That enquiry again — a colleague gave the very same explanation only yesterday.
It was in a Monday-morning meeting room that Oda, who heads the customer support function at a B2B service company, let that slip. Six months earlier, as enquiry volumes climbed, the firm’s first-line responses across phone, email and chat could no longer keep pace. Supervisors were forever allocating tickets and fighting fires, while operators wore themselves thin caught between “replying quickly” and “replying correctly”. Sales complained that they “couldn’t read how customers were feeling”, development that “faults and feature requests arrived all jumbled together”, and customers that “the answer changes depending on who you get”.
Today the firm uses AI to classify enquiries, draft first responses, summarise case histories and lay the groundwork for VOC analysis, and has shifted to a design in which people concentrate on the harder conversations — churn concerns and specification queries among them. Over the past three months, reviewing operations across 1,240 email and chat enquiries, the classification work done before the first reply fell from an average of six minutes per case to two. That said, the figure reflects this firm’s particular channels, enquiry types and state of knowledge management, and it would be a mistake to assume every organisation will see the same result.
This article sets out, for the CS leaders, supervisors and operations designers keen to push through customer support reform, how to redesign the contact centre on the assumption of CS AI — covering channel design, first response, knowledge management, quality control and escalation design. The aim is to separate what you hand to automated-response AI from what people attend to, and to pursue greater support efficiency while holding the line on response quality.
A word of caution, though: dropping AI in does not, on its own, change anything on the ground. Only with quality control, training, knowledge updates and clear rules for cross-team coordination in place does it translate into repeatable improvement. If any of this rings true, do read on with your own CS operation in mind.
Why customer support work so readily reaches breaking point
Customer support sits closest to the customer, yet it is also work whose load is hard to see.
As enquiry volumes rise, the priority on the ground becomes, first and foremost, “clearing the backlog”. Answer the phone, reply to the email, react to the chat. Absorbed in the case in front of them, staff tend to put off reviewing response quality, updating the FAQ, tidying the knowledge base and considering how to stop problems recurring.
Let this carry on and three problems, chiefly, take hold.
- Inconsistency between staff
- To the same question, one member of staff explains the background carefully while another fires back something short and businesslike. Neither means any harm, but to the customer it can read as “the company’s answers aren’t consistent”.
- Over-reliance on supervisors and leaders
- Every time a slightly tricky enquiry lands, the floor asks “how should I word this?”. Supervisors get bogged down in individual judgements, leaving little time for coaching, quality improvement or knowledge upkeep.
- Enquiry data left untapped
- Customers’ frustrations, requests, stumbling points and signs of churn are all there in the day-to-day enquiries. Yet if the information is buried the moment a case is closed, none of it feeds back to sales, development or marketing.
What matters in customer support reform is not simply handling more cases. You need to think about three things at once: reducing enquiries, steadying the first response, and putting the customer’s voice to work in improving the business.
The point of adopting AI is not to cut headcount but to change roles
Hear the words “CS AI” or “automated-response AI” and it is sometimes taken to mean “automate enquiry handling, the lot of it”. In real customer support, though, a design that hands everything to AI is simply not realistic.
Every customer’s situation is different. The weight of the same enquiry shifts with the contract, the usage history, past exchanges, the emotional temperature and how much the account matters to the business.
So the first thing to settle when adopting AI is not “how much do we automate?” but “what is the work people genuinely need to attend to?”.
| Who handles it | Main tasks | Design considerations |
|---|---|---|
| Work readily handed to AI | Classifying enquiries, drafting first responses to common questions, searching past knowledge, summarising long enquiries, tidying case histories, preparing the groundwork for VOC analysis | Begin with areas that are straightforward to handle on the basis of set rules or existing knowledge. |
| Work people should own | Responses that attend to the customer’s feelings, judgements involving contracts or money, exceptions, defusing complaints, conversations with customers at risk of churning, and complex enquiries that need internal coordination | It calls for more than returning the right information — it demands the ability to keep the customer’s trust intact. |
A 2025 Gartner study likewise found that leaders across service and support feel the pressure to adopt AI and to spend more on it. At the same time, a BCG study notes that only a minority of firms are managing to draw measurable business value from generative AI.
In short, redesigning customer support around AI is not a matter of swapping people out for machines. It is about moving the preparation, the tidying and the first drafts to AI so that people can concentrate on judgement and conversation.
Start by rethinking your channel design
When you redesign customer support, the first thing to revisit is channel design.
Phone, email, chat, enquiry forms, FAQs, help pages. At many companies, customers can get in touch through several channels at once. But having lots of channels and being easy for the customer to use are two different things.
An urgent fault, for instance, may well need to be taken by phone. Checking a procedure or confirming a basic specification, on the other hand, can often be resolved through chat or the FAQ. Enquiries about contract changes or billing, which need identity and history checks, are sometimes better received through a form that gathers the necessary details up front.
What matters is to design channels around the customer’s situation and the difficulty of the response, rather than carving them up to suit the company.
The same goes for AI: there is no need to drop it into every channel in the same way. It is more realistic to start where enquiry volumes are high and the answer patterns are comparatively easy to standardise.
You might, for example, use AI for the first response in chat, classifying what the customer is asking. Where AI can answer, it drafts the first response and a member of staff checks it before replying. Anything needing judgement goes straight to a supervisor or specialist team.
By deciding AI’s role channel by channel in this way, you can pursue support efficiency without doing much harm to the customer experience.
In first responses, watch not just for speed but for consistency
In enquiry handling, the speed of the first reply gets a lot of attention. Not keeping the customer waiting does matter. But what truly counts in that first reply is not speed alone.
The first answer a customer receives shapes how much they trust the company. If it is vague, or the explanation differs from one member of staff to the next, the customer starts to wonder “is this company’s support up to it?”.
For an AI-assisted first response, you need to design at least the following three things.
- What may be answered
- Anything set out plainly in the FAQ or manual — procedural guidance, explanations of basic specifications — is an area where AI can readily draft a first response.
- What needs checking
- Contract terms, individual settings, billing and the impact of incidents are matters where, even if AI drafts the reply, human review should be mandatory.
- What must not be answered
- Legal judgements, decisions on compensation, unconfirmed specifications and special arrangements for individual customers should all be designed so that AI does not state them as fact.
Even when you use a tool that combines AI chat, summarisation and training-data management — our own Kanata, say — simply instructing it to “answer the enquiry” is not enough.You need clear rules: “if you don’t know, say you don’t know”, “check the underlying knowledge”, “send anything touching contracts or money to a member of staff”.
The purpose of the first response is not merely to reply quickly. It is to reach a state where, whoever handles it, the basic information the customer needs is all there.
AI use without knowledge management does not last
One thing readily overlooked when adopting CS AI is knowledge management.
If the information AI refers to is out of date, it may well answer on the basis of that stale information. An un-updated FAQ, an old product manual, untidy case histories — drop AI into that and you simply add to the checking burden on the floor.
If you are to redesign customer support around AI, you need to treat knowledge as something to be operated, not merely stored.
In practice, that means organising your common enquiries, answer templates, product specifications, incident procedures, escalation criteria and records of past judgements — and setting, for each, an owner responsible for updates and a review frequency.
| Type of knowledge | Example review timing |
|---|---|
| FAQ | Monthly |
| Knowledge tied to product specifications | With each release |
| Incident-handling procedures | After each incident |
AI’s accuracy is not settled by the model alone. Knowledge management that the floor can keep updating is precisely the foundation of CS AI use.
Use escalation design to lighten the load on people
One reason the floor wears itself out in enquiry handling is that “who to consult, and from what point on” is left vague.
Every time a difficult enquiry arrives, a member of staff judges it on their own, asks over chat, checks with a supervisor, then consults development or sales. When this plays out each and every time, response times stretch out and the customer’s answer is delayed too.
An AI-led escalation design starts by settling the axes along which enquiries are classified.
- Is the urgency high?
- Is the customer’s emotional temperature running high?
- Is there any impact on contracts or revenue?
- Is technical investigation needed?
- Has the same enquiry come up before?
- Are there signs it could lead to churn or a complaint?
AI can help with these classifications on the basis of the enquiry text. The final judgement, though, must rest with a person. Even just having that initial sorting can lower the checking burden on supervisors and staff.
In escalation design, you also make the destination clear.
| Enquiry type | Main point of referral |
|---|---|
| Specification queries | Development |
| Contract terms | Sales or customer success |
| Billing | Finance |
| Faults | Technical staff |
| Complaints | Supervisor |
More important still is feeding the outcome of an escalation back into the knowledge base. If a judgement, once made, is never recorded, the same checking happens all over again next time. By asking, after a case, “could this be handled at first line next time?”, “should it go into the FAQ?”, “should AI be allowed to refer to it?”, enquiry handling improves little by little.
Shift quality control from checking every case to focused review
With AI, you can produce first-response drafts, summaries and classification results in short order. But the more output there is, the harder it becomes for people to check every bit of it in detail.
That is where quality-control design comes in.
Conventional quality control typically had supervisors review a set number of case histories, checking the wording and the content of answers. Once AI is in use, that approach alone may not keep pace.
You need to narrow what you look at.
Among the answers AI drafts, for instance, you review those touching contracts, pricing, faults, churn or complaints with particular care. Meanwhile, the likes of procedural guidance and basic-specification checks — things set out plainly in the knowledge base — you handle as sample checks.
Quality control also needs to confirm not just “is it correct?” but “will it land with the customer?”. An AI answer can be grammatically tidy yet fall short of attending properly to the customer’s concern.
Including the following sorts of points among your check items makes the operation easier to run.
- Does it follow the underlying knowledge?
- Does it state as fact anything that should not be stated as fact?
- Is the explanation suited to the customer’s situation?
- Are there too many technical terms?
- Where needed, can it switch over to human review?
Automated responses and first-response drafts from AI do not do away with the need for quality control. If anything, by making clear where to focus your attention, the supervisor’s role becomes more important still. From an AI risk-management standpoint too, it is worth operating with a defined purpose of use, anticipated risks and a means of monitoring. For reference, see the NIST AI Risk Management Framework and the OECD AI Principles.
Use VOC analysis to make support a starting point for business improvement
Customer support is where the customer’s voice gathers.
This screen is hard to follow.
I trip up on the same setting every time.
It’s different from what I was told before I signed up.
With this feature, I’d find it easier to stay.
Such remarks are more than mere enquiries. They are valuable information that can feed into product improvement, a rethink of sales materials, better onboarding and revised marketing messages.
Yet caught up in the day-to-day, teams often never get round to VOC analysis. Enquiries are marked resolved and end up buried in the history.
Here, too, AI helps. It can summarise and classify a period’s worth of enquiries, sorting them into customer frustrations, requests, questions, churn signals, suspected faults and the like. Rather than having staff read back through each one, you can have AI do the groundwork and let people judge the important trends.
You might, for example, produce a monthly VOC report covering the following.
- The themes with the most enquiries
- Themes that have grown since the previous month
- Operations customers tend to stumble over
- Matters that suggest something went unexplained at the sales stage
- Faults and improvement requests to share with development
- Items to add to the FAQ or help pages
Organised like this, customer support shifts from “a function that processes enquiries” to “a function that returns customer understanding to the business”.
Where AI-led CS redesign tends to come unstuck
Adopt AI and the hoped-for results may still fail to materialise. More often than not, the cause lies not in the AI itself but in the design and the way it is run.
- Adopting AI while the knowledge base is still in disarray
- An out-of-date FAQ, manuals no one can locate, answer templates that differ from one member of staff to the next. In that state, AI’s output will not be stable either.
- Not deciding what to leave to AI
- Try to have AI answer absolutely everything and misanswers and overconfident assertions become more likely. You need to make clear what may be answered, what needs checking, and what must not be answered.
- Ignoring the review burden on the floor
- An operation in which people must check every answer AI produces can, on the contrary, mean more work. Separating focused review from sample review is key.
- Losing sight of the customer experience
- Prioritise the company’s processing efficiency alone and the customer may feel they are being handled mechanically. Even where AI is in play, you need to design the conditions for switching to a person and the way that is communicated to the customer.
- Not designing for cross-team coordination
- There is a limit to what CS can improve on its own. Without working alongside sales, development, marketing, IT and leadership, the root causes of enquiries simply persist.
An AI-led contact-centre redesign is a rethink of how the work is organised, not merely a tool rollout. Get this wrong and AI, far from helping the floor, becomes something that creates fresh checking work.
Start small with implementation and widen it as you go
You do not have to change everything about customer support in one go. If anything, starting small makes it easier to bed in.
-
Take stock of your enquiries
Gather the past one to three months of enquiries and classify them by theme. Find the ones that come in high volume, the ones whose answers are easy to standardise, and the ones where answers vary by member of staff.
-
Decide what to hand to AI
Candidates include enquiry classification, drafting first responses, pulling out FAQ candidates and summarising case histories. Rather than auto-replying directly to customers from the outset, starting with internal drafts and classification makes the risk easier to contain.
-
Run a trial
Settle the target channels, target enquiries, the staff doing the checking, the review method and the evaluation metrics. Among those metrics, include response time, time to first reply, escalation rate, customer satisfaction and the checking burden on staff.
-
Update the knowledge base
Enquiries AI could not answer well may point to gaps in the FAQ or manuals. Conversely, enquiries it answered reliably become candidates for standardising the first response.
-
Widen the scope
From chat to email, from first response to VOC analysis, from within CS to coordination with sales and development. Widening it in stages makes it easier to bed AI use in while keeping disruption on the floor to a minimum.
In choosing tools, look at how easily they support work design and knowledge management
When choosing CS AI or automated-response AI, it is best not to judge on answer accuracy alone. In real operation, what matters is integration with your existing enquiry-management tools and CRM, permission management, log inspection, the ease of updating knowledge, and a UI the floor can keep using.
A company with a well-kept FAQ and help pages already, for instance, is suited to an AI chat or retrieval-augmented arrangement that can refer to them readily. A company with many enquiry histories that is spending a great deal of time on analysis will benefit from a tool strong on summarisation and classification. Where you want to set things up right through to internal training and onboarding, the reuse of learning content and prompts becomes a selection criterion too.
A service such as Kanata, which lets you combine AI chat, AI summarisation and training-data management, is one option where you want to push ahead with enquiry classification and first-response drafting while referring to your internal knowledge. Whichever tool you use, though, it is important to settle, before adoption, “which work are we improving?”, “which information do we let it refer to?” and “who reviews it?”.
A tool will not do the thinking about work design for you. Only with the way work is divided, the way knowledge is kept and the mechanism for quality control in place can you put a tool’s effect to the test.
In summary: AI-led customer support is a redesign meant to raise the value of human response
Redesigning customer support around AI is not an exercise in mechanically automating enquiry handling.
It is about revisiting the whole flow: classifying enquiries, making first responses consistent, updating the knowledge base, sorting out escalation, and returning VOC to business improvement.
Work that can be left to AI, you leave to AI. Work people should check, people check. Customers people should attend to, people attend to. Only once that line is drawn can you pursue support efficiency while holding response quality steady.
The customer support floor is where a company’s understanding of its customers gathers most densely. By drawing on AI, you can keep that voice from being buried and turn it towards a better customer experience and business improvement.
AI is no cure-all, mind. Leave old knowledge, vague operating rules and divisions between teams as they are, and results will not hold steady. What counts is to take the adoption of AI as the occasion to rethink the business processes themselves.
Q&A
When bringing AI into customer support, where should we start?
It is more realistic to start not where you auto-reply directly to customers, but with internal classification, summarisation and the drafting of first responses. Take stock of the past one to three months of enquiries and pick the areas that come in high volume, are easy to standardise and tend to vary by member of staff — they are easier to test.
What is the difference between enquiries you can leave to automated-response AI and those people should handle?
Procedural guidance, explanations of basic specifications and common questions set out plainly in the FAQ or manual are areas AI can readily support. Contract terms, billing, the impact of faults, compensation decisions, complaints and churn concerns, on the other hand, should presuppose human checking or judgement. What matters is to define in advance what may be answered, what needs checking and what must not be answered.
Once we adopt AI, does knowledge management become unnecessary?
It does not. If anything, AI use raises the importance of keeping the knowledge base in order. If the FAQ or manuals are left out of date, AI may answer on the basis of stale information too. It is important to settle update timing and owners — the FAQ monthly, product specifications with each release, incident-handling procedures after each incident, and so on.
In quality control after adopting AI, what should we check?
Check whether AI’s output follows the knowledge base, whether it states as fact anything it should not, whether the explanation suits the customer’s situation, whether there are too many technical terms, and whether it switches to a person where it should. Rather than checking every case, a design that puts high-impact areas — contracts, pricing, faults, churn, complaints — under focused review is effective.
By what metrics should we measure the effect of CS AI?
You need to look at a combination: time to first reply, the time spent classifying enquiries, time to resolution, escalation rate, the proportion resolved by the FAQ, customer satisfaction and the checking burden on staff. Watch processing time alone and you risk overlooking a decline in the customer experience or response quality. It is important to confirm both efficiency and quality.
