Customer Service Outsourcing After AI: Where Human Agents Win Loyalty

The commercial case for automating customer service rests on an assumption that customers will accept the substitution. The evidence does not yet support it. Gartner surveyed 5,728 customers in December 2023 and found that 64% would prefer that companies did not use AI in customer service, and that 53% would consider switching to a competitor if they learned a company planned to use it. The most common concern was that AI makes it harder to reach a person.

Qualtrics reinforces the point in its 2026 Consumer Experience Trends Report, based on more than 20,000 consumers in 14 countries that include the UK, the UAE, Australia and New Zealand. Half of consumers (50%) worry that company AI use will stop them reaching a human, and 53% fear misuse of personal data. Nearly one in five who had used AI for customer service saw no benefit, a failure rate almost four times that of AI use in general. These are vendor and analyst surveys, but they agree in direction.

For a Chief Customer Officer the implication is that deflection rate is an incomplete measure of success. A contact that AI closes cheaply, but which leaves the customer unable to reach anyone, is a saving this quarter and a churn risk next year. Qualtrics also finds that nearly half (47%) of bad customer experiences lead to decreased spending, which gives a way to price the downside: each automated failure carries an expected loss of revenue, not only a repeat contact.

The evidence is not confined to one geography. Qualtrics spans fourteen countries, including the UK, the UAE and Australia, and the UK index shows customers rewarding quality. We would therefore resist a single global automation policy. Test acceptance market by market and journey by journey, and report switching intent as a managed risk beside the cost saved, so the board sees both sides of the trade.

What to do: before extending automation to a new contact type, estimate the revenue exposed in that journey, test the AI route against a human control group, and track repeat contact and later spend alongside containment. Treat the 64% as a design constraint. The service has to be good enough that customers stop caring whether a person or a machine answered.

The workforce story is less dramatic than the headlines. Gartner's survey of 321 service and support leaders, fielded in September and October 2025 and published in April 2026, found that 85% are adding tasks and responsibilities to frontline agent roles and 75% are moving agents into entirely new roles. Only 31% have implemented or are planning frontline layoffs through the first quarter of 2027, while 63% are reducing headcount gradually through attrition.

That is consistent with Gartner's June 2025 prediction that, by 2027, 50% of organisations that expected to significantly reduce their customer service workforce will abandon those plans. A March 2025 poll of 163 leaders found that 95% plan to retain human agents. Pressure to move is nonetheless intense: 91% of leaders report pressure from executive leadership to implement AI in 2026. The gap between that pressure and operational reality is where outsourcing decisions are most exposed.

Klarna is the most cited reversal. The company reported in 2024 that its AI assistant handled 75% of customer chats. As reported by Entrepreneur in May 2025, its chief executive then described AI as cheaper than human staff but of lower quality, and said: "I just think it's so critical that you are clear to your customer that there will always be a human if you want." One company's experience proves little on its own, but it shows what happens when cost is the only test.

The 63% reducing through attrition is the quieter and more important number. Attrition lets a company shrink the frontline without a redundancy programme and without the sudden loss of knowledge that comes with one, but it changes the mix of work faster than the mix of skills. The agents who remain inherit the contacts AI could not close, which are by definition the harder ones. The consequence is that average handling time on the human tier will rise even as total volume falls, and contracts that penalise handling time will punish exactly the right behaviour.

What to do: do not size an outsourcing contract on a straight-line assumption that AI removes a fixed share of volume. Model three scenarios for contained, reclassified and retained work, and ask each provider which roles its agents will move into. Gartner reports that 58% of leaders plan to upskill agents as knowledge management specialists. That matters because AI is only as reliable as the knowledge it draws on.

The most useful routing rule is not whether a contact arrives by chat, voice or email. It is how emotionally loaded the moment is and how much revenue or risk rides on it. Customers will accept a machine for a delivery update or a password reset. They are far less willing to for a bereavement, a disputed charge, a safety concern or a renewal that decides a large account.

Gartner's April 2026 release reports that 54% of US customers, in a survey of 5,801 conducted in January and February 2025, trust human agents more than AI for product or service recommendations, while 32% trust AI more. Trust is divided and context-dependent. That argues for routing on the customer's situation and the stakes involved, rather than on whichever channel is cheapest to staff.

The UK offers a useful reading of what customers value. The Institute of Customer Service's January 2026 UK Customer Satisfaction Index stands at 78.2 out of 100, up 2.1 points on a year earlier, and 35.6% of customers say they prefer excellent service even if it costs more, up 4.3 points. Willingness to pay for quality is rising just as automation spreads. In our view that makes service a differentiator again, and not simply a cost line to be minimised.

A further point concerns the customer who stays silent. Qualtrics finds that only 29% of customers now communicate directly with an organisation after a bad experience, down 7.5 points from 2021, and that 30% say nothing at all. If fewer unhappy customers tell you, a poor automated experience will not show up in complaint volumes. It shows up later, as reduced spending. That is one more reason to send high-stakes moments to people, who can detect distress and ask the follow-up question that a bot rarely will.

What to do: build a contact taxonomy of 30 to 50 types and score each on emotional load, financial value, regulatory exposure and complexity. Automate the low-low cell without apology, assist agents in the middle, and reserve named, empowered humans for the high-high cell. Review the scoring quarterly, because AI capability moves and the boundary will move with it.

The largest offshore hub shows what a mature response looks like. The Philippine industry body IBPAP reports, as carried by Philstar and Malaya, that IT-BPM revenue rose from $32.5 billion in 2022 to $35.5 billion in 2023 and $38 billion in 2024. For 2026 it expects $42 billion in export revenue and nearly 1.97 million jobs. Growth continued through the very period in which generative AI arrived in contact centres.

IBPAP's stated priorities are instructive: scaling AI responsibly, and upskilling the workforce to take on higher-value work. The business model of a serious provider is shifting from seat-hours to resolution, and from handling volume to managing exceptions, quality and knowledge. Buyers should expect proposals that price outcomes and bundle AI into the service, not ones that leave the client to procure and integrate it separately.

The implication is that a provider's AI roadmap is now a selection criterion on a par with its price per hour. A provider that automates its own work well can pass savings on. One that does not will protect seat revenue and slow your transition. Ask which of your contact types the provider would automate first, what it expects to retain with humans, and how savings are shared once automation goes live.

Cost still matters, and the offshore model is not obsolete. Klarna's experience, in which AI was cheaper but of lower quality, shows that the cheapest option per contact is not the cheapest over a customer lifetime. Offshore and nearshore hubs remain the right home for Tier 1 volume where processes are well documented and escalation paths are short. What changes is the proportion of work that stays there, and the expectation that the provider operates AI tooling, knowledge management and quality analytics as part of the service rather than as chargeable extras.

Procurement teams should also test the claim that AI savings will be passed on. A provider paid per seat has little incentive to remove seats. Ask for a transparent baseline of volumes and handling effort by contact type, agree which automation gains are shared and on what schedule, and include a clause that re-bases pricing as containment rises. Without it, the client funds the AI investment and the provider keeps the productivity gain.

What to do: run selection as a joint design exercise. Ask two or three providers to propose a human-AI operating model for the same sample of 500 real contacts from your own history, then compare containment, resolution and escalation quality rather than slide decks. Prefer providers willing to accept outcome-linked fees on the human tier.

A promise to always offer a human is moving, in some markets, from brand choice to legal requirement. In consumer credit, the EU's Consumer Credit Directive (Directive (EU) 2023/2225, dated 30 October 2023) states that the consumer should have the right to obtain human intervention on the part of the creditor. It is a narrow rule, but it signals how European lawmakers think about automated customer-facing decisions.

We should not overstate this. Most service contacts are not covered by a statutory right to a person, and we are not aware of a general equivalent in the UK, the Gulf or the United States. Regulators in financial services, utilities and health do, however, already examine how firms treat vulnerable customers, and an automated flow with no route to a human is hard to defend in that setting. The direction of travel is consistent.

Where the law does not require it, customers increasingly do. Qualtrics finds that half of consumers worry AI will prevent them from connecting with a human being, and Klarna's chief executive has made the same point from the supplier side. A published, tested, easy route to a person is therefore among the cheapest trust assets a company can build, and it costs little if the human tier is sized sensibly.

Data protection compounds the issue. Qualtrics finds that 53% of consumers fear misuse of personal data when companies automate interactions. An outsourced AI layer that processes personal data across several jurisdictions raises questions about where data is held, who can read transcripts and whether they are used to train models. These belong in the contract, not in a later audit. Legal counsel in each market should review them, and we would treat the answers as part of supplier risk assessment alongside price.

For the UK specifically, the commercial argument is already visible even where the legal one is not. The UK index records the highest share of experiences ever right first time, at 83.2%, and a rising share of customers willing to pay for excellent service. A company that automates in ways that damage first-time resolution would be moving against the market, not with it. That is a reputational as well as a financial exposure.

What to do: map each contact type to the rules that apply in each market you serve, including financial-services conduct and data-protection requirements. Hard-code a human escape route into every automated flow, test it monthly with mystery contacts, and publish the commitment. If an outsourcer runs the AI layer, make the escape route a contractual service level with credits when it fails.

Where the human tier sits matters as much as whether it exists. The default for two decades has been to place people wherever labour is cheapest and add technology on top. After AI, the human tier handles the hardest contacts, with the most context, the most discretion and the highest stakes. Those are the contacts least tolerant of poor language fit, slow escalation or an agent without authority to act.

Our view is a four-tier design. Tier 0 is AI for authenticated, reversible, rules-based tasks. Tier 1 is AI-assisted agents in a scaled hub for routine and moderately complex work. Tier 2 is specialist agents in the customer's region, time zone and language for disputes, complaints and retention. Tier 3 is named owners inside the company for strategic accounts and policy exceptions. Outsourcing can serve Tiers 0 to 2; Tier 3 should stay close to the business.

Gartner's finding that 85% of leaders are adding responsibilities to frontline roles, and 75% are moving agents into new roles, points the same way: the human tier is becoming smaller, more skilled and more empowered. In Saudi Arabia and the UAE, where Arabic dialect and expectations of personal service raise the bar, we would expect an in-region Tier 2 to matter more than in English-language Western markets. We have not found a robust public benchmark of regional contact-centre economics, so we would test that assumption in a pilot rather than assume it.

Hiring and retention follow from the design. If the human tier is smaller and handles harder contacts, the profile of the agent changes: more judgement, more product knowledge, more tolerance of ambiguity. Gartner reports that 58% of leaders plan to upskill agents as knowledge management specialists, which suggests agents will increasingly maintain the knowledge that AI uses. Pay, progression and training should reflect that, and attrition in this tier should be watched closely, because each departure removes context the system cannot easily replace.

What to do: assign decision rights as deliberately as headcount. A Tier 2 agent who must seek approval for a modest goodwill gesture is a Tier 1 agent with a better title. Set refund, credit and exception limits by tier, and measure first-contact resolution and time to a decision for each. If a provider cannot operate with those limits, it cannot run the tier.

Cost per contact rewards the wrong behaviour once AI is in the flow, because containment can be inflated simply by making humans harder to reach. Contracts should move to cost per resolved issue, with quality-weighted incentives, escalation-time service levels, and shared savings on automation the provider delivers. Penalties should attach to avoidable repeat contacts and to failures of the route to a human.

Benchmarks help. The UK index records 83.2% of customer experiences as right first time, the highest level in its history, which shows that the better organisations are raising the bar on resolution, not merely on speed. Your own baseline for first-contact resolution, by contact type, is the right starting point for any outcome-linked fee.

Operational measures are not enough. Gartner's finding that 53% of customers would consider switching if they learned a company used AI, and Qualtrics' finding that 47% of bad experiences reduce spending, both point to the same lesson: pair containment and resolution with a loyalty measure. Track renewal, repeat purchase and share of wallet for customers who experienced each route, and report them beside cost.

Governance is the last piece. With 91% of service leaders under pressure from executives to implement AI, someone must be accountable for the customer-side consequences of automation and not only the cost-side benefits. We recommend a single named owner for the end-to-end service model, with authority across in-house teams and providers, and a quarterly review that looks at loyalty outcomes as carefully as at savings. Without that owner, each function optimises its own metric and the customer absorbs the difference.

Finally, plan for reversibility. Gartner expects many organisations to abandon workforce-reduction plans by 2027, and Klarna's experience suggests that rebuilding a human tier is slower than dismantling one. Keep a minimum viable human capability, including trained leads and documented procedures, even for contact types you have fully automated. Contracts should preserve the right to move work back, with step-in rights over people, knowledge and data.

What to do over 90 days: classify the top 40 contact types, pilot one automated and one human-led flow with a control, rebuild the provider scorecard around resolution and loyalty, and agree the escape-route commitment. Decide at the end which contact types move up or down a tier. This is a design exercise first and a procurement exercise second.

Karter McKinley helps customer leaders design the contact taxonomy, the routing logic and the provider scorecard that turn this argument into a contract. Through our AI and technology work we can support implementation alongside your technology and outsourcing partners, with the human tier and its decision rights specified as carefully as the automation.