Every enterprise technology conversation right now circles back to AI agents. And for once, the excitement isn't running too far ahead of reality. According to a Zapier survey of over 500 enterprise leaders, 72% of enterprises are already using or testing AI agents, and 84% plan to increase their investment over the next 12 months.
Those numbers are big. But they also raise a question that doesn't get asked enough: what exactly are companies doing with these agents, and are they actually getting value from them?
The Adoption Numbers Are Real
This isn't a "maybe someday" technology anymore. Of the 72% with AI agents in motion, 40% already have multiple agents running in production. Not piloting. Not exploring. Running. Another 32% are in active pilots, and 14% more have adoption on the roadmap. That's 86% of enterprises with agents deployed, testing, or planned.
And where are they showing up? Customer support leads at 49%, operations is right behind at 47%, then engineering (35%) and marketing (31%). Sales and finance are earlier in the curve — 26% and 24%, respectively — but they're moving.
The most common use case won't surprise anyone who's spent time in the weeds of enterprise operations: data management. 47% of enterprises are using agents for data entry and extraction. Document analysis and summarization comes in at 41%, tied with customer support triage. In other words, the work nobody wants to do manually is exactly the work agents are getting first. That tracks.
The Gap Between Deploying and Trusting
Here's where it gets interesting. Despite all this adoption, most enterprises aren't exactly letting agents run free. Human-in-the-loop remains the most common management approach at 38% — meaning a real person reviews and approves agent output before it moves forward. Only 20% say their AI systems operate with minimal oversight.
I'd be more worried if those numbers were reversed. Agents are excellent at pattern-based work — data extraction, triage, routing. They're less good at judgment calls, edge cases, and anything that requires the full context of a business situation. Keeping humans in the loop isn't timidity. It's honesty about where the technology actually is right now.
And honestly, this is something I think about constantly in my own work. I build HR workflows that touch real people's careers — onboarding, performance, hiring. The cost of getting it wrong isn't a bad data field. It's someone's experience at their company. Human-in-the-loop isn't a limitation I'm waiting to outgrow. It's a design choice I'm making on purpose.
The biggest barrier to faster adoption? Security and data privacy, cited by 18% of leaders. That tracks with what I hear from IT leaders constantly — and what I see firsthand inside our own People team. They're not opposed to agents. They just need confidence that deploying them won't create compliance headaches or expose sensitive data. That's not resistance. That's good judgment.
Where the Money Is Going (and Why)
The investment signal is strong. 84% of enterprise leaders say they'll likely or certainly increase AI agent spending in the coming year. 36% say the increase is certain. Just 2% say they definitely won't invest more, and only 1% are pulling back.
When asked where they see the biggest opportunity, 30% of leaders pointed to routine workflow automation. Another 17% said improving customer experiences, and 12% mentioned strategy and decision-making support. Only 5% said they don't see value in AI agents at all.
Those priorities tell you something about how enterprises are thinking about agents right now. They're not chasing moonshots. They're looking at the boring, repetitive work that eats up their teams' time and asking whether an agent can handle it reliably enough to be worth the investment.
What This Means for IT Leaders
If you're a CIO or IT leader watching these numbers, a few things are worth thinking about.
First, the build-versus-buy question is genuinely open. The survey shows enterprises split fairly evenly across cloud provider platforms (53%), open-source tools (48%), enterprise AI platforms (48%), and orchestration frameworks (46%). Only 26% build custom agents from scratch. There's no dominant approach yet, so your choice should depend on your existing infrastructure and how much control you need over agent behavior.
Second, governance matters more than speed. The fact that human-in-the-loop is the most popular approach tells you enterprises are prioritizing trust over velocity. If you're rolling out agents, build approval workflows and monitoring from day one.
Third, start where the work is most repetitive and least ambiguous. Customer support triage, data entry, document summarization: these are tasks where agents perform well because the inputs are structured and expectations are clear.
What Comes Next
AI agents are moving from experiment to infrastructure. That's clear from the data. But the organizations that get the most value won't be the ones that deployed the most agents or spent the most money. They'll be the ones that matched agents to the right work, kept humans involved where judgment matters, and built-in governance from day one — not bolted it on after something broke.
The pattern I keep seeing, in our own work and in the companies we partner with, is that the teams who treat agents as tools in service of their people — not replacements for them — are the ones actually shipping results. Not pilots. Not proofs of concept. Results.
If you're figuring out where to start, start with the work that's most repetitive and least ambiguous. Customer support triage, data entry, document summarization — the stuff your team already wishes they didn't have to do manually. Build human review into the workflow from the beginning. Measure what it frees up, not just what it automates. And then let your people reinvest that time into the work that actually requires them.
That's the whole game. Not more agents. Smarter ones, in the right places, with humans still driving.