With tools like ClawBot, AI agents—including their tricks and risks—have become tangible for consumers. This has ushered in a new era for agentic AI. It's high time for a TechUpdate on agentic AI.

 

tech update is the Ziptone section on emerging technology relevant to customer engagement. We briefly explain what it is, how it works, why it is relevant, what the pitfalls are and we ask the reader's opinion: Hot or Not?

What is it

Until recently, the marketing noise dominated, but gradually, examples of applications are becoming available. Agentic AI is the next step in the development of generative AI. Agentic AI is the name for applications based on generative AI, in which an artificial agent can operate autonomously. An AI agent can make decisions within predefined frameworks and without human intervention, independently perform (complex) tasks, and adapt to environments, changing information, and user needs, escalating to a staff member when necessary.

In other words, agentic AI is AI that can act purposefully. An agentic AI application can independently execute a command formulated in natural language, such as "change my subscription and confirm the new price." The system then determines the necessary steps.

Agentic AI is therefore the counterpart to deterministic software that produces a standardized output according to fixed process steps.

The difference with an AI-based chatbot (which allows you to converse in natural language) is that the AI ​​agent also has access to external systems and an action-oriented layer. Agentic AI includes action models that can interact with systems like CRM, billing, order management, or identity management.

What can you do with it?

Agentic AI can fully execute processes. In customer contact, this offers applications in self-service and follow-up, such as checking the impact of a change, scheduling appointments, escalating a case, placing an order, or processing a return. It is a promising automation solution in customer contact.

For consumers, agentic AI also offers new possibilities – apart from their interactions with customer service. For example: with the OpenClaw's agentic solution Can you arrange for consumers to be asked via WhatsApp what they want to eat in the coming week? After which the bot will create a menu, fill the supermarket shopping basket, choose the delivery date and finalise the order?

Why is it relevant

For customer contact environments, agentic AI is the next step in automating customer contact processes. While chatbots focus on dialogue (questions, answers), AI agents can also perform tasks. This is a game-changer for customer contact, because a large part of the work of human customer contact professionals is focused on looking up data, navigating systems, checking statuses, scheduling appointments, arranging refunds, or adjusting tickets. On the consumer side, the availability of agents means that customers can also put "their" agents to work.

Customer service is the ideal testing ground for enterprise agents. It can reduce the handling time of customer processes that are currently handled by customer contact professionals or back-office employees.

Recent developments like OpenClaw and ClawBot demonstrate that agentic AI is more than just 'AI that talks'; both platforms demonstrate that 'AI does things'. OpenClaw is an agentic AI assistant that can run locally and perform real actions through various tools, such as clearing email, sending messages, and executing browser actions. This brings the concept of agentic AI to a tangible level for consumers; it also significantly increases the widespread interest in agentic AI. The derailments, in particular, are attracting attention. For example, WIRED published an article about an AI agent that ultimately started exhibiting phishing-like behavior after it gained broad access to systems.

How it works

An agentic AI solution roughly consists of five building blocks.

1. Goal, policy, and guardrails. You define what the system can and cannot do, with limits, escalation rules, and approvals. In customer contact, this often involves: human-in-the-loop for exceptions, large amounts, sensitive personal data, or policy deviations.
2. Reasoning and planning. The system breaks down a goal into steps, chooses the appropriate route, and monitors progress. This is a key difference from a flowchart in deterministic software: multiple paths to the same goal exist, and all kinds of context and information can be taken into account along the way.
3. The agent calls tools: CRM APIs, knowledge bases, order systems, identity checks, payment providers, ticketing.
4. Correct context and data. For example, if CRM, ERP, and billing don't have the most up-to-date information at the same time, an AI agent might produce incorrect results.
5. Logging, auditing, and monitoring. A mature agentic environment ensures the recording of steps, including the answer to the question: why, with what data, and with what result.

Systems in which multiple specialized AI agents collaborate are usually organized according to a hybrid design. A triage agent handles the intake process: think of intent recognition, summarization, prioritization, and follow-up. Various specialized agents, each with their own skills, handle individual tasks or processes, with limited access to data. An orchestrator agent determines which specialist is called in and when to escalate to a human.

AI agents only deliver value when they are connected to processes and systems. To function properly, they must rely on unambiguous information. For example, a "return" must be an unambiguous concept.

Furthermore, the scope of AI agents must be clearly defined for security reasons, which is why AI agents are often specialized in a single task. Specialization makes it easier to properly define guardrails, exceptions, and approvals. Multiple agents, each with their own distinct specializations, can collaborate with each other where necessary. Moreover, narrow tasks are easier to develop, test, and monitor, including performance.

A general categorization could be: agents that perform transactions (change, cancel, reimburse, identify) have narrow, limited authority. If an AI agent primarily supports tasks (search, explain, summarize), it can have broader operations. When dealing with financial transactions or advice, identity, or compliance, it's best to combine a specialized agent with a human-in-the-loop when exceptions arise.

Just pay attention

1. Several companies have invested heavily in developing and promoting solutions for agentic AI in recent years. There are significant vested interests involved. CCaaS and CRM providers know that automation puts pressure on their (FTE-based) licensing models. At Salesforce, agentic AI is clearly part of its reorganizations and product strategy.

Gartner states that more than 40% of agentic AI projects is expected to be discontinued before the end of 2027, partly due to costs, unclear business value and insufficient risk management.

2. Gartner also explicitly mentions "agent washing": products that are relabeled (assistants, RPA, chatbots) without true agentic capabilities. True agentic AI meets the following four characteristics:

• The AI ​​agent can demonstrably perform actions independently in systems (CRM, billing, logistics, IAM);
• The AI ​​agent works with explicit guardrails, permissions and audit logs (who did what, when, with what data, why);
• The AI ​​agent can handle exceptions: escalate, reclaim, or stop in case of uncertainty;
• The AI ​​agent runs on current, validated data.

Solutions that are scripted or that can only deal with the happy flow are not agentic AI.

3. While the AI-based chatbot relied on defined, well-organized knowledge and information environments, agentic AI must be able to work with data. An AI chatbot can quickly get started with a good knowledge base, but unleashing agentic AI on business is of a different order of magnitude.

In addition to data quality, latency can also be an issue for the safe operation of AI agents.

An AI agent must be able to consult the right source at the right time, regardless of the data's location, and must be able to rely on the information being the most up-to-date at that moment. Agentic AI becomes unreliable when working with outdated copies of data, or when systems are out of sync with the leading source (system of record). For example, the AI ​​agent must be able to work with (and rely on) all information from phone calls that occurred immediately before the incident.

4. Major risks

The risks are high precisely because of the access of agentic AI applications to systems, something for which the Dutch Data Protection Authority has already warnedFor example, agentic AI can also open the door to malware. Getting started with agentic AI places very high demands on the security of organizations, including within the development framework.

Hot or Not?

Featured, Technology, tech update