AI agents for customer service: from the traditional chatbot to the intelligent agent

For a decade, *automating customer service* meant drawing a decision tree. If the customer types invoice, show these four options. If you do not recognize what they typed, repeat the menu. It worked in the demo and it collapsed on the first real day, when someone wanted to change the delivery address on an order that was also incomplete.

The first two describe the present: self-service resolves little, and human agents spend the working day not talking to customers. The last two describe the tension ahead: the market expects mass automation and customers demand a way out to a person. A serious deployment lives inside that tension.

The traditional chatbot failed for a structural reason: it tried to enumerate in advance everything a customer might want. Every exception became one more branch. A year in, the tree had hundreds of nodes nobody understood end to end, and the one person who knew where to touch it had already left.

The symptom shows up in the self-service numbers. Gartner found that only 14% of customer service and support issues are fully resolved without human help; even among the cases customers themselves rate as *very simple*, only 36% close in self-service ¹. It is worth reading that precisely: the survey measured self-service as a whole (portals, knowledge bases, IVR and chatbots), not the failure rate of chatbots. But it marks the ceiling of what the previous generation achieved.

Something similar happens on the other side of the counter. According to the sixth edition of Salesforce's State of Service report (5,550 service professionals across 30 countries, December 2023 to January 2024), contact center agents estimate they spend only 39% of their week working with customers; the remaining 61% goes to administrative tasks, internal meetings and training, manually logging case notes and other activities ¹³. It is self-reported perception (not an objective productivity measurement), the sample is global rather than Mexican, and Salesforce sells software in the category; but the pattern is recognizable.

An AI agent is not a chatbot with better prose. The technical difference is that it has no script: it has an objective, knowledge sources it can consult, and a set of actions it can execute against your systems. Faced with a query, it decides what it needs to find out, looks it up, and if something has to be done (check an order in the ERP, reschedule an appointment, issue a credit note within an authorized range), it does it and confirms the result.

That last row is the most underestimated. A chatbot that does not understand leaves the customer annoyed; a badly bounded agent confidently states a returns policy that no longer exists. The engineering work is not in making it speak well (that part is already solved), but in defining what it queries, what it executes and where it stops.

It helps to separate three things that usually get cited as if they were the same: controlled field studies, customer cases reported by vendors or consultancies, and analyst predictions. Only the first is evidence in the strict sense, and the strongest study on customer support is the one by Brynjolfsson, Li and Raymond, published as an NBER working paper. In a single Fortune 500 business process software company, with 5,179 chat support agents, staggered access to a generative AI conversational assistant increased productivity (issues resolved per hour) by 14% on average; 34% among novice or lower-skilled agents, with minimal effect on the most experienced ones. The paper itself reports signs that the tool may reduce the quality of the conversations of the most skilled agents, and documents improvements in customer sentiment and employee retention ³.

Reported cases promise more, with less rigor. McKinsey documents a fast-growing Asian bank that, after a twelve-month AI-enabled service transformation, doubled or tripled the use of its self-service channels, cut service interactions by 40% to 50% and lowered its cost-to-serve by more than 20% ¹²: a single anonymized case, in banking, reported by the consultancy itself, with no control group and no published methodology. Verizon, in turn, deployed an assistant with Google Cloud built on Vertex AI and Gemini models for 28,000 customer care representatives and store staff, with 95% answerability (the ability to answer queries comprehensively), according to the joint announcement ¹⁴. That is not a customer-facing agent but a copilot: the person still handles the conversation and the AI works behind them.

Most of the dashboards I have seen measure what is easy to count, not what matters. These are the metrics that can hold up a decision, each with the trap it carries:

And a warning about the point of comparison: if you measure the agent against the old chatbot, almost any result looks good. The honest comparison is against your current human operation, in the same categories and the same hours.

Escalation is not the fallback plan: it is part of the product, and it is where customer perception is decided. Gartner found that 87% of customers consider it essential that companies using generative AI offer the option to reach a human agent, in a global survey of 3,566 B2B and B2C customers fielded between February and March 2026 ⁶. This is not a soft preference: it is a condition of entry.

1. 1

   Define the triggers in writing, before you build

   Three families at minimum: explicit intent (the customer asks for a human), risk signal (formal complaint, legal mention, high amount) and technical limit (no support for the answer in the knowledge base). If it is not written down beforehand, it gets improvised afterwards.

2. 2

   Keep the exit visible at all times

   Hiding the *talk to a person* button to inflate deflection is the most expensive shortcut there is. With 87% of the customers surveyed rating that option as essential ⁶, hiding it trades a pretty metric for a real loss: 63% of the consumers surveyed by Zendesk say they are willing to switch to a competitor over a single bad experience ¹¹.

3. 3

   Hand over context, not the raw transcript

   The human should receive three lines: what the customer wants, what was attempted and what data was validated. Pasting the whole conversation forces them to read it while the customer waits.

4. 4

   Route to the right human, not the first one available

   The agent has already classified the case; use that classification in the routing. Escalating a billing issue to technical support adds one more handoff and erases the time advantage you gained.

5. 5

   Close the loop with whatever the human resolved

   Every escalation points to a gap in the knowledge base or in the catalog of actions. If nobody reviews them every week, the agent freezes at its day-one level.

Klarna is the most cited case in the world and it is almost always cited only halfway. In February 2024 the company announced that, in its first month of global operation (23 markets, more than 35 languages), its AI assistant held 2.3 million conversations, two thirds of its service chats, a volume the company itself calculated as equivalent to the work of 700 full-time agents. It also reported that repeat enquiries fell 25% and projected a $40 million impact on its 2024 profit ⁴. These are self-reported figures in a press release, with no methodology and no independent verification, issued while the company was preparing its IPO; the $40 million was a projection, not an audited result.

The second half arrived a little over a year later. In May 2025 Klarna started a pilot to hire remote human agents so that its customers can always talk to a person: the AI still handles the routine work and humans take the complex cases. It was not an abandonment of AI (anyone telling it that way is also telling half the story), but a design correction ⁵.

The aggregate data points the same way. Only 20% of service and support leaders report having reduced their agent headcount because of AI, while 55% keep headcount stable while handling higher volumes, according to a Gartner survey of 321 leaders from October 2025 ⁹. The same firm predicts that by 2027 half of the organizations that expected to significantly cut their service workforce will abandon those plans, given the difficulty of hitting staffing targets *without agents* ⁸.

The dominant move is not layoffs but a reshaping of the role, although both things coexist. In a global Gartner survey of 321 leaders (September to October 2025), 85% are expanding the responsibilities of their human agents as AI reduces contact volume and 75% are moving agents into entirely new roles; in the same survey, 63% are reducing front-line headcount gradually through natural attrition and 31% have already carried out or plan layoffs through the first quarter of 2027 ¹⁰. Anyone selling you an agent while promising only one of those two halves is not describing the data.

There are scenarios where I recommend not doing it, and saying so early saves months. It is not worth it if volume is low and every query is different: maintaining the knowledge costs more than the savings. It is not worth it if the knowledge base does not exist or is out of date: the agent does not invent correct policies, it amplifies the incorrect ones. And it is not worth it as the first piece if the underlying process is broken: if most of your contacts are about orders arriving late, the agent will elegantly explain why they arrived late instead of fixing the logistics.

One detail almost nobody designs is tone. In that same global Zendesk survey, 64% of the consumers surveyed say they are more likely to trust AI agents that show traits such as friendliness and empathy ¹¹. That is not solved with a generic instruction to *be friendly*: it is solved by deciding how your brand responds when the customer is already angry, and testing it with real conversations before you open the channel.

The deployments that do reach production look alike. They start with the three or four categories that concentrate most of the volume in their own operation (typically order status, billing, appointment changes, access resets), not with the long tail. They build the knowledge base and the read integrations first, and only then enable write actions, one at a time and with limits. They measure against the previous human operation from week one, and they have escalation designed before the agent is ever exposed to a customer.

Gartner predicts that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention, with a 30% reduction in operational costs ⁷. It is an analyst prediction, not a measurement, and the adjective *common* is doing all the work. The useful reading is not that the human team disappears: it is that the repetitive part is going to move, and whoever builds the knowledge and the escalation well today will get there with an advantage.