AI in Customer Service: What Changed in 2026

At the start of 2026, most conversations about AI in support were still about capability: could the bot answer accurately, would customers accept it, how much of the queue it might take. By autumn, those questions had mostly settled into engineering problems. The arguments that replaced them are commercial and organizational. What exactly is a vendor paid for? Who owns the AI agent your team depends on, and what happens when that changes? What do you have to tell customers, and what happens to the agents whose work the AI takes over?

Five developments drove that shift. None of them is a new model or a new feature, and each one changes how a support leader should plan for 2027.

Vendors started billing for outcomes, and the definition became the contract

Per-resolution pricing was not new in 2026, but it became the market's reference point that year. In May, at its annual Relate conference, Zendesk moved its AI agents to billing for verified resolutions: a separate AI evaluation model checks each interaction, and only those judged resolved are charged, with low-value exchanges excluded. Fin, the company formerly known as Intercom, already billed per outcome. Contract-based vendors increasingly offer a resolution option alongside conversation-based pricing.

On paper, this moves risk from the buyer to the vendor, since an AI agent that fails to help is not paid. In practice, it moves the negotiation to a single sentence: the definition of a resolution. Vendors define it differently. Some count a conversation as resolved when the customer does not return within a set window. Some include certain handoffs in the billable unit. Zendesk now relies on its own model to make the call, so the check is still the vendor grading its own work.

The practical consequence is that pricing comparisons in 2026 stopped being about the rate. A lower rate under a looser definition can cost more than a higher rate under a stricter one. Before signing, teams need three things in writing: what counts as resolved, what happens to the charge when a ticket reopens or the customer writes again about the same issue, and whether the vendor's resolution reports can be checked against the help desk's own data.

The platforms consolidated

The second change is about who owns the tools. Three deals in six months redrew the market.

In March, Zendesk announced and then, on March 26, completed its acquisition of Forethought, an AI agent platform that had been sold to companies on many help desks. Zendesk now offers Forethought's agents as its own product and says they will keep working on other platforms.

In May, Intercom renamed itself Fin, after its AI agent, a sign of how much of the company's growth the agent had come to represent. Fin had built its position partly on running on top of other help desks, including Zendesk and Salesforce.

In June, Salesforce announced an agreement to acquire Fin for about $3.6 billion, with the deal expected to close in the final quarter of Salesforce's fiscal year 2027, which ends in early 2027. Salesforce has said Fin will complement Agentforce, its own AI agent product.

For buyers, consolidation raises a question that rarely came up in evaluations before: what happens to a tool you rely on when its owner changes, especially if it runs on a help desk now owned by a competitor? Nobody knows yet how these products will evolve, and the acquirers have made reassuring statements. But a support team that runs an AI agent from one company on top of a help desk from another now has a roadmap risk worth writing into the contract: notice periods for changes to integrations, a right to export configurations and conversation data, and termination terms if a supported integration is withdrawn.

The work moved from answering questions to completing requests

The third change is quieter, and it follows from the first two. Once vendors charge for resolutions, the most valuable tickets are the ones an AI agent can actually close. For informational questions, a good answer is the resolution. For refunds, cancellations, billing disputes, plan changes, and order problems, which make up a large part of the queue in subscription and ecommerce businesses, the resolution is an action in another system.

During 2026, that distinction moved to the center of how AI agents were sold and evaluated. Vendors that once demonstrated fluent answers from the help center now demonstrate write actions: a refund issued in the payment provider, a subscription changed in the billing system, a CRM record updated. The evaluation questions changed with them. Which systems can the agent write to? Who builds and maintains those integrations? Which actions can it take on its own, and which require a person to approve them? What does the audit log show when finance asks why a refund was issued?

This is where the real cost of AI in support now sits. The language model is the commodity part. The integrations, the written business rules, and the controls around actions that move money are the work, and someone has to do it: the support team, the vendor, or a mix of both.

Disclosing AI became a legal requirement in the EU

On 2 August 2026, the transparency obligations in Article 50 of the EU AI Act began to apply. For customer service, the relevant part is simple to state: people must be told when they are interacting with an AI system, unless it is obvious from the context. Formally, the duty falls on the provider of the AI system. In practice, the company that puts the agent in front of its customers controls the widget, the email templates, and the handoff messages, so the two have to work together.

The Digital Omnibus package agreed this year postponed the AI Act's obligations for high-risk systems to December 2027 for standalone systems, but it did not postpone the transparency rules. Fines for breaching them can reach EUR 15 million or 3 percent of worldwide annual turnover, whichever is higher.

For most support teams, compliance is a configuration task with a few details that are easy to miss. The disclosure should appear at the start of the interaction, on every channel where the AI replies, and in every language the AI uses. Handoffs should make clear when a person has taken over. Companies outside the EU are covered too if they serve customers there. The details of what counts as "obvious from the context" are worth confirming with legal counsel rather than with the vendor's marketing page.

The headcount story turned out differently than predicted

For two years, the dominant expectation was that AI would shrink support teams quickly. The data published in 2026 points to something slower and more interesting. In April, Gartner released a survey of 321 customer service and support leaders, conducted in late 2025. Eighty-five percent said they were adding new tasks and responsibilities to frontline roles as AI takes on more contacts, and 75 percent were moving agents into entirely new roles within service. Reductions are happening, but mostly gradually: 63 percent were shrinking frontline headcount through attrition, while 31 percent had made, or were planning, AI-related layoffs through early 2027.

The same research found that customers still trust human agents more than AI for recommendations, by 54 percent to 32 percent in a separate survey of US customers.

What this means for planning is that the agent role is being redesigned rather than removed. When the AI closes the repetitive requests, agents are left with disputes, complex cases, and upset customers, plus new work that did not exist before: reviewing AI decisions, approving actions the AI prepares, maintaining the rules and knowledge it runs on. Targets built for the old queue, especially handle time, stop making sense. Teams that redesigned roles and metrics in 2026 are in a better position than those that simply waited for attrition to do the work.

What did not change

Several things stayed exactly as they were, and they still decide whether AI works in a given support operation.

An AI agent is still only as good as what it reads. Outdated help-center articles still produce confident wrong answers, and policies that exist only in agents' heads still cannot be automated. AI still speeds up the process you already have: if refunds take ten days because finance releases them in batches, customers will hear that faster and be no happier. And the most honest measure of success is still the share of customers who do not have to contact you again about the same problem, which no vendor dashboard reports by default.

How Automated Support Platforms Fit Into Existing Workflows

Support automation platforms are increasingly designed to handle customer requests beyond generating responses. These systems can retrieve customer and payment information, apply predefined business rules, and execute actions through APIs. Common applications include refunds, cancellations, billing inquiries, subscription changes, order status updates, and account management. Depending on the complexity of a request, the system may resolve it automatically or escalate it to a human agent.

For example, platforms such as Soportio integrate with existing help desk software and connect to billing and customer relationship management (CRM) systems. This approach allows businesses to introduce automation without necessarily replacing their current support infrastructure. Organizations can establish rules that determine which requests are handled automatically, which financial transactions require approval, and which situations require human intervention.

Another consideration is how these platforms are implemented and maintained. Some solutions offer usage-based pricing tied to automated resolutions, while others use subscription or per-agent pricing models. Implementation responsibilities can also vary, with certain providers managing workflow configuration and ongoing maintenance while others allow internal teams to control automation rules directly. Evaluating these differences helps organizations determine how support automation fits their operational requirements, technical resources, and existing systems.

Five things to do before 2027 planning

• Get the billing definition in writing: For every AI tool you pay per outcome, write down what counts as resolved and how reopened tickets are treated, then check a month of the vendor's reported resolutions against repeat contacts in your help desk.
• Read your contracts for a change of owner: Look for terms on integration changes, data export and termination if a supported integration is dropped.
• Check your AI disclosure: If you serve EU customers, confirm the notice appears on every channel and in every language, and that handoffs to a person are clear.
• Tag your queue by request type: Separate the tickets that need an answer from the ones that need an action in another system. The second group is where the largest gains are, and where integrations and controls matter most.
• Redesign the agent role on purpose: Decide which new responsibilities agents take on, from approving AI actions to maintaining the rules, and change the metrics they are measured on before the queue changes under them.