Real-world use cases for agentic AI

Remember those simple days of yore, when generative AI meant sending a question to an AI model and getting an answer in return? You might add in a vector database to provide some context for the question and some guardrails for safety and security. That sounded hard at the time, but in retrospect it was a walk in the park.

Today, the trending technology is agentic AI systems. Instead of a chatbot, a vector database, and a guardrail, you now have an endless selection of datasets, large and small models of all kinds running in all possible locations, and instead of a simple prompt-response interaction with a human on one end and an LLM on the other, there’s an army of agents connected by a complex — and dynamically evolving — logical workflow. Or probabilistic workflow, as the case may be.

There are new protocols connecting data and agents, new protocols connecting agents to other agents, and orchestration frameworks to chain it all together.

With all this complexity, you might think that companies would be slow to adopt agentic AI. You’d be very wrong.

In a Cloudera survey of 1,500 enterprise IT leaders in 14 countries released in mid-April, 57% of respondents say they’ve already implemented AI agents, and 96% say that they plan to expand their use of AI agents in the next 12 months.

[ Agentic AI’s impact on the enterprise: ongoing coverage ]

Other surveys show similar results.

According to a SnapLogic survey of over 1,000 IT decision-makers in the US, UK, Germany, and Australia released in February, 50% are using AI agents. In addition, 92% of respondents are confident that AI agents will deliver meaningful business outcomes in the next 12 to 18 months, and 79% are planning to invest over $1 million in AI agents over the next year.

According to Gartner, agentic AI is the top strategic trend of 2025. By 2029, 80% of common customer services issues will be resolved autonomously, without human intervention. The firm also predicts that 33% of enterprise software applications will include agentic AI by 2028, and 15% of all day-to-day work decisions will be made autonomously.

“It’s certainly not just marketing hype,” says Gartner analyst Sid Nag. “It is something that’s going to be of very high importance for automating many tasks in many environments.”

What is an AI agent, really?

There is a bit of “agent washing” happening in marketing departments right now. Just as, over the past three years, companies have added the “AI” label to every application, so now everything with a chatbot anywhere near it is being labeled an agent.

But in general, the way that technology leaders differentiate an AI agent from a chatbot is that the agent can take autonomous action.

No longer limited to answering questions, AI agents can carry out tasks on our behalf — sometimes extremely complicated tasks that require extensive interactions with other agents and systems.

Here’s how enterprises are putting AI agents to use today.

Software engineering with agentic AI

Software development was one of the breakout use cases for generative AI — and is also a top use case for agentic systems.

A GitHub survey of 2,000 developers in the Brazil, Germany, India, and the US found that 97% were using AI coding tools by mid-2024. And according to a HackerRank survey of more than 13,000 developers across 102 countries released in March, AI now generates, on average, 29% of all code.

There’s a wealth of public code bases on which models can be trained. And larger companies typically have their own code repositories, with detailed change logs, bug fixes, and other information that can be used to train or fine-tune an AI system on a company’s internal coding methods.

As AI model context windows get larger, these tools can look through more and more code at once to identify problems or suggest fixes. And the usefulness of AI coding tools is only increasing as developers adopt agentic AI. According to Gartner, AI agents enable developers to fully automate and offload more tasks, transforming how software development is done — a change that will force 80% of the engineering workforce to upskill by 2027.

Today, there are several very popular agentic AI systems and coding assistants built right into integrated development environments, as well as several startups trying to break into the market with an AI focus out of the gate.

The most popular agentic coding platforms today include Devin from Cognition Labs, Cursor, and Windsurf. There’s also a free, open-source option, Cline.

OpenAI is expected to release its own agentic software engineer platform soon, A-SWE, which stands for agentic software engineer.

Established players are getting into the game as well. GitHub Copilot announced an agentic mode in February. Amazon announced an enhanced CLI agent for its Q Developer platform in March. VS Code rolled out an agentic mode in April. Google also has an agentic AI development platform, Firebase Studio, that the company announced in April.

Agentic AI code development platforms are a significant advance over chatbot-based code assistants. With a chatbot, a developer asks a question and gets a code snippet. But an agentic AI platform can plan an entire project, write the components, create tests and check that the code works, and iterate until it meets all the project objectives.

At cybersecurity firm Abnormal AI, between half and three-quarters of the company’s 350 engineers are currently using these tools, says Dan Shiebler, the company’s head of machine learning.

“We’re making very substantial investments in making our engineers more effective,” he says. The company is currently using Cursor and is experimenting with other platforms. “And there are a number of things built internally.”

Not every use case requires a full agentic system, he notes. For example, the company uses ChatGPT and reasoning models for architecture and design. “I’m consistently impressed by these models,” Shiebler says.

For software development, however, using ChatGPT or Claude and cutting-and-pasting the code is an inefficient option, he says.

“The next step up is the Cursor type of interface, where you have a box where you tell it what to do, and the agent responding to you has context of the code and can make changes based on the instructions you give it, and you can review it.”

But the latest evolution is where the coding system can generate an entire application without a human touching the code at all. It can use APIs and provision infrastructure — and there are several areas where Abnormal is already using such tools.

“Bolt, v0, and Lovable are three tools in this category,” Shiebler says. “I personally like Lovable, but we’ve seen a lot of success with v0 for interface design, where it’s taken the place of Figma in a lot of user workflows.”

Any company that’s serious about developing technology needs to be using agentic AI software development tools, says Kevin Merlini, VP of product and CoCounsel for tax, accounting, and audit at Thomson Reuters. “If they’re not, I don’t know why they’re not doing that,” he says.

Thomson Reuters’ software engineers use various AI-powered coding tools. “We have a multi-model approach so we’re not locked in,” he says. “And, broadly, we have a multi-vendor approach.”

Being flexible allows companies to be able to ride the wave of innovations that’s happening now, he says. “Everyone should be employing multi-prong strategies, exploring products, and trying to understand it themselves.”

AI agents for research and document analysis

Thomson Reuters isn’t just using agentic AI internally for things like software development and research. It’s also building agents into its customer-facing offerings.

Specifically, the company has created the CoCounsel genAI assistant for legal, tax, audit, and accounting professionals. More than 240,000 customers now use CoCounsel, with the greatest usage related to legal research and document analysis skills.

“Agentic technology is supercharging the way we can deliver value for customers,” says Merlini. “I look at it as a new category of software.” It goes far beyond what can be accomplished with a simple chatbot interface, he says.

“With a basic chatbot using RAG and one folder of files, you’re getting a prompt and giving an answer,” he says. “There’s not too much autonomy. But what if you have dozens of different repositories? How does it know which repositories to access? What if you have multiple tools and capabilities, taking actions in some systems, pulling data from an API?”

Even a straightforward task like research can benefit from an agent approach, he says. “It seems simple on the surface,” he says. “But what if someone has a question that requires multiple steps, and the answer isn’t just in one source?”

AI is in a feedback loop right now, he says. “All these building blocks are coming together, giving the system more capabilities and more tools that it can use,” he says. “It’s opening up more use cases. And it’s definitely the direction we’re going.”

Agentic AI for customer service

Customer support chatbots can answer simple questions. AI agents can tackle more complex challenges — and can even act to solve problems.

There’s a lot of risk here. It’s bad enough if a chatbot gives a customer incorrect information or promises a discount that the company can’t deliver. But what if the AI can act autonomously, can place or cancel orders or can give discounts and refunds?

That’s why, for its initial deployment, Bosch Power Tools is using agentic AI to assist human agents, not replace them — and is keeping humans in the loop as a safety precaution.

“The users will be our agents,” says Victor Nguyen, the company’s project lead for genAI in business operations. End customers won’t be exposed to the new agentic AI systems directly. “We’ll have autonomous AI agents supporting our human agents.”

Bosch is using Cognigy.AI as its AI platform, which supports integration with multiple back-end AI models. “At the moment we’re using [OpenAI’s] GPT 4.0 and [Google’s] Gemini,” says Nguyen. “We’re quite flexible.”

 It’s also integrated with the company’s CRM system and ticketing system. “We have also integrated it with a translation service, so we can translate email text or document attachments,” Nguyen says.

The system is currently in the second pilot phase, he says, and will be used by live human agents for real cases starting in May. In June, it will be deployed to the first customer service center, out of 23 at the company.

The eventual goal is to have the platform be widely used across the company, he says. “Bosch is such a huge company; Power Tools is just one division,” he says. “We will join forces with other Bosch groups to make it a scalable solution. We’re closely collaborating with our central IT to make sure this is globally scalable.”

The biggest challenge, he says, isn’t the agentic technology but the lack of company-wide standardized processes.

“In Germany, say, there might be a different process for changing an order than if someone in Latin America was doing it,” he says. “And there are different systems being used. We reviewed screens and made sure we standardized them as much as possible, though there will always be some country-specific stuff.”

Nguyen recommends that companies looking to roll out agentic AI for customer service start standardizing data and systems as soon as possible.

“Most people think that AI is the solution, that AI will fix everything,” he says. “That’s not the case. The first homework to do is to get the good data, good quality data, and make sure it’s maintained. It’s not just a one-time task to upload the data somewhere.”

AI agents for document processing

Enterprises have been using chatbots to process documents for years. Generative AI is good at, say, summarizing, or pulling out specific information.

But with agentic AI, an entire document-focused workflow can be automated.

Marketing firm Route Three Digital recently built an AI agent for a customer using Google’s Vertex platform and Gemini genAI models to automate a process that used to take seven days as the client’s users collected documents and information to create a proposal.

“We wrote the code and scripted it to capture all the pivotal information into one master document, then use Gemini to clean up the text and make it more readable,” says Sharmilla Singh, the company’s chief marketing and operations officer.

It’s still not completely foolproof, she says, and there is still a human involved to review the final document and tailor it as needed. But when the tool launched last year, the client saw a multi-day workflow reduced to a few hours.

The next step, she says, is to have an AI agent that does everything. “The goal is to remove the human,” she says.

Marketing is a relatively low-risk use case for agentic systems, Singh says. “It’s not going to take down your company.”

Other use cases for AI in marketing include search engine marketing and online advertising. “If you don’t stay on top of new methodologies, you could lose market share,” she says.

Agent democratization

Google’s Vertex AI is just one of many AI agent building platforms that’s trying to make it easier to build and deploy AI agents. In April, Google also announced that its Agentspace platform, first launched in December, now has a no-code agent designer and pre-built agents for tasks like deep research and idea generation.

Google has also launched an agent marketplace and opened it up to partners. As of this writing, there are 138 agents offered on the platform, from companies like Deloitte, VMware, Amdocs, Palo Alto, Wipro, and Dun & Bradstreet.

But Google is just starting to catch up to the 800-pound gorilla that is Microsoft’s Copilot Studio. It has already been used by more than 160,000 organizations to build agents, said Charles Lamanna, Microsoft’s Corporate VP of Business and Industry Copilot, in a March update. More than 400,000 custom AI agents have been created in the previous quarter alone, he added.

Other companies offering AI agents include AWS, with its Bedrock Agents, as well as Salesforce, ServiceNow, Workday, and SAP.

What’s more, AI model makers are beginning to bake agentic capabilities into their core products. OpenAI, for example, just announced two new reasoning models with agentic AI functionality and tool use built right in. In the future, businesses may not even need third-party agents or agentic platforms.

But agentic AI technology is still so new that “it’s a little too early to get any real feedback from enterprises” about their experiences with it, says Gartner’s Nag. “I don’t think it’s ready for prime time yet, or even if it’s ready for prime time, it’s not something that people are adopting wholesale.”

And there’s still a lot of healthy skepticism about the technology, he says. “I think that will be mitigated over time and you’ll see it become more pervasive in various functions — IT operations, sourcing, procurement, finance, and a whole bunch of other things.”

“It’s not there yet,” he adds. “But it’s becoming a little bit more real.”