives said they would likely increase headcount because of the changes agents bring. The evidence so far supports a narrower claim than most headlines: AI employees change which work people do before they change how many people a company employs.
About one in five McKinsey respondents say AI operating costs, including token costs, have constrained their use of AI. KPMG reports that 74% of large US companies now include cost reviews in AI approval processes, up from 61% a quarter earlier, and 43% have usage or token budgets.
This matters for AI employees because the economics of an owned job differ from the economics of a chat. A job-scoped agent can be measured on cost per completed case against the human baseline, which is the comparison a CFO will ask for.
Much of the confusion in 2026 research comes from counting different things under one word. Gartner has a name for part of the problem: "agent washing," the rebranding of assistants, RPA, and chatbots without real agentic capability. It estimates only about 130 of the thousands of vendors claiming agentic AI are genuine (Gartner, June 2025).
Attribute | Copilot | RPA bot | AI agent | AI employee |
|---|---|---|---|---|
Who starts the work | A person, each time | A schedule or trigger | A goal or prompt | Standing triggers tied to a job |
Handles variable inputs | Yes, with a person driving | No, needs a fixed path | Yes | Yes |
Takes actions in systems | Rarely, through the user | Yes, scripted | Yes, per task | Yes, across the full job |
Owns the outcome | No | No | Usually one task | Yes, the recurring job |
Identity and permissions | The user's | A bot account | Often borrowed | Its own non-human identity |
Escalation | Not applicable | Fails or stops | Varies | Named owner, evidence attached |
Measured on | User productivity | Runs completed | Task success | Job outcomes versus baseline |
An AI employee often coordinates RPA steps, deterministic checks, and specialist agents inside one job. For a deeper comparison, see AI employee vs. AI agent and AI agents vs. RPA.
Where agents run today depends heavily on industry. McKinsey reports that respondents most often scale agents in IT, knowledge management, and software engineering. In PwC's survey, more than half of companies were using or planning agents within six months in customer service (57%), sales and marketing (54%), and IT and cybersecurity (53%).
Finance is the clearest case of high interest, deliberate pace. Gartner's 2025 AI in Finance Survey of 183 CFOs and senior finance leaders found 59% of finance functions using AI, barely changed from 58% in 2024 after a jump from 37% in 2023 (Gartner via CPA Practice Advisor, surveyed May to June 2025).
The adopted use cases are telling. Among finance teams using AI, the most common applications were knowledge management (49%), accounts payable process automation (37%), and error and anomaly detection (34%). And 91% reported low or moderate impact at first, which Gartner ties to the time it takes to move from pilot to production.
Agentic AI in finance specifically is earlier still. In a Deloitte Center for Controllership poll of more than 3,300 finance and accounting professionals in January 2025, 13.5% said their organizations already used agentic AI, and 33.6% were building or planning it. Trust was the top barrier (21.3%), ahead of systems integration (20.1%) (Deloitte, July 2025).
The most useful finance number in this data set comes from the same poll: 59.7% of those finance professionals trust AI agents to make decisions only within a defined framework, with people making the judgment calls. Only 2.7% would let agents make every decision. That is close to a written description of the AI employee operating model, where authority is set by action, amount, entity, and risk, and anything outside it goes to a named approver. PwC found the same caution from the executive side: only 20% trust agents with financial transactions.
The finance jobs where AI employees are furthest along share three traits: high volume, variable documents, and a clear policy for exceptions. They include invoice processing and accounts payable, cash application and collections, reconciliation, vendor onboarding, and billing. Our guide to AI employees for finance teams covers each process in detail.
Procurement intake, order management, compliance checks, and document-heavy operations have the same profile as finance: many systems, messy inputs, and rules that humans already follow. Zamp's own published case studies, including a multi-entity billing deployment and an enterprise-wide rollout across five functions at a pharmaceutical company, show what this looks like in practice. They are vendor case studies, so read them as examples of the operating model rather than market-wide evidence.
These functions lead agent scaling in McKinsey's data. Coding agents have gone furthest: about two in ten organizations are scaling them, 31% at larger enterprises, and 32% say they decided against buying at least one software product because they could build it with agentic coding tools. Customer service agents are common too. In our read of vendor offerings, many remain closer to chat-based assistants than to AI employees that own a case from intake to resolution.
The research agrees on the main risks. Gartner names escalating costs, unclear business value, and inadequate risk controls as the reasons projects get canceled. PwC ranks trust as a top-three challenge for 28% of executives. Microsoft finds 86% of AI users treat AI output as a starting point rather than a final answer. That habit works for a copilot, but software that acts on its own needs controls built into the job instead of a person checking every output.
For an AI employee, governance needs to answer five practical questions before production:
Regulation is arriving on a longer timeline than adoption. Under the EU AI Act, obligations for general-purpose AI model providers began applying on August 2, 2025. The Digital Omnibus amendments deferred the main high-risk system deadlines to December 2, 2027, for stand-alone systems and August 2, 2028, for systems embedded in regulated products (European Parliament Legislative Train). In the US, rules remain a patchwork of state laws and sector regulators, covered in our AI employee laws in the US guide. Zamp's approach to these controls is documented in how Zamp's AI employees are governed, audited, and secured.
The published forecasts for the next two years point in two directions at once.
Our forecast is that 2027 will sort the market by accountability rather than by capability. Projects that assigned an agent a real job, with identity, authority, escalation, and job-level measures, are the ones most likely to survive Gartner's projected cancellation wave. Projects that bought agentic features and hoped for transformation are the likely casualties. We also expect "AI employee" to become a procurement category with its own evaluation criteria, as described in our enterprise AI employee platform comparison.
Zamp builds AI employees for back-office operations, with finance and accounting as the deepest focus. Each AI employee runs one defined job under an Agent Operating Procedure, with scoped permissions, human approval gates, and a case-level audit trail, as described in how Zamp's AI employees work. Deployment options are covered in on-prem, BYOC, or SaaS deployment for AI employees. Zamp is a vendor in this market, so weigh the analysis above on its sources, not on our position.
If you are a finance or operations leader, start with one recurring job that has clear inputs, policies, and an owner, baseline it, and measure the AI employee against that baseline. If you are writing about the category, cite the primary sources linked above, and please link back to this report as the aggregate view. We will update the figures as new survey waves are released and record each change in the page's update note.
