Every product keynote this year has promised some version of the same thing: an AI agent that handles the busywork on its own while employees focus on higher-value work. That promise builds on a wave of broader AI adoption most workplaces have already gone through. Agents are the next step after prompting a chatbot a paragraph at a time. But what is it like using them on the ground, outside the scope of presentation decks?
We surveyed 1,000+ employees who use AI at work about how they actually use agents day to day, how often those agents work as advertised, and who ends up responsible when they don't.
To see how that lines up with what's happening at the leadership level, we pulled in perspectives from CTO Studio, Howdy's own interview series with CTOs and engineering leaders. Taken together, we can see how agentic AI is faring from the bottom and the top.
54% of workers feel confident with AI, but nearly 1 in 10 still feel behind
Just over half of workers surveyed (54%) say they feel very confident in their use of AI, but it did take a decent amount of time to get there. Only about a quarter (26%) reached that confidence within their first three months of using AI tools. Nearly a third (31%) needed three to six months, 17% needed six months to a year, and another 17% took more than a year to feel comfortable. 8% say they still feel behind.
Once workers get there, AI settles into a fairly consistent set of jobs. Searching for internal information is the most common use (75%), followed closely by summarizing documents and brainstorming (70% each), writing (64%), and data analysis (53%). That pattern tracks with broader AI adoption trends in the workplace, which have leaned toward retrieval and synthesis over tasks that require independent judgment.
Managing the cost of AI hasn't kept pace with adoption. 19% of workers surveyed say they've struggled to budget AI tokens, and 14% have been hit with a surprise bill from an AI platform that far exceeded what they expected to spend. Of those who got the surprise bill, 76% say their company has since revised its AI spending policy.
CTOs are watching the same instinct play out at the leadership level, where token spend has become its own kind of vanity metric:
“Token maxing is the newest version of counting lines of code: it is easy to pull up a dashboard and read a number, but engineers are smart, and if you reward token spend they will manufacture it in ways that do nothing for your business,”
- Stephen Poletto, Field CTO at Span, on CTO Studio
71% say fewer than half their department's AI agents reach long-term production
Two-thirds of workers who use AI regularly (67%) have moved beyond prompting a chatbot and now use agents. Agents are AI systems that carry out multi-step tasks with less hands-on direction. 16% of those surveyed built their own agent from scratch, and 40% had to train the agent themselves.
“You have to treat the agent like it's your smartest dumb employee — the most brilliant junior-level person you have, faster and smarter than you, but with no experience and no wisdom whatsoever,”
- David Ting, Founder at a stealth AI startup and former CTO of Bespin Global, on CTO Studio
Once agents are running, the payoff is close to universal: 96% of workers say agents save them time, and 84% say the efficiency gains outweigh what the agents cost to run.
Workers also have plenty of firsthand evidence of agents falling short. Just under half (48%) say their agents rarely or never fail, 40% say failures happen sometimes, and 12% say they fail frequently or all the time.
Much of that failure traces back to something more foundational than the AI itself: 93% of workers surveyed have hit a problem caused by bad internal documentation or unclear instructions, and 1 in 5 say their agents simply aren't deployed efficiently at their workplace. The upkeep shows up on workers' calendars too, at an average of 3 hours a week spent troubleshooting agents, on top of the tool sprawl many AI-heavy workplaces already deal with.
Scaled up, that time cost may explain why so few agents make it to a stable home: 71% of workers say less than half of their department's AI agents ever reach long-term production, including 29% who say a quarter or fewer make it and another 42% who put the figure at 26–50%.
Inconsistency is the leading reason to ditch an agent
A quarter of workers surveyed (25%) have had to deactivate an agent entirely. Despite national headlines, the real reasons are more mundane than civilizational uprisings: 58% point to inconsistency, 22% say the agent never worked in the first place, 7% say it went rogue, and the remaining 13% cite some other reason. Workers reported spending just over a full work day trying to make an agent work– 8.4 hours – before deciding to scrap it.
One CTO points to a specific failure mode behind that inconsistency: agents that claim, without saying so outright, to have done something they haven't.
“The single quality that becomes the most critical is judgment. The AI is sometimes too eager to please, and it tells you it's complying with your request when there are parts it's actually not,”
- Dileepan Narayanan, Chief Product & Technology Officer at Afiniti, on CTO Studio
Only 28% of workers say their workplace is clear on who owns AI output
29% of workers use AI agents specifically for automation, letting the agent complete a task without a person handling each step by hand. That automation isn't always clean: more than half of workers surveyed (53%) say they sometimes or frequently run into redundant code or low-quality AI output, what's become known as "AI slop." The full breakdown is:
- 46% who see AI slop sometimes
- 39% who see it rarely
- 7% who see it frequently
- Just 9% who say they never encounter it
Handing tasks to an agent hasn't dented most workers' sense of ownership over their own work. 54% say using AI agents has had no effect on how much ownership they feel over their deliverables, and 28% say it's actually increased. Only 18% say it's decreased.
What's murkier is who's accountable when something goes wrong. Just 28% of workers say their workplace is extremely clear about who owns AI-generated outcomes on a team. That ambiguity pairs with a separate pressure: 34% say their employer values output quantity over quality, a combination that makes it easy to ship more AI-assisted work without anyone being clearly on the hook for its accuracy.
Ownership and quality are exactly the two things that get fuzzy first when a task moves from a person's to-do list to an agent's. One CTO argues the fix isn't a better tool; it's a policy that never lets that fuzziness happen in the first place:
“An AI agent is autonomous, but it still needs a scoped identity tied to an accountable human owner, because when something goes wrong an organization will not take the liability on its own,”
Bhavna Bhatnagar, CTO and Head of Engineering at VigourSoft Global Solutions, on CTO Studio
How to get more AI agents into long-term production
The data points to three fixable problems, not one unfixable technology. The first is documentation: with 93% of workers hitting issues traced back to unclear internal instructions, cleaning up the internal docs an agent references is one of the cheapest fixes available. The second is ownership: with only 28% of workplaces extremely clear about who's accountable for AI output, naming an owner for each agent-touched workflow closes that gap without touching the underlying technology. The third is staffing the build itself: 40% of workers who use agents had to train them personally, on top of their existing workload, which is a lot to ask of someone who wasn't hired to be an AI agent trainer. Fixing all three comes down to who's actually doing the work, which is why the CTOs we spoke with kept steering the conversation back to hiring:
“The people who make it through this AI shift will be the ones with curiosity and integrity, because knowledge I can teach but those two qualities I cannot,”
- Prakash Narayan, CTO at 3K Global AI, on CTO Studio
That's the piece Howdy is built to help with: hiring engineers who already know how to build, train, and troubleshoot AI agents means fewer agents stall out in the multi-hour troubleshooting cycle workers describe above. If agents are stalling out in your organization before reaching production, book a demo to see how the right hires close that gap.
Methodology & fair use
In August 2026, we surveyed 1,002 full-time US employees who use AI at work about their experience with AI agents. Ages ranged from 20-68 with an average age of 40; there were equal numbers of men and women.
For media inquiries, contact media@digitalthirdcoast.net.
When using this data, please attribute by linking to this study and citing Howdy.com.
