A founder running a 12-person professional services firm asked me recently whether they should add an AI assistant to their customer support. They’d seen the Klarna figures: 2.3 million conversations handled in the first month after launch, resolution times cut from 11 minutes to under 2, and a profit impact estimated at $40 million across 2024. “Could we do something like that?” they asked. The honest answer is: sometimes yes, often in part, and sometimes not yet. The question worth sitting with is which situation you’re actually in.
What does “Klarna-style AI support” actually mean for a smaller firm?
“Klarna-style AI support” means an LLM-based assistant connected to your live systems: able to answer customer questions, handle routine requests such as refunds or rebookings, and pass complex cases to humans. UK platforms including Twig and Intercom Fin now offer this to smaller firms. The choice you’re weighing is whether your volume, content quality, compliance posture, and operational readiness are in the right place to make it worthwhile.
Klarna’s setup has three layers. A generative AI front-end handles the conversation using a large language model. Back-end integrations give it access to live customer data, including balances, order history, and case records, so it can act as well as answer. A confidence-based routing layer passes queries to human agents when risk is high or the situation falls outside what the AI can handle reliably. Klarna ran this across 23 markets and 35 languages, around the clock.
UK platforms replicate the pattern at SME scale. You supply the content layer: your FAQs, refund policy, booking rules, and case history. The platform provides the AI, the retrieval mechanism that searches your documents in real time, and the helpdesk integration. Many deployments for firms of five to fifty people go live in weeks once the documentation is prepared and routing rules are confirmed.
The starting-mode decision matters most: co-pilot, where the AI drafts and a human reviews before sending, versus autonomous, where certain low-risk queries are handled end-to-end by the AI. Both are legitimate. The sequence is where many firms go wrong.
When does AI support make sense for a UK SME?
The pattern works well when a significant share of your support tickets are routine: password resets, booking changes, basic billing queries. Klarna directed exactly these to its AI and kept human agents for complex or premium cases. For a UK SME, the practical threshold is clear: at least one person’s time absorbed by repetitive support, clear written policies to draw on, and errors in those topics that are cheap to fix.
Beyond volume, two other conditions matter as much. First, your content has to be in good shape. Klarna’s speed improvement and 25% reduction in repeat contacts came partly from well-structured system data and consistent policies. If your FAQs are out of date or your refund rules live mostly in your team’s heads, the AI fills gaps with guesses, and guesses are where errors come from.
Second, you need a way to measure whether it’s working. The ICO’s guidance on automated decisions requires transparency and human oversight where personal data is involved. The FCA’s Consumer Duty applies to any customer-facing channel in a regulated firm, AI-mediated or not. That means you need escalation metrics, complaint rates for AI-handled tickets, and a clear list of topics the AI must never handle autonomously, such as hardship cases, regulated advice, or formal complaints.
Running 30 to 60 days in co-pilot mode before removing human review gives you an evidence base. You want to know your override rate before you extend the AI’s autonomy, not after you have already done so.
When should you hold off?
Three situations tip the balance against deploying AI support autonomously. If you operate in a regulated sector, the FCA’s Consumer Duty and ICO requirements mean accountability for automated decisions stays with you, not the platform. If your documentation is thin or inconsistent, the AI fills gaps with guesses. And if your reputation rests on high-touch service, cutting handling time can quietly erode the thing customers actually pay for.
Each situation deserves a specific response. Regulated firms in financial services, health, or legal need to treat AI customer support as part of their governance framework. The FCA has been clear that accountability for customer outcomes cannot be delegated to technology providers, and that AI-mediated interactions are subject to the same Consumer Duty obligations as any other channel. The EU AI Act classifies certain AI systems in financial services as high-risk, relevant to UK firms serving European customers, with strict requirements around human oversight and technical documentation.
Sparse or inconsistent documentation creates a different problem. Without clean policies and well-structured FAQs, AI retrieval produces errors rather than answers, and you spend more on remediation than you save on handling time.
The brand risk is subtler but real. Klarna recognised it. After its initial AI-led approach, the company moved toward an “Uber-style” flexible human layer for complex and premium interactions, acknowledging that some customer situations are better served by people. If high-touch care is central to what your clients pay for, the same logic applies to your firm.
What does it cost to get this wrong?
The consequences are concrete. In 2024, a Canadian tribunal held Air Canada liable for incorrect refund advice its chatbot gave a customer, ordering the airline to honour the promise. A UK council took its AI chatbot offline after it produced inappropriate responses. ICO fines under UK GDPR reach up to £17.5 million or 4% of global turnover.
The Air Canada ruling is particularly instructive. The airline argued that its chatbot was effectively a separate entity and that its errors were not the airline’s responsibility. The tribunal rejected that argument. If your AI makes a promise to a customer, you are held to it.
The ICO has shown a similar reach on data. When HMRC used a voice identification system without adequate consent, the regulator required deletion of 5.1 million records and a change in approach. The point is direct: “we used a vendor’s technology” is not a compliance defence.
Operationally, a platform outage or model error at scale can leave you with a surge of unresolved tickets and no human capacity to absorb them. The FCA’s operational resilience framework requires firms to think through exactly this scenario, setting impact tolerances for important business services before a disruption happens, not during one.
The cost of getting it wrong runs in three phases: immediate firefighting and refunds, medium-term remediation and added controls, and long-term customer attrition that you may never fully trace back to the root cause.
What should you ask a vendor before you sign?
Before signing a contract, ten questions will tell you more than any demo. Group them under data protection and security, governance and control, and commercial terms. The answers will also show whether the vendor has deployed this with firms like yours or is effectively asking you to be an early adopter. If they cannot give you before-and-after metrics from comparable SMEs, that is already your answer.
On data protection: ask where your customer data is processed and stored, whether you can disable training on your tickets so they do not feed the general model, and what audit logs you receive from prompt through to response. The ICO and NCSC require you to understand these answers before deployment, not after a breach.
On governance: confirm you can set topic-level policies, so the AI never handles formal complaints autonomously and always escalates any vulnerability flag. Ask how the platform manages hallucinations, what the confidence threshold for routing to a human is, and what the SLA looks like for a platform outage.
On commercial terms: ask for actual before-and-after numbers from deployments similar to yours rather than headline claims based on Klarna’s scale. Pricing models vary considerably, from per-seat subscriptions to per-resolution fees, and your ticket volume determines which works better for you. Ask whether you can pilot in co-pilot mode with a clear exit option and data-deletion guarantees before any autonomous deployment begins. Firms that skip this step often find they have no negotiating position once the integration is live.
If you want to think through where your business sits on this decision, Book a conversation.
