The AI agent answered the overnight queue on its own. The refunds it granted, the prices it quoted and the customers it handed back are still for the support team to stand behind.
AI Skills for Customer Service Teams: What a Support Team Needs to Be Able to Do
By James Cotton · Last updated · 12 min read
Part of our topic guide on AI Skills for Business.
By James Cotton, Founder, iO-Sphere
One day on a desk that runs an AI agent
The day starts with the overnight queue. The AI agent has handled most of the chat on its own, the dashboard shows a count of resolutions, and a handful of conversations wait for a person, one from a customer who could not use the chat. Nobody can read every answer the bot gave, so the first decision is which ones somebody does.
A desk like this is on many firms' plans. In AI Adoption Research from the Department for Science, Innovation and Technology (DSIT), published in January 2026 from 2025 fieldwork, 46% of businesses using or planning to use AI name customer service as an area that uses AI today or plans to.
Guidance on AI agents from the Competition and Markets Authority (CMA), published on 9 March 2026, uses this desk as a worked example. It says "Someone with appropriate experience should regularly review how the AI agent responds to different queries, and any complaints or feedback from customers", a sentence the guidance repeats, with small changes, in all four of its examples.
The same example names the failure to watch for, an agent that "keeps giving people the wrong information about the features of a product or their cancellation rights", and the response: "quickly refine the relevant prompts and workflows". Someone in the team has to own those prompts and the knowledge the bot answers from.
Mid-morning, a customer asks the bot for a refund on goods bought under a returns period the business extended last month. The CMA's refund example asks a reviewer to "check that the AI agent's decisions properly reflect the nature of a customer's refund request, and any differences in time limits or amounts due".
Its illustration of a problem is this case: an agent "not always taking into account your extended returns policy when deciding whether a customer is entitled to return goods and claim a refund". Catching it takes a person who knows the returns policy better than the bot's instructions do.
In the afternoon a team leader reviews recorded calls and chat transcripts. Guidance on monitoring workers issued by the Information Commissioner's Office (ICO), whose functions passed to the Information Commission on 30 September 2026, says: "It is not usually proportionate to monitor or record the content of calls in all cases." That guidance is under review after the Data (Use and Access) Act 2025.
Its own example is a call centre that monitors helpline calls "for training and quality control purposes", telling staff through a policy and customers during the call. Elsewhere it says: "You must tell these people that you are recording the call and why."
DSIT's 2025 fieldwork found that 84% of businesses using AI report at least some input or checking of what it produces, and its qualitative research found those checks "were usually down to the individual who had been using the AI". On a support desk that individual is often the agent, which is where the judgement has to be built.
On our reading, two data protection duties reach this desk. Article 5 of the UK GDPR does not mention transcripts, yet its principles cover the personal data in them: kept "accurate" and held "for no longer than is necessary". Since 19 June 2026, section 164A of the Data Protection Act 2018 requires an organisation to acknowledge a data protection complaint within 30 days, and the queue is where many arrive.
The AI skill most firms are asking for is getting further with your own work, and on this desk that work is the queue, the refund and the review; what the same skill means for finance teams and people and HR teams is on their own pages.
What the law makes of the bot's answer
For the overnight answers, the CMA's guidance sets out what it expects of the agent that gave them. It should be prompted to "respect customers' statutory rights and the terms of contracts" and to "avoid misleading customers (both through what they do and do not say)", and tested, "for example through A/B testing or unit testing".
The guidance wants "a human in the loop, actively checking that the AI agent is making correct decisions and generating expected results", and quick action where an agent reaches large numbers of people or vulnerable customers. Its penalty is a ceiling: "You could be fined up to 10% of your worldwide turnover, and possibly forced to compensate affected consumers." How companies govern agentic AI takes the question company-wide.
The statutes behind the guidance do not mention AI. The guidance's customer service example points first to the Digital Markets, Competition and Consumers Act 2024, whose section 225, in force since 6 April 2025, says "Unfair commercial practices are prohibited". Section 226 counts false or misleading information as a misleading action, including information that, "although true, is presented in a misleading way".
The refund at mid-morning brings in the Consumer Rights Act 2015. For services, section 50 treats as a term of the contract "anything that is said or written to the consumer, by or on behalf of the trader, about the trader or the service", where the consumer takes it into account in deciding to enter the contract or in a later decision about the service.
If the bot decides refunds with "no meaningful human involvement", Articles 22A to 22D of the UK GDPR, which replaced Article 22 in full on 5 February 2026, add rules for decisions with "a legal effect" or "a similarly significant effect"; refund, eligibility and account decisions are the kind to flag for review, and our guide to AI in recruitment and HR decisions sets out those rules.
For the overnight customer who could not use the chat, the law to know is the Equality Act 2010. In England, Wales and Scotland, a service provider "must not discriminate against a person requiring the service by not providing the person with the service", and its duty to make reasonable adjustments is owed to "disabled persons generally" (Schedule 2).
For information, the reasonable steps include providing it "in an accessible format" (section 20).
What the vendor counts, and what it bills
The resolution count on the morning dashboard is also the basis of the bill, and three help-desk vendors count three different things for their AI agents; every figure in this section is as the vendor's own page displayed it on 8 October 2026.
| Vendor | What is billed | Price as displayed |
|---|---|---|
| Fin (Intercom) | An "outcome", charged "at most once per conversation" | $0.99 per outcome; billed in US dollars only |
| Zendesk | An "automated resolution": a request "successfully resolved by the AI agent, without any escalation to a human agent" | £1.30 per committed resolution; US$2.00 pay-as-you-go |
| Freshdesk | A "session": "a unique interaction between an end-user and an AI Agent" | $49 per 100 sessions after the first 500 (US dollars) |
Fin's outcome is its billing unit. The pricing page says: "Never pay if Fin can't answer, fails to perform an action or human support is requested." When "A resolved conversation is reopened", "The earlier outcome is refunded." Used with another help desk, Fin has a $49 per month minimum that includes 50 outcomes, and no seat costs.
Zendesk sorts AI agent conversations into tiers. An "Assisted escalation" and a "Contained resolution" do not count against the allowance of 5 or 10 automated resolutions per agent per month. For a "Verified resolution", when the conversation ends, "a verification process is performed by an LLM", a large language model, "that evaluates the text of the conversation to confirm that the customer's request was satisfactorily resolved".
In Zendesk's count, an email conversation ends "72 hours after the first email in the conversation", a messaging one 2 hours after the first message by default (up to 72 hours if raised), and a voice one "immediately upon hangup".
Freshdesk counts sessions. An email session is "a 72-hour window starting from the customer's first email", and "All AI responses within that window count as a single session". The first 500 sessions are free, "limited to once per account".
Fin's homepage claims resolution rates "averaging 76% across 12,000+ customers". Zendesk's pricing page offers to "Resolve up to 80% of even the most complex service issues across any channel." Each is the vendor's claim, in its own terms.
None of these units is defined as a customer confirming the problem is fixed. Reading the bill is a team skill: knowing what each unit counts and when its conversation ends, sampling the conversations counted, and checking each against what the customer needed and whether they came back.
What customers say they want
Late in the afternoon, a customer in the chat asks for a person. Gartner's survey behind the 87% figure asked 3,566 business and consumer customers in February and March 2026. When it asked those unwilling to engage with AI what might change their mind, the most common response was "the ability to switch to a human agent if needed".
Gartner's advice from the same research is that "Service leaders should not use GenAI as a mandatory first step for every issue." The same survey finds that 50% of customers say their interactions are easier when companies use generative AI.
The Institute of Customer Service, a membership body, published "Hype or Reality?" in 2026. Its chief executive's statement of 2 March 2026 says "24/7 availability, faster responses and easy and efficient access to accurate information" matter to customers, while their concerns "revolve around the loss of human contact and oversight, data privacy, and job displacement".
The research finds that "As many as 40% of people doubt businesses can keep their data secure in AI applications", and that "the consumers most optimistic about AI are also the strongest advocates for ethical frameworks and employee training." What agents may put into an AI tool belongs in a staff policy on generative AI.
How the agent's job is changing
In the UK, an estimated 453,500 people work in customer service occupations, 1.4% of everyone in employment, in the Annual Population Survey for April 2025 to March 2026. The official classification puts call and contact centre occupations and customer service supervisors in that group, and leaves out people who serve customers inside other occupations.
In Gartner's survey of customer service and support leaders, conducted in October 2025, nearly 80% of organisations plan to move at least some agents into new roles, "driven by expected automation of routine tasks, and the need for human expertise in complex or emotionally sensitive interactions". Who builds those automations is the subject of AI automation for business processes.
The same research found that 58% of service leaders aim to upskill agents into knowledge management specialists, "acknowledging the need for accurate, continually updated content to support both AI systems and customer self-service interactions". These are aims reported by leaders, and they describe the morning's job: keeping what the bot answers from accurate.
In the Institute's 2026 survey of employees in organisations with more than 50 staff, only 24% believe their organisation has seen a return on AI investment, and even those positive about AI "want clear communication about how it will be used and how it will affect their roles".
Virgin Media O2 announced in May 2026 "a bespoke and officially recognised training programme for contact centre staff", built with the Institute around techniques such as emotional intelligence; its AI tool, Lumi AI, launched in 2025, gives agents "real-time prompts and insights".
The AI Skills for Business Competency Framework (version 3, March 2026) describes a Worker persona through its duties. Two fit this desk: "Engage with and adopt AI tools (e.g. office productivity, HR systems, finance platforms, customer service chatbots) to support day-to-day tasks", and "Apply professional judgement when interpreting AI outputs, ensuring decisions remain fair, accurate, and accountable by remaining compliant with regulatory and legal policies and practices."
The framework's AI Change Champions are "trusted peers within teams who help translate an organisation's AI ambition into everyday practice"; on a support desk, the agent who reviews the bot's transcripts is a natural first one. Skills England's benchmark sets out six AI foundation skills for any worker in England; we look at what AI literacy training should cover separately.
Training so far, and where we would start
The Office for National Statistics (ONS) finds that only 11% of businesses with 10 or more employees that use AI, or are not sure whether they do, have trained more than half their workforce on AI (June 2026, official statistics in development); training everyone is the design question in planning AI upskilling for a whole workforce.
A June 2026 report for England co-developed by Skills England, Skills for AI: what works, has its author argue that AI skills training should be "practical, linked to real tasks and decisions". A support desk has those tasks in its queue every day.
Our recommendation is to train the team on its own queue and transcripts: our data and AI fluency training has agents practise on that work and on realistic case studies and simulations, which we call Prism, and a custom programme can take a team into automating its routine queries, each priced per engagement.
For the person who will redesign the service around the bot, we would choose the AI Transformation (Level 4) apprenticeship, and for whoever does the team's analysis work, Data & AI Essentials (Level 3), matching each to the duties in the job.
Age at the start and levy position decide the employer's share (the levy is a charge on employers with a pay bill over £3 million); the sums for an employer outside the levy are in our guide to AI training for small businesses.
By evening, what the day's reviews found, from a returns rule the bot did not apply to a customer who needed a person, has gone back into the bot's prompts and the knowledge it answers from, and tonight's queue will be answered from what the team corrected.
Frequently asked questions
When does a chatbot have to tell customers it is a bot?
The CMA's guidance says to tell customers when dealing with AI rather than a person "might affect people's decisions". From 2 August 2026, EU AI Act Article 50 requires providers to design such systems so people are told unless that is obvious to a reasonably well-informed, observant and circumspect person, and reaches non-EU providers under Article 2(1) "where the output produced by the AI system is used in the Union".
Can an AI agent issue refunds?
Yes: the CMA's guidance names processing refunds among the uses a business might make of an AI agent. Its refund example expects the design to take account of "consumers' statutory rights under the Consumer Rights Act 2015 (for example, if goods are faulty)", the Consumer Contracts Regulations 2013 for a customer who changes their mind, and the business's own terms, such as extended returns periods.
Who is responsible if the bot quotes the wrong price?
It depends on whether the customer took the bot's answer into account, and what else they were told at the time. The CMA's example says "you must respond accurately to queries about prices, products and rights". Section 225 of the 2024 Act counts "a person acting in the name of, or on behalf of," a business as a trader, and section 50 covers what is said on its behalf.
Do customer service agents need to learn prompting?
Yes, though on a support desk the prompts that matter most are the bot's own instructions, which someone in the team maintains. Skills England's benchmark lists "writing clear instructions for AI tools" and "adjusting how an AI tool works to get better results" among its six foundation skills, and the AI Skills for Business framework's Worker keeps track of "where human intervention is required".
Upskilling a whole team?
Tailored data and AI training for organisations, from data literacy to technical upskilling, built around your team's real work.