You know the feeling from the customer’s side. You had one simple question for a supplier, the chatbot misread it three times, and you ended up typing “speak to a human” in capital letters. You were in good company. In 2023, 36% of UK consumers said chatbots made getting service more difficult. Now you are on the other side of the counter, weighing up AI for your own support inbox, and the question you typed into Google was probably some version of “what is the best AI for customer support”. That question has a better version. The tool you pick matters far less than where you put it.
What choice are you actually facing?
The real decision is placement, whether AI sits in front of your customers as a bot, or behind your team as an assistant that drafts replies, summarises tickets, and finds the right answer in your help content. Both routes use similar technology. They carry very different risks, and choosing the wrong one for your situation is how good firms end up delivering bad service.
The scale of the gap is worth sitting with. In 2024, 54% of UK contact centre operations were using AI or automation on at least one customer channel, yet only 24% rated their own use of it as highly effective. That gap owes more to deployment decisions than to the software. The main platforms, Zendesk AI, Intercom’s Fin, and Salesforce’s Einstein tools, all offer both modes, so picking a vendor settles very little. The question that actually shapes your customers’ experience is which mode you switch on, for which queries, and with what route back to a person. For an owner-managed business without a dedicated technology team, getting that placement call right matters more than any feature on the comparison chart.
When does a customer-facing bot earn its place?
A customer-facing bot works when the queries it handles are simple, informational, and reversible, when your help content is accurate and current, and when there is always a fast, obvious route to a human. If any of those three conditions is missing, the bot will frustrate more customers than it helps, whatever the vendor demo suggested.
Customer tolerance backs this up. Ipsos found 67% of UK consumers were comfortable with AI answering simple FAQs, but only 17% were comfortable with AI making decisions about financial products without a human involved. BT designed its digital assistants around exactly that line. The bots handle password resets and simple billing questions, and they pass complaints and vulnerable customers to a person quickly.
The gentlest version of a customer-facing deployment is AI search over your existing help pages, a widget that surfaces the right article rather than improvising an answer. When the underlying content is clean, this can be configured in days, and vendor case studies report ticket deflection of up to 30%. The same mechanism cuts the other way. If your help pages are outdated, the AI will return wrong answers with complete confidence, and ContactBabel names poor data quality as one of the biggest barriers to AI working in customer service. Whatever you deploy, keep it well away from binding decisions such as account closures, eligibility calls, or refund refusals.
When is AI behind your team the better call?
Agent assist, AI that helps your people rather than replacing them, is the better call when queries are sensitive or regulated, when your records and help content are imperfect, or when you simply cannot afford a public mistake. Your team stays on the front line while the AI drafts replies, summarises tickets, and surfaces the right knowledge before each response goes out.
When ContactBabel asked UK customer experience leaders about their current AI use, the top answer was supporting agents with real-time guidance rather than replacing them. Early deployments report agents saving 20 to 30% of their time on email drafting and note-taking. The gains are less dramatic than a bot’s headline numbers, but so are the failures. A person reads every AI suggestion before it reaches a customer, which catches the fabricated answer, the wrong tone, and the edge case the model missed.
That safety only holds if your team treats suggestions as drafts. When staff rubber-stamp whatever the AI proposes, errors reach customers anyway, just with a human signature on them. Build review into the workflow, audit what the AI suggested against what was actually sent, and make it easy for agents to flag bad suggestions. This route also suits regulated work. If you handle queries about money, health, or employment, keeping a person on the front line aligns with the ICO’s expectations on human oversight and keeps you clear of the automated-decision rules entirely.
What does getting this call wrong cost?
A bad automation call costs you three ways, in lost customers, regulatory exposure, and security risk. Roughly one in four UK consumers has abandoned a business after a poor chatbot experience, and the fines available to the ICO for unlawful automated decision-making run to £17.5 million or 4% of turnover, whichever is higher.
The churn is the visible cost, and you rarely see it coming, because frustrated customers tend to leave rather than complain. The regulatory costs sit further back. Under UK GDPR Article 22, customers have the right not to be subject to solely automated decisions with legal or similarly significant effects, which catches bots that reject claims, close accounts, or refuse service without human review. The ICO has warned that customer-facing AI often involves large-scale processing of personal data, which points towards a data protection impact assessment before launch. If you are FCA-regulated, the Consumer Duty applies to automated support in full, and you will need to evidence that your bot delivers fair outcomes, including for vulnerable customers. Firms selling into the EU face the AI Act on top, which treats systems influencing credit, insurance, or access to essential services as high-risk.
Security completes the picture. The NCSC warns that bots wired into back-office systems open routes for prompt injection and data theft, and the cost of a breach response comfortably exceeds whatever the automation saved.
What should you ask before you decide?
Five questions settle the decision better than any feature comparison. What share of your contacts are genuinely simple? How does a customer reach a human, and how fast? Is your help content accurate enough to build on? What personal data will the AI see? And can you export your data and leave if the vendor lets you down?
Honest answers to the first two rule out more bots than any procurement process. Once you have a mode in mind, put these to every vendor on your shortlist:
- Does the system answer only from our content, or can it improvise? Look for a strict retrieval mode you can enforce.
- Can we see which articles the AI uses, so we can fix the gaps it exposes?
- Where is our customer data hosted, and can we switch off training on our conversations?
- How does hand-off to a human work, and can we set escalation targets the system has to honour?
- What do we get back if we leave, and in what format?
The last question carries more weight than it appears to. The CMA has flagged how much of the AI market depends on a small number of foundation model providers, and an exit route protects you if your vendor’s prices or terms shift underneath you.
And if the honest answer on your help content is no, that is your project for this quarter, ahead of any purchase. Fix the help pages, standardise the reply templates your team already uses, and clean the customer records. Every one of those steps improves service on its own, and each one makes whatever AI you eventually buy work dramatically better. The customer typing “speak to a human” in capital letters is telling you exactly where the boundary sits. Place the AI on the right side of it.
