88% of enterprises now use or test AI agents, but most still need a human in the loop

Zapier's 2026 AI Agents survey of over 500 enterprise leaders found that 88% of organizations are using or testing AI agents, up from 72% in 2025, and that production use rose from 40% to 71% in one year. Only 30% of leaders at organizations using AI agents say the technology completes more than 60% of assigned tasks without human intervention, and 90% report at least one positive outcome from AI agents, with nearly half reporting or expecting payback within a year.

A robot vacuum should be a simple delegation. The floor is dirty. The robot cleans the floor. I go do literally anything else. Instead, the promise of automated cleaning comes with a surprising number of supervisory responsibilities: rescue it from a charging cable, move a sock from its path, ponder why the bathroom threshold is apparently an impassable geological formation. The floor gets cleaner. I remain involved. Enterprise AI agents are roughly in the same spot. In Zapier's 2026 AI Agents survey, 88% of enterprise leaders say their organizations are using or testing AI agents, up from 72% in 2025. But only 30% of leaders at organizations using AI agents say the tech handles a majority of their tasks without human intervention. For the second year, Zapier surveyed over 500 enterprise leaders about how their organizations use AI agents, where they're seeing results, and what's keeping them from giving those agents more autonomy. Here's how things have changed since last year. Table of contents: - AI agents in production jumped from 40% to 71% in one year ai-agents-in-production - Adoption is up, but autonomy is limited: most AI agents still need a human to finish the job limited-autonomy - 90% report at least one positive outcome from AI agents, and nearly half report or expect payback within a year positive-outcomes - Enterprises plan to increase investment in AI agents in 2027, but security and accuracy concerns linger increase-investments AI agents in production jumped from 40% to 71% in one year In 2025, 40% of enterprise leaders said their organizations had AI agents in production, while 32% were still piloting or testing them. In 2026, production use reached 71%, and testing fell to 17%. Enterprises may still have reservations about agents, but most now have them doing real work. AI agent use also grew across every department, with operations and customer support remaining the largest users. Finance, the furthest behind in 2025, made the biggest jump +17 points , followed by HR +13 points . The most common uses stayed consistent: document analysis and summarization, along with data extraction and entry, led in both years. | AI agent use grew in every department | 2025 | 2026 | Change | |---|---|---|---| | Finance | 24% | 41% | +17 pts | | HR | 30% | 43% | +13 pts | | Operations | 47% | 59% | +12 pts | | Sales | 26% | 37% | +11 pts | | DevOps / Engineering | 35% | 44% | +9 pts | | Customer support | 49% | 57% | +8 pts | | Marketing | 31% | 39% | +8 pts | This wider adoption across departments doesn't mean finance has handed an AI agent the company credit card or HR has automated every consequential personnel decision. Department-level adoption tells us where agents are being used, not everything they're allowed to do. For that, we need to look at autonomy. Adoption is up, but autonomy is limited: most AI agents still need a human to finish the job There's a lot of human labor hiding in the phrase, "we use AI agents." Only 30% of leaders say their agents complete more than 60% of assigned tasks without human intervention. Most organizations also haven't connected their agents to broader automated workflows multi-step processes completed mostly by software . And the difference is stark. When more than half of an organization's agents are part of broader automated workflows, 77% of leaders say their agents complete more than 60% of tasks without human intervention. When half or fewer agents are connected to broader workflows, that drops to 8%. Consider how this could look in action: an agent reads a new support ticket, checks the customer's account information, and classifies the issue. Then the rest of the automated workflow takes over. Predefined rules send billing disputes to a person for review and route routine requests to the right queue. The agent handles the judgment call, and the workflow handles everything around it, including when to bring in a human. With those checkpoints built in, the agent can work on its own without someone watching every step. That's how agentic workflows https://zapier.com/blog/agentic-workflows/ work in Zapier. You can add AI steps directly into a Zap workflow with AI by Zapier https://zapier.com/ai . Human oversight happens at the checkpoints you set, not whenever someone notices something that looks weird. 90% of leaders report at least one positive outcome from AI agents, and nearly half report or expect payback within a year Despite the limits on autonomy, 90% of leaders say their organizations have seen at least one positive outcome from AI agents. The financial picture is encouraging, too. Of leaders whose organizations use AI agents, 46% say their most mature agent has paid for itself or is expected to within 12 months of deployment. Enterprises plan to increase investment in AI agents in 2027, but security and accuracy concerns linger In the 2026 survey, 90% of leaders say their enterprises are certain or likely to increase investment over the next year, up from 84% in 2025. That signals continued appetite for AI agents heading into 2027. It doesn't mean every budget has been approved, but enthusiasm hasn't evaporated after contact with production environments. But the concerns haven't evaporated either. When asked what holds organizations back from adopting AI agents, leaders most often cite data privacy, security, and regulatory compliance https://zapier.com/automations/legal/compliance-management risks 27% , followed by accuracy, reliability, and errors 24% . One way to limit failures is to pair AI with deterministic automation https://zapier.com/blog/deterministic-ai/ : predictable, rules-based steps that run the same way every time. In Zapier, an AI step can interpret an ambiguous request while the rest of the Zap controls where information goes, which actions need approval, and what happens when required data is missing. That doesn't make the AI infallible. It just gives the surrounding process fewer opportunities to turn one mistake into six. And since Zapier logs every Zap run, teams get a built-in record of errors and exceptions that most teams aren't tracking yet. Give AI agents more responsibility with Zapier Your agents are probably already doing some useful work. The real question is how much you let them do on their own. Give them a clear process to work from, and that question gets a lot less scary. Add Zapier to your AI assistant, like Claude or ChatGPT, then describe what should happen: what kicks it off, which apps are involved, and where AI needs to make a call. Your assistant drafts the Zap and deploys it to Zapier. The Zap kicks off when something happens in your apps or on a schedule and runs the same steps every time, bringing in AI only for the steps that need judgment. Every run is logged, and if something breaks, Zapier finds the problem and proposes a fix. It all works across 9,000+ apps https://zapier.com/apps . You choose which actions it can take, like drafting an email without sending it. And every action runs through Zapier, so your credentials never reach the model. For an added layer of protection, AI Guardrails https://zapier.com/apps/ai-guardrails-by-zapier-ca231637/integrations checks AI-generated text for personal data, prompt-injection attempts, and toxic language before it proceeds to the next step. If it flags something, your workflow can stop or route the item to a person for review. This keeps you in control without micromanaging your AI agent at every step. Related reading: - What are agentic workflows? https://zapier.com/blog/agentic-workflows/ - How to build safe and trustworthy AI agents with Zapier https://zapier.com/blog/safe-trustworthy-ai-agents/ - What is AI orchestration? A guide to intelligent systems https://zapier.com/blog/ai-orchestration/ - Zapier MCP: Perform tens of thousands of actions in your AI tool https://zapier.com/blog/zapier-mcp-guide/ Methodology The 2026 and 2025 surveys were both conducted by Centiment for Zapier. The 2026 survey was fielded between September 16 and September 21, 2026, and is based on 513 completed surveys. The 2025 survey was fielded between October 23 and October 27, 2025, and is based on 514 completed surveys. Respondents for both surveys were U.S. managers and above, from managers to C-level executives, at organizations with at least 1,000 employees who have used AI professionally. The data for both surveys are unweighted, and the margin of error is approximately ±4% for the overall sample at a 95% confidence level. This article was originally published in December 2025. The most recent update, with new survey data, was in November 2026.