AI customer support for UK businesses
An AI support layer grounded in your real help content and order data — answers the common questions instantly, assists your agents on the rest, and hands off cleanly to a human the moment it's out of its depth.
The same questions, over and over, all day.
Most support queues are dominated by a small set of repetitive questions — where's my order, how do I return this, what's your policy on X. Agents answer them all day, which leaves no time for the genuinely hard cases that need a human. AI support grounded in your real content can answer the repetitive ones instantly, freeing agents for the cases that actually need them — without the customer-trapping failures that give AI support a bad name.
- Repetitive questions eat the day — A handful of question types make up most of the volume. Answered by hand, every time, they crowd out the cases that need real attention.
- Scripted bots that don't help — The old chatbots followed decision trees and frustrated everyone. Customers learned to type 'agent' immediately. That's not what good AI support is.
- Agents without context — Agents tab between systems to find the order, the history, the policy. The answer exists; finding it is the slow part.
Resolution over deflection. Always a way to a human.
We build AI support to resolve, not to deflect for its own sake. It's grounded in your real content so it doesn't make things up, it hands off to a human the instant it's uncertain or asked, and we measure satisfaction and resolution — not how many customers we stopped reaching a person.
- Independent — no commissions — We use your helpdesk's native AI where it fits or build a custom RAG layer where it doesn't, picking what's right for your stack. No helpdesk or AI vendor commissions.
- Grounded, with citations — RAG grounds every answer in your real help content and order data, with sources — so it answers from your actual policy and this customer's actual order, not a guess.
- Clean handoff, every time — The AI hands to a human the moment it's out of its depth or the customer asks. No loops, no trapping. A human is always one step away.
Six AI support builds
Grounded AI applied where it genuinely speeds resolution — for customers and for agents.
A four-phase engagement, priced flat
No hourly billing. No scope creep. You know what you're paying and what you're getting before we start.
We analyse your ticket mix — what's repetitive and genuinely AI-answerable, what needs a human — and assess your help content and order-system access. Output: a build plan with the handoff design and a running-cost forecast.
We decide native helpdesk AI vs custom RAG, define what the AI answers vs escalates, and design the handoff and guardrails. You sign off on scope and cost.
We build the RAG layer grounded in your content, integrated with your order systems and helpdesk, with clean handoff. We test against real tickets and measure resolution and accuracy before it faces customers.
Documentation, training and a check-in 90 days after launch to measure resolution rate, CSAT and agent time saved. After that, fractional CAO retainer or done.
What AI customer support actually is
AI customer support uses a large language model, grounded in your own help content and order data, to answer customer queries and assist human agents. The grounding is the important part: this is not a chatbot guessing from general knowledge, and it's not the scripted decision-tree bot everyone learned to hate. It's an AI that retrieves your actual returns policy and this customer's actual order, and answers from that — citing the source.
The honest framing, consistent with our whole approach to AI: the AI handles what it's genuinely good at — the repetitive, answerable questions — and a human handles everything else, with the AI assisting them. It resolves; it doesn't trap. This page is the AI build; the broader function, including the rule-based parts, is on the customer service automation page, and it's one pattern within our AI automation practice.
RAG: why grounding matters
RAG — retrieval-augmented generation — is the technique that makes AI support trustworthy. Instead of answering from its general training (where it can confidently invent a policy you don't have), the AI first retrieves the relevant information from your content — help articles, policies, the customer's order record — and generates its answer from that, with citations.
The difference is night and day for support:
- Without RAG: the AI guesses from general knowledge. It might invent a 30-day returns window when yours is 14, confidently and plausibly. Dangerous.
- With RAG: the AI answers from your actual returns policy and this customer's actual order date, and shows where it got the answer. Trustworthy and auditable.
We never build customer-facing AI support without RAG grounding. A support AI that makes things up damages the brand faster than a slow queue ever would.
The three failure modes we design against
Most AI support people have suffered fails in one of three ways. We design against all three explicitly:
- The bot that won't let you reach a human. Ours hands off the moment the customer asks or the AI is uncertain. A human is always one step away — never buried.
- Deflection chased as a metric. Optimising for 'tickets deflected' incentivises hiding the contact button and frustrating people. We optimise for resolution and satisfaction. Deflection is a by-product of genuinely answering, not a goal.
- Confident wrong answers. RAG grounding plus confidence handling means the AI answers from your real content and escalates when unsure, rather than inventing a plausible-sounding wrong answer.
What it builds on
Good AI support needs three things connected:
- Your help content — articles, policies, FAQs — as the knowledge the AI retrieves from. (If this is thin, improving it is part of the work, and the AI's gap-detection helps.)
- Your order and customer data — Shopify, CRM, ERP — so it answers with real, specific information, not generic responses.
- Your helpdesk — Gorgias, Zendesk, Intercom, Freshdesk, Front, Help Scout — so AI and human agents work in one place, with clean handoff.
We use the helpdesk's native AI (Gorgias Automate, Zendesk AI, Fin by Intercom) where it fits and a custom RAG layer where it doesn't — independent, no commissions.
What it costs
- Focused build: £8k–£15k fixed.
- Broader engagement (deep integration + agent-assist): £15k–£25k fixed.
- Plus ongoing AI usage and any helpdesk AI add-on fees (forecast in discovery).
- Fractional CAO retainer: £5k–£15k per month.
The £1,500 Discovery Sprint is the paid scoping step. Flat fees, no commissions.
When AI support isn't the answer
If your volume is low, or your queries are mostly complex and human-judgement-heavy, the ROI isn't there — better to invest in good agents and rule-based customer service automation (triage, WISMO, routing) without the AI layer. And if your help content doesn't exist, that comes first — the AI can only ground in content you have. We'll tell you honestly in discovery.
How this fits with the wider Watermelon model
This is one of four AI build patterns under the AI automation hub, alongside AI agents, document AI and internal AI assistants. It's the AI layer within the broader customer service automation function; for ecommerce specifically see ecommerce & DTC.
Ready to talk?
Bring your ticket volume and the questions you answer most. The free 30-minute call will tell you what AI can genuinely resolve, what should stay human, and what we'd build.
What can AI genuinely resolve?
30 minutes. No deck. Bring your ticket volume and your most common questions. We'll tell you what AI can resolve, what stays human, and what we'd build.