An AI agent is software that takes a goal, decides its own steps, and uses tools like a browser, a code editor or a company database to get there. Most lists of AI agents use cases read like a brochure. The evidence is messier. Agents are doing measurable work in a few places, mainly writing code and handling routine support and IT requests. They still fail often enough elsewhere that at least one well-known company has walked its plans back in public.
This piece sorts what’s holding up from what isn’t, using surveys, benchmarks and company statements from the last two years.
What counts as an agent
A chatbot answers a question. An agent carries out a task: it breaks the job into steps, calls tools, checks what happened and keeps going until it thinks it’s done.
The label gets stretched. Gartner has warned about “agent washing,” where vendors rebrand chatbots and old automation as agentic AI. Its June 2025 release estimated that only about 130 of the thousands of vendors using the term have real agentic capability. If a product page says “agentic,” ask what it actually does without a person clicking through each step.
Where agents are doing real work
Writing and changing code
This is the strongest case. Google CEO Sundar Pichai says about 75 percent of the company’s new code is now generated by AI and then reviewed by engineers, up from 25 percent in October 2024 and 50 percent by fall 2025, according to The Decoder. Two limits on that number: it’s Google’s own, and the report doesn’t say how the share is measured.
A controlled trial points the other way. METR, a research nonprofit, had 16 experienced developers work on 246 real issues in large codebases they knew well, randomly allowing AI on some tasks and not others. With AI allowed, tasks took 19 percent longer. The developers had predicted a 24 percent speedup, and afterward they still believed they’d been about 20 percent faster. DX’s write-up of the study notes the developers accepted fewer than 44 percent of what the AI generated, and the authors say the result is specific to that setting and doesn’t mean the tools are useless elsewhere. The tools in the trial were also from 2025.
Put together, agents help most on well-scoped work and new code. The gain shrinks when you’re inside a big, old codebase that you understand better than the tool does.
Routine customer support
In February 2024, Klarna said its AI assistant was doing the work of 700 agents and had handled 75 percent of customer chats, about 2.3 million conversations, in its first month. By May 2025 the CEO was telling Bloomberg that leaning on cost had produced lower quality, and the company started hiring human support staff again, as Entrepreneur reported.
Gartner’s surveys show the same pattern across the industry. In a survey of 321 customer service leaders in October 2025, only 20 percent had reduced staffing because of AI, and 55 percent were holding staffing steady while handling more volume. An earlier Gartner survey of 163 leaders found 95 percent planned to keep human agents, and that 51 percent of customers trust a human most to resolve a problem, against 7 percent who trust AI most, per The Register’s summary. Gartner’s own view in that report was that a contact center with no human agents is “not yet technically feasible, nor is it operationally desirable.”
So agents take the repetitive layer of support, and people keep the rest.
Internal IT, lookups and engineering support
McKinsey’s 2026 survey of 1,719 people in 97 countries, fielded between May 4 and June 8, found that agents are most often scaled in IT, knowledge management and software engineering. Among large companies (over $1 billion in revenue), 40 percent are scaling agents, up from 27 percent a year earlier. Smaller companies stayed flat at 22 percent.
The money is lagging the adoption. In the same survey, 37 percent of respondents attributed at least some profit impact to AI, unchanged from last year, and only about 6 percent qualified as high performers. That figure covers AI in general, not agents alone, but it suggests the spending is running ahead of the payoff.
Shopping and payments
Agents that buy things for you are the use case we’ve covered most closely. We’ve gone through who pays when an agent orders the wrong product, whether an agent ever sees your real card number, and how well agents handle coupon codes.
Driving
The physical-world version is still very small. Driverless trucks haul paid freight in Texas, but with a few dozen vehicles per company against millions of drivers, as we laid out in our piece on whether AI will replace truck drivers.
Where agents fall down
Multi-step office work
Carnegie Mellon researchers built a simulated software company and gave agents everyday jobs: browsing, writing code, messaging coworkers. The best model they tested, Gemini 2.5 Pro, completed 30.3 percent of tasks. Claude 3.7 Sonnet completed 26.3 percent. The failures are the interesting part. Agents skipped an instruction to message a colleague, got stuck on browser pop-ups, and in one case, an agent that couldn’t find the right person in the chat tool renamed a different user to match the name it wanted.
Salesforce ran a similar test on sales and service tasks. Leading agents succeeded about 58 percent of the time on single-turn tasks and about 35 percent when the work took several back-and-forth turns, and the models showed near-zero awareness of confidential information. Both studies are covered in The Register’s report.
Those were mid-2025 models, and newer ones score higher, so read the scores as a snapshot. The failure types are what’s worth keeping: missed instructions, brittle handling of messy screens, shortcuts that look like success, and worse reliability as a task gets longer.
Cost and unclear payoff
Gartner predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value or weak risk controls, in the same June 2025 release. That’s a forecast, not a measurement, and the survey behind its adoption figures drew on 3,412 webinar attendees, which isn’t a representative sample of businesses. Still, it matches what the customer-support numbers show: the savings arrive more slowly than the pitch.
How to tell whether a use case will hold up
The examples above suggest a few checks, and they apply whether you’re evaluating a vendor or trying an agent yourself:
- You can tell quickly whether it worked. Code either passes its tests or it doesn’t. A support ticket is either resolved or reopened. Agents do best where success is visible.
- A mistake is cheap to undo. An agent that drafts is safer than one that sends, pays or deletes.
- The task is narrow and repeated. The wins in IT and support come from the same few request types arriving thousands of times.
- A person sees the output before it matters. Google’s code is reviewed by engineers, and that review step is part of why the 75 percent figure is believable.
- There’s an easy route to a human. Klarna’s reversal came down to customers who couldn’t get one.
If a proposed use case fails two or three of those, expect the cancellation numbers to apply to you.
What’s coming
Gartner expects at least 15 percent of day-to-day work decisions to be made autonomously by agents by 2028, up from zero in 2024, and a third of enterprise software to include agentic features. Treat both as predictions. The measured picture today is narrower: agents are useful on bounded, checkable work, and the companies that tried to replace people wholesale are the ones now rehiring.
Frequently Asked Questions
What are AI agents used for?
The best-supported uses are writing and reviewing code, answering routine customer support requests, and handling IT and internal knowledge lookups. Shopping and payments agents exist but are newer, and physical-world agents like driverless trucks are still a few dozen vehicles per company.
What’s the difference between an AI agent and a chatbot?
A chatbot responds to what you type. An agent takes a goal, plans steps, uses tools such as a browser or database, and acts on the result, often without asking you at each step.
Are AI agents reliable enough for business use?
For narrow, checkable tasks, often yes. On longer, multi-step work the evidence is weaker: Carnegie Mellon’s simulated office tests had the best model finishing about 30 percent of tasks, and Salesforce found success fell from about 58 percent on single-turn tasks to about 35 percent on multi-turn ones. Those were 2025 models, so current numbers are likely higher.
Will AI agents replace customer service teams?
Not so far. In Gartner’s October 2025 survey, only 20 percent of customer service leaders had cut staff because of AI, and Klarna started rehiring human agents after saying its AI assistant had done the work of 700.
How much code do AI agents write?
Google says about 75 percent of its new code is AI-generated and then reviewed by engineers. That’s a company-reported figure with no published method. A separate 2025 trial found experienced developers working in large codebases took 19 percent longer with AI than without it.
Why do so many AI agent projects fail?
Gartner points to rising costs, unclear business value and inadequate risk controls, and predicts over 40 percent of agentic projects will be canceled by the end of 2027.
