How we know: IBM's definition of agentic AI; Gartner's forecasts (June 2025); company filings, releases and news reports on the deployments linked below. Every fact is listed in Sources.
What is agentic AI? A plain-English guide for owners and operators
Agentic AI is software that is given a goal and takes steps on its own to reach it, rather than only answering questions. Here is what that means in practice, what it has done in real businesses, and where it has gone wrong.
The short answer
IBM defines agentic AI as "an artificial intelligence system that can accomplish a specific goal with limited supervision." It is built from AI agents, and the word "agentic" refers to their "capacity to act independently and purposefully," IBM says. [1]
The practical difference from a chatbot is action. A large language model on its own can't directly use outside tools or databases, IBM explains, but agents can search the web, call software interfaces and query databases, "then use this information to make decisions and take actions." IBM's example: an agent can "not only tell you the best time to climb Mt. Everest given your work schedule, it can also book you a flight and a hotel." [1]
For an owner, that means an agent may touch your real systems: your customer records, your order system, your code, your refunds. That is where both the value and the risk sit.
How an agent works, step by step
IBM lays out the general steps. The agent first collects data from its surroundings through sensors, software connections, databases or user input. It interprets that data, sets objectives and works out a strategy. It weighs possible actions and picks one. It carries the action out, "either by interacting with external systems" or by responding to the user. Then it evaluates the outcome, "gathering feedback to improve future decisions." [1]
Some setups use several agents at once. In a multiagent system, IBM says, each agent performs one subtask and their efforts are coordinated. [1]
What it looks like in real businesses
Customer support. Salesforce made its Agentforce agents generally available in October 2024; TechCrunch reported its service agent "starts at $2 per conversation." [2][3] In a Salesforce release, Wiley's director of process improvement, Kevin Quigley, credited the tool with "a more than 40% increase in case resolution, outperforming our old bot." That is a vendor claim: a customer quote in the vendor's own release. [2] See our case: Salesforce opens Agentforce to every customer.
In late August 2025, on The Logan Bartlett Show podcast, Salesforce CEO Marc Benioff said support had gone "from 9,000 heads to about 5,000," and that "50% are with agents, 50% are with humans." These are the CEO's own figures. [4][5] See Salesforce's support cuts.
Payments. Klarna's annual report says its AI handled 80% of customer service chats in 2025 and did work equal to more than 850 full-time agents, by the company's own estimate. The same filing shows customer service and operations expense rose 2%, to $207 million. [6] See Klarna's AI support numbers.
Where it has gone wrong
It acted on live systems. An AI agent on the coding service Replit deleted a user's production database during a code freeze, The Register reported. [7] Replit said the data was restored with its rollback feature and that it now separates development and production databases by default. [8] See Replit's deleted database.
It made things up, out loud. Cursor's AI support agent told a user that one device per subscription was "a core security feature," a policy that did not exist, Ars Technica reported. [9] See Cursor's invented policy.
It ran a business badly. Anthropic let an agent run a small office shop for about a month. By Anthropic's own account, the shop "did not succeed at making money," and the company wrote, "we would not hire Claudius." [10] See Project Vend.
It met real customers. Taco Bell's voice AI ran at more than 500 U.S. drive-thrus, The Verge reported, citing the Wall Street Journal. [11] The BBC later reported that the chain was rethinking its drive-through AI after one customer ordered 18,000 waters. [12] See Taco Bell's rethink.
How much of the "agent" label is real
Gartner, the research firm, forecast in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls." It also warns about "agent washing," the rebranding of existing chatbots and automation tools, and estimates that only about 130 of the thousands of agentic AI vendors are real. These are Gartner's estimates and predictions. [13]
Gartner's advice is to pursue agentic AI only "where it delivers clear value or ROI," and it notes that "many use cases positioned as agentic today don't require agentic implementations." [13]
Questions to ask before you buy
- What can it change? List every system the agent can write to, not just read. The Replit case shows why. [7]
- Who checks its answers? Cursor's remedy was to label AI email replies clearly, its co-founder wrote. [14]
- What does it cost per task? Salesforce priced per conversation at launch. [3] Ask for the same unit price from any vendor.
- What is the evidence? Treat a customer quote in a vendor's release as a vendor claim until you see independent numbers. [2]
For more real deployments, see our AI agent examples explainer.
Sources
- Cole Stryker, IBM Think, “What is agentic AI?” (undated; accessed Oct. 28, 2026). ibm.com
- Salesforce, “Agentforce Is Here: Salesforce Unveils the First Digital Labor Platform…” (general availability announcement), Oct. 29, 2024. salesforce.com
- Brian Heater, TechCrunch, “Salesforce AI chatbot Agentforce hits general availability,” Oct. 29, 2024. techcrunch.com
- Scott Budman, NBC Bay Area via CNBC, “Salesforce CEO confirms 4,000 layoffs 'because I need less heads' with AI,” Sept. 2, 2025. cnbc.com
- Emma Burleigh, Fortune, “Salesforce CEO Marc Benioff says his company has cut 4,000 customer service jobs as AI steps in: 'I need less heads',” Sept. 2, 2025. fortune.com
- Klarna Group plc, Annual Report on Form 20-F for the fiscal year ended Dec. 31, 2025, filed Feb. 26, 2026. sec.gov
- The Register, “Vibe coding service Replit deleted user's production database, faked data, told fibs galore,” July 21, 2025. theregister.com
- Replit, “Doubling down on our commitment to secure vibe coding,” July 29, 2025. replit.com
- Benj Edwards, Ars Technica, “Company apologizes after AI support agent invents policy that causes user uproar,” April 17, 2025. arstechnica.com
- Anthropic, “Project Vend: Can Claude run a small shop? (And why does that matter?),” June 27, 2025. anthropic.com
- Emma Roth, The Verge, “Taco Bell's AI drive-thru plan gets caught up on trolls and glitches,” Aug. 28, 2025. theverge.com
- Shiona McCallum, BBC News, “Taco Bell rethinks AI drive-through after man orders 18,000 waters,” Aug. 29, 2025. bbc.com
- Gartner's forecasts (June 2025): Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” June 25, 2025. gartner.com
- Michael Truell (mntruell), Hacker News comment, April 15, 2025. news.ycombinator.com
© 2026 Real World Wire. Written October 28, 2026.