Agentic AI: The autonomous revolution that is reshaping enterprises in 2025

Agentic artificial intelligence (AI) is transforming the enterprise landscape by introducing truly autonomous decision-making capabilities that transcend current AI applications. This next evolution of AI moves beyond simple query-and-response systems to AI agents that can independently execute complex tasks, make nuanced decisions and adapt to changing business environments with minimal human oversight.

Deloitte predicts that 25% of companies currently using generative AI (GenAI) in their operations will launch agentic AI pilots in 2025. That number is predicted to grow to 50% in 2027. Read on to learn why leading research firms like McKinsey & Company, Gartner and Forrester have named agentic AI one of the top technology trends for 2025, and what organizations that haven’t already implemented AI agents can do to prepare for this transformative technology.

How is agentic AI different from generative AI?

Generative AI is sophisticated technology capable of producing outputs in response to prompts. The way in which GenAI reacts to a particular prompt is by mimicking patterns learned from its training data. While powerful, GenAI can’t autonomously complete complex tasks, which is exactly what agentic AI can do.

Based on outcomes, AI agents can plan, strategize and adjust their approach to achieve a goal. For example, a human engineer can input a prompt that details a new software feature. Agentic AI can output the code, test it, orchestrate multiple iterations of the code in response to the tests and deploy the tool it helped create.

Steve Nemzer, director of AI growth and innovation for TELUS Digital, defines agentic AI as “systems that can do more than converse and solve abstract problems. They can make decisions and take action.” He outlines its main characteristics as follows:

1. The technology can action a high-level goal by breaking it down into smaller sequences or a series of logical steps.
2. It can use tools like optical character recognition (the conversion of images of text into machine-encoded text) or operate complex frameworks like enterprise resource planning software applications to analyze data and make decisions.
3. Agentic AI can gather new information from the environment and analyze that data to make decisions.
4. By cooperating with other artificial intelligence and using chain-of-thought reasoning, it can improve on its own over time. It can also explain its actions with conversational capabilities.

While generative AI’s main purpose is content creation and data synthesis, agentic AI’s main purpose is autonomous problem solving and task execution. Agentic AI builds on GenAI’s capabilities, enabling AI agents to run autonomous, multi-step workflows.

How does agentic AI work?

Agentic AI’s additional technologies and capabilities enable it to act independently. Here’s a general overview of how it works in practice:

• The user provides instruction by entering a prompt. The AI agent interprets the intent and asks for further clarification, if needed.
• The AI agent transforms the prompt into a structured workflow by breaking it down into smaller tasks.
• To identify the features, objects and entities relevant to the task at hand, the AI agent collects and processes data from its environment. Sources include databases, sensors and digital interfaces.
• Some AI agents are built on large language models (LLMs), in which case the underlying LLM acts as the central “reasoning” engine that understands the specific task. The AI agent coordinates with various specialized models that carry out functions like content creation, vision processing or recommendation systems. To do so, these models may use multiple techniques such as retrieval-augmented generation (RAG) to access proprietary data sources so they can deliver accurate outputs. The AI agent is able to rapidly execute tasks based on its formulated plans by integrating with these external tools and software via application programming interfaces (APIs).
• Some agentic AI applications can continuously improve through a feedback loop, which inputs the data generated from the agent’s interactions back into the system to enhance it.

Agentic AI use cases

The implementation of agentic AI is reshaping enterprise workflows, making them more efficient, productive and faster. The following are some examples:

• Customer service: AI agents can handle more complex customer inquiries and act autonomously to resolve those issues. For example, a company that sells wireless audio products could leverage an AI agent to help with a variety of initial tasks customers may encounter post-purchase. To assist a customer with setting up a new product, the AI agent would begin by asking them about the product, the environment in which it’s located and basic diagnostic information. The AI agent would then instantly access all of the relevant data (a process that would require a human customer service representative to access various systems, manuals and programs) and guide the product set-up until it can confirm the customer is satisfied.
• Financial services: In terms of productivity, AI agents can be used to automate repetitive tasks such as compliance checks and transaction processing to reduce human error and leave employees free to do more strategic work. For example, consider a customer who wants to change the payment due date on their credit card. A GenAI bot will respond with instructions on how to do this and likely link the client to the page where they can do so. However, an AI agent will explain how to change the payment due date, link the client to the page where they can do so, it might wait for the customer to make the change and then put the client on an outreach list for a potential cross-selling opportunity — all with one prompt.
• Cybersecurity: According to a 2023 report by the World Economic Forum, the global shortage of cybersecurity professionals is close to four million. Further, 52% of organizations cite this resource shortage as the biggest challenge in safeguarding their digital infrastructure and computer systems. AI agents can assist cybersecurity work done by humans by, for example, autonomously detecting cyber attacks, helping software engineers pinpoint vulnerabilities in new code, detailing how to solve a cybersecurity problem and more.
• Coding: By 2028, 90% of enterprise software engineers will use AI code assistants, according to Gartner. By solving complex problems like making unfamiliar code more easily understood, generating test cases and scaling deployment, AI agents can free up developers’ time to focus on more strategic work.

AI agents have the potential to automate multi-step processes across business functions, making operations more efficient and employees more productive. In order to do so, businesses must first undergo the implementation process.

Best practices for implementing agentic AI in your organization

The agentic AI implementation journey is not dissimilar to the GenAI one. However, unlike implementing a responsive generative AI tool, you are developing an autonomous system capable of independently making decisions, taking action and, in many cases, continuously learning. Compared to GenAI, agentic AI involves more complex architectures and the interaction of advanced models.

The first step in your agentic AI implementation process is to determine your goals, says Nemzer. An AI agent requires a clearly defined business scenario and defined sequence of actions to guide its decisions and determine how it should prioritize tasks. To set your goals, Nemzer recommends assessing your current workflows for automation potential. “Look for task sequences that generally have a limited range of outcomes but may require sophisticated decision rules that were too complex for previous prescriptive automation attempts,” he says.

Once your use cases are established, determine which agent framework approach will best suit your needs. Some custom use cases will require AI agents to be built from the ground up, while for other more common use cases, AI agent templates and tools from companies like Salesforce (Agentforce) and Microsoft (Copilot) can be used as starting points.

Evaluating the technical stack components is the next step. Your AI agent needs access to a carefully curated knowledge base (a collection of information sources and databases that will inform its decision making and task execution). This includes data specific to your industry, regulatory requirements, informational feeds via RAG, internal policies and more.

The next phase is testing your model outputs for accuracy. Inevitably, your model will require adjusting through further supervised fine-tuning and reinforcement learning from human feedback (RLHF) to arrive at the optimal output. During the RLHF stage, “your group of testers should be diverse and include skilled exploratory software test engineers trained in finding edge cases, security experts, testers with industry-specific domain expertise and laypeople who represent your target users” according to Nemzer.

A critical step is determining which model decisions will require human-in-the-loop oversight, which Nemzer suggests for all model decision-making and autonomous actions during the initial implementation phase. “For lower-risk actions, where consequences of flawed outcomes are minimal, you’ll eventually dial back the human oversight,” he says. “However, for automations that have significant consequences, human oversight will always be required.”

Prior to agentic AI implementation in your organization, a combination of automated testing and a second stage of human-in-the-loop testing, such as RLHF, is necessary to ensure your application produces safe and consistent results. Only then should you integrate AI agents and tools into your workflows. An ongoing requirement will be to continuously monitor and evaluate your model, adjusting as needed.

Nemzer offers two key pieces of advice when implementing agentic AI. First, set modest initial goals. “It’s tempting to make big leaps, but in this fast-moving environment, smaller, more agile steps are likely to be more effective.” Second, he recommends overcommunicating — with partners, employees and especially customers. Being transparent about how and when agentic AI (or any AI) is being used is key to not only building trust, but also to practising responsible AI. Be open about what the AI agent is good at, what its limitations are, what data it was trained on and the fact that there is human oversight of model output.

Risks and challenges in agentic AI

Agentic AI offers significantly advanced capabilities that could vindicate business investments in the technology; however, implementing AI agents into your workflow does come with some considerations.

The potential risks associated with agentic AI implementation are, again, not dissimilar to those encountered when implementing generative AI. For example, security and privacy concerns, bias and fairness considerations, hallucinations and more. However, agentic AI requires further considerations since you’re incorporating AI agents that can reason and act autonomously into your workflows. While all of these risks can be enough to give you pause, they can be mitigated by implementing the following four criteria prior to any AI implementation:

1. First and foremost, focus on putting strong data governance and cybersecurity policies in place.
2. Balance risk with reward by determining the level of autonomy and data access your business is comfortable with allowing AI agents to have. As Nemzer mentioned above — start small. Consider implementing low-risk use cases that require access to non-critical data first and build from there.
3. Because of the autonomous nature of agentic AI, it is essential to ensure that human oversight and guardrails that define and limit an AI agent’s scope of action are built into the workflow. Guardrails are frameworks that ensure AI systems operate within set boundaries. They are implemented to prevent model output that is inappropriate, factually incorrect (a hallucination), that goes against regulations specific to an industry and more.
4. Continue to carefully evaluate and question agentic AI’s capabilities as it continues to evolve. Each evolution will come with new challenges, and these could take some time to resolve. Maintaining a healthy level of skepticism can serve organizations well.

Through autonomously making decisions and taking real actions, agentic AI is poised to revolutionize enterprise operations. The time to seize the potential of this transformative technology is now.

Turn your agentic AI ideas into reality

Regardless of where you are on your path to AI implementation — from just starting out to making the leap to agentic AI — TELUS Digital is here to help. Our end-to-end solutions empower business transformation every step of the way. Reach out to learn more.