AI-Assisted Customer Support

What we automate, and what we will not

Our position on AI in customer support is narrower than most suppliers advertise, and it is deliberate.

We automate work that removes typing and lookup from an agent's day, where the cost of an error is low and a human checks the output. We do not automate the judgement, the tone, or the decision about what happens to a customer. Every capability on this page sits on one side of that line, and we will tell you which side.

The reason is not caution about the technology. It is that the failure mode of automated support is expensive and delayed: the contacts a bot fails to resolve arrive later, angrier and harder to handle, frequently after the customer has already complained publicly.

What AI does on our accounts

• Conversation summarisation - condensing a long thread before escalation, so the next tier starts with the case rather than the transcript.
• Intent classification - tagging inbound email, chat and tickets by reason and urgency on arrival, with an agent confirming.
• Knowledge retrieval - surfacing the relevant article while the agent is still reading the question.
• Draft replies - a first version for a human to review, edit and send. Never sent unreviewed.
• Sentiment and escalation signals - flagging interactions that merit a supervisor's attention or a quality review.
• Quality review selection - choosing which interactions a human reviewer should examine, which materially raises the hit rate of a limited review budget.
• Translation assistance - supporting an agent working across languages, with a fluent speaker responsible for what is sent.
• After-contact work - drafting the case note and disposition for the agent to confirm.

Each of those is bounded, checkable, and reversible before a customer sees it. That is the test we apply.

What we do not do

We do not put an unsupervised bot in front of your customers as a deflection strategy. Automated deflection reduces contact counts, which looks like success, and frequently raises total cost. The contacts it fails to deflect return worse, and the ones it wrongly resolves become complaints you find out about later.

We do not let automation decide refunds, goodwill, cancellations, complaint outcomes or anything involving money. We do not let it handle a customer who is distressed, vulnerable, or describing a safety issue. We do not send generated text to a customer without a human reading it. And we do not use your customers' conversations to train general-purpose models.

Where you want a customer-facing assistant, we will help you scope one - with a clear handover to a person, an obvious route to bypass it, and honest measurement of what it actually resolves rather than what it deflects. But we will not tell you it replaces the team.

Grounding: the control that matters

A model asked not to invent things will still invent things. A system prompt is a request, not a control.

So anything AI produces on our accounts is grounded in an approved source - your knowledge base, your policy documents, the customer's own record - and the agent can see what it was drawn from. Where the source does not support an answer, the correct output is "no answer found", not a plausible sentence.

This is the same principle we apply to our own site. The optional assistant on this website answers only from a curated, approved knowledge base, and a post-generation validator rejects any reply containing a price, percentage, headcount, certification or client name that cannot be traced back to an approved source - replacing it with a fallback rather than showing it. It is disabled unless deliberately configured, and it stays disabled without curated content behind it. We would rather show a visitor nothing than an invented figure.

What the agent's day actually looks like

Worth being concrete, because "AI-assisted" is otherwise a slogan.

A ticket arrives already tagged by intent and urgency. The agent opens it and sees a summary of the customer's previous three contacts rather than reading them. The knowledge panel is already showing the two articles most likely to be relevant. A draft reply is available, which the agent reads, corrects and personalises - or discards, which happens often and is fine. When they close the case, the note is drafted and they confirm or amend it.

The time saved is real and it goes back into attention rather than into throughput targets. That distinction is the point: used to increase handled volume per agent, assistance produces the rushed contacts and repeat volume it was supposed to prevent.

Data protection

Automation is subject to the same access, retention and processing rules as a person. Where we are your processor, any AI assistance in scope is documented in the services agreement rather than introduced quietly, and we do not send your operational or customer data to a general-purpose service outside those agreed arrangements.

Where automated processing touches personal data, that belongs in your privacy notice, and we will tell you what to say in it. See security and data protection.

Measuring whether it helps

Assistance is worth having only if quality holds or improves. So we report the quality measures alongside the efficiency ones: first contact resolution, repeat contact, reopen rate, satisfaction and quality score, next to handling time and volume.

If handling time falls while repeat contact rises, the assistance is producing faster wrong answers and we say so. Also reported: draft-reply acceptance rate, classification accuracy, and knowledge-retrieval hit rate - because a suggestion agents routinely discard is a cost, not a benefit. See the measures we report. Figures published elsewhere on this site are historical or representative and campaign-dependent.

Getting started, and what it costs

AI assistance is part of how our teams work rather than a separate product, so it is included in the dedicated agent rate - published rates and what moves them. Where you require a specific platform or licence, that is itemised separately.

Bring your handling time, repeat contact rate and knowledge base state. The knowledge base is the binding constraint: assistance grounded in thin or outdated content produces confident wrong answers, so improving it is usually the first and highest-return piece of work. Related: technology, quality assurance outsourcing. Talk to us.