Automating customer service easily improves contact centre performance: it shortens calls, handles some enquiries and reduces the need for agents. The problem arises when a company measures the cost per contact but fails to check how much it actually cost to resolve the customer’s issue.
The cheapest contact can lead to more expensive service
Most CX automation projects start with a simple calculation. If a bot handles some of the calls, the cost per contact will fall. At scale, even a small saving on a single interaction can add up to a significant amount.
However, this calculation does not take into account what happens after the conversation ends. A customer who has not received a response may return via a different channel, ring the helpline, lodge a complaint or cancel the service. The contact disappears from the chatbot’s statistics, but the problem remains within the organisation.
Therefore, a high rate of conversations concluded without human intervention does not necessarily indicate that automation is working. It merely indicates that the customer was not transferred to an agent. It is not known whether they found a solution or simply gave up trying.
Qualtrics points out that the widely used ‘containment rate’ does not reflect service quality. The system may keep the conversation within the automated channel whilst failing to understand the customer’s issue. It is precisely the feeling that the problem has been properly identified that has the strongest impact on satisfaction and continued loyalty.
This implies a change in the fundamental unit of measurement. Instead of the cost per contact, what is needed is the cost of a lasting resolution to the issue, encompassing repeat enquiries, escalations, financial adjustments, complaints and customer churn.
AI delivers real productivity gains when it empowers staff
The risk of poorly designed automation does not negate its economic potential. A survey of 5,179 customer service staff showed that access to a generative AI assistant increased the number of cases resolved per hour by an average of 14 per cent. Among less experienced staff, the increase reached 34 per cent, whilst top-performing agents saw a significantly smaller improvement.
The key takeaway from this study is not about replacing people. Above all, AI accelerates the dissemination of knowledge previously held by the most experienced staff. It helps find answers more quickly, suggests ways to solve problems and reduces the time needed to document conversations.
It is in such applications that the return on investment is most predictable. The automation of summarising, classifying cases, searching for information and preparing responses reduces service costs without depriving the customer of access to the person responsible for the outcome.
Klarna reported that between the first quarter of 2023 and the first quarter of 2025, the cost of processing a single transaction fell by 40%, whilst maintaining the declared level of satisfaction. The company had previously also reported a reduction in the time taken to resolve a case from 11 minutes to less than two minutes, as well as a 25 per cent drop in the number of repeat enquiries. These are the company’s own figures, but they demonstrate that automation can simultaneously improve both cost efficiency and customer experience, provided it actually resolves the customer’s issue.
What matters, therefore, is not whether the conversation is conducted by a human or by AI. What counts is whether the customer achieves the expected result without any extra effort.
A bot ceases to deliver savings when it becomes a gateway
Customers do not reject automation simply because there is an algorithm on the other end. They accept it when they receive a response faster than via a traditional channel. Frustration arises when the bot fails to resolve the problem, yet simultaneously hinders access to a human.
Researchers describe this phenomenon as ‘gatekeeper aversion’. The customer avoids a channel where they first have to go through an imperfect automated stage, and may then be directed to an expert and have to explain the whole situation all over again. This aversion grows in line with the importance and risk of the matter. It is, however, reduced by clear information about the bot’s capabilities, the expected response time and quick access to a human agent following an unsuccessful attempt.
For the company, this means that the escalation mechanism is not merely a technical add-on to the chatbot. It forms part of the product and directly impacts the economics of the entire process.
A well-designed handover includes the contact history, customer details, steps already taken, and the reason why the bot was unable to resolve the matter. If an agent begins by asking “How can I help?”, the prior automation has not saved any time; it has simply added another step.
The greatest risk arises when teams are given the target of maximising the number of conversations retained by AI. The system then begins to optimise against escalation, even if handing the matter over to a human would be the best solution for both the customer and the company.
Fewer straightforward cases does not mean fewer skills are needed
Automation is also changing the way contact centres operate. With AI handling simple enquiries, agents are now primarily dealing with exceptions, complaints, disputes, financial issues and situations not covered by the standard workflow.
As a result, the average duration of a human-handled call may increase. This does not necessarily mean a decline in productivity; it simply means that the structure of enquiries has changed.
Gartner’s data shows that staff reductions are not the predominant outcome of AI implementations today. Only 20 per cent of customer service leaders surveyed reported reducing the number of agents due to AI. In April 2026, however, 85% of leaders indicated that they were expanding their employees’ scope of responsibility, as automation is shifting their work towards higher-value tasks.
For senior management, the more important question than the number of jobs lost is therefore whether the organisation can make better use of its staff’s skills. If agents are handling more complex cases, they need broader authorisation, better access to data and the ability to make decisions that have not been delegated to an automated system.
The savings resulting from reducing the first line of support can quickly disappear if the remaining team lacks the resources to handle exceptions. This leads to longer queues, an increase in the number of complaints, and longer resolution times for the most valuable or high-risk cases.
An autonomous agent changes the risk profile
A chatbot that answers questions may provide incorrect information. An AI agent with access to the CRM, payment system, orders and email may carry out an incorrect operation.
This is a qualitative change. The system not only affects customer perception but can also alter data, grant a discount, cancel a service, initiate a refund or disclose protected information.
The Air Canada case demonstrated that a company remains liable for the content conveyed by its chatbot. A Canadian court ruled that the company was liable for incorrect information regarding the possibility of obtaining a refund, even though the correct terms and conditions were set out elsewhere on the website.
For businesses, this means that a response generated by a model is not merely a casual technological suggestion. In the eyes of the customer, it constitutes a message from the company.
As AI is integrated into operational systems, the importance of cybersecurity also increases. OWASP points out that prompt injection can lead, amongst other things, to the disclosure of confidential information, unauthorised access to functions, and the execution of commands in connected systems.
The level of control therefore needs to increase in tandem with autonomy. For a simple information bot, an up-to-date knowledge base and monitoring of responses may suffice. An agent carrying out actions requires restrictions on permissions, limits on the scope of operations, a record of decisions, and approval for outcomes that are difficult to reverse.
From 2 August 2026, companies operating in the EU will also be subject to the obligation, arising from the AI Act, to inform users that they are interacting directly with an AI system. At the same time, the use of customer data to operate or develop models remains subject to the provisions of the GDPR.
Transparency is not merely a formal obligation here. Customers make different decisions when they know they are speaking to a bot, are aware of its limitations and can see the option to switch to a human.
There is no single ‘correct’ level of CX automation
Order status, invoice copies or simple data updates lend themselves well to full automation. They are repetitive, based on unambiguous information, and the consequences of an error can be quickly reversed.
A financial dispute, suspected account takeover, identity theft or a breach of contract present a completely different set of challenges. A single error can cost more than thousands of correctly handled enquiries.
Therefore, the decision to automate should not depend solely on the volume of enquiries. The consequences of a mistake, data quality, the number of exceptions, the sensitivity of the information and the ability to reverse an action are also important.
In this model,the CIO’s role extends beyond simply selecting a platform and language model. They are responsible for the entire chain: knowledge sources, integrations, permissions, activity logging, post-change testing, and the ability to quickly disable automation.
The board, meanwhile, requires metrics that link cost savings to customer outcomes. Alongside service costs and the proportion of automation, it is worth monitoring the solution’s effectiveness at first contact, customers returning with the same issue, failed escalations, complaints, financial adjustments and post-interaction churn.
CX automation delivers the greatest value not when it eliminates the most human interactions, but when it removes the most reasons why customers contact the company again.
