What AI can and cannot do in support
Support is one of the more dependable places to apply AI, because much of the work is answering questions that have been answered before. A model given your help content can reply at once, at any hour, in the customer's language. It does not tire of the hundredth question about delivery times.
It also has firm limits. It knows only what is written down and what your systems tell it. It cannot make an exception, calm an upset customer the way a good agent can, or notice that a problem is new. Implementations that fail tend to share one cause: the AI was placed between customers and people in order to reduce contact, instead of being used to resolve the simple requests so that people have time for the rest.
Choose which requests to cover
Export a few months of tickets or chats and sort them by reason for contact. For each reason, note the volume and what an agent does to resolve it.
| Kind of request | Example | Fit for AI |
|---|---|---|
| Information that is the same for everyone | Opening hours, delivery areas, how to return an item | Strong. Answerable from help content |
| Information about the customer's own case | Where is my order, when is my appointment | Strong, once identity is verified and the system is connected |
| Guided procedures | Resetting a device, setting up an account | Good, if the steps are documented and current |
| Simple account actions | Change an address, reschedule, cancel within policy | Possible later, with rules and confirmation |
| Complaints and disputes | A damaged item, a billing error, a poor experience | Hand over. AI can gather details first |
| Exceptions and judgement | A refund outside policy, a vulnerable customer | A person, always |
Begin with two or three reasons from the top rows that together make up a meaningful share of contacts. Narrow scope done well earns trust, and the list can grow every month.
Get the knowledge right first
The AI will repeat whatever your help content says, including what is out of date. Before launch, review the articles behind the chosen requests with the agents who handle them. They know which articles are wrong, which are missing and what customers really ask.
- One answer per question. Remove or merge articles that contradict each other, and retire old versions.
- Policies with their conditions. Refund periods, exclusions and regional differences stated explicitly, because the AI will be asked about the edge.
- Agents' unwritten knowledge. The saved replies and personal notes agents rely on belong in the shared content.
- An owner and a date. Someone responsible for each article, and a routine for updating content when a policy or product changes.
- Things it must not answer. Legal, medical or financial advice, speculation about faults, promises about compensation. List them.
Our guide to preparing company documents for a knowledge assistant goes through this work in detail.
Design the handoff to a person
- 01
Make the route obvious
A customer who asks for a person, in any wording, gets one. Hiding the route produces angry customers and worse reviews than having no AI at all.
- 02
Hand over on signals, too
Repeated rephrasing, strong negative language, a topic on the handover list or the AI's own uncertainty should all trigger an offer of a person.
- 03
Pass the whole conversation
The agent sees what was asked, what was answered and any details collected, with a short summary. The customer must not be asked to start again.
- 04
Be honest about waiting
Outside staffed hours, say so, take the details and promise a reply by a stated time through a channel the customer chooses.
- 05
Tell the customer what they are talking to
State at the start that this is an automated assistant. People are more forgiving of a machine that says it is one, and some jurisdictions require the disclosure.
Identity, account data and actions
The most valuable answers are about the customer's own order, booking or bill. Those require two things general answers do not: knowing who the customer is, and a connection to the system that holds the facts.
- 01
How is the customer identified?
What to decideSigned in to their account, a one-time code to their registered phone or email, or details matched against the order. Decide what each level permits.
- 02
What may be shown at each level?
What to decideAn order status may need little. Address, payment or personal details need proper verification.
- 03
Which systems are connected, and read or write?
What to decideStart with reading. Our guide on connecting AI to your CRM and ERP covers the design.
- 04
Which actions may the AI take?
What to decideEach one with its rule, such as cancel only before dispatch, and a confirmation step shown to the customer.
- 05
What is logged?
What to decideEvery lookup and action, linked to the conversation, for disputes and audits.
A worked example
A furniture retailer brings AI into support
This example is invented to show a staged rollout. A furniture retailer's support team of nine is overwhelmed each time a sale ends.
- RequestsThree months of contacts are sorted. Delivery status and delivery scheduling make up almost half. Assembly questions are next. Damage claims are a tenth, and are marked for handover.
- KnowledgeAgents review 64 help articles. 11 are outdated and 7 topics have no article. The delivery policy for remote areas existed only in a team leader's notes.
- Stage oneFor a month, the AI drafts replies inside the agents' own tool. Agents edit and send. Their edits show where the content is still weak.
- Stage twoThe assistant answers customers directly on delivery and assembly questions, on the website chat only. Order status is read from the order system after a one-time code. Anything else is handed over with the conversation.
- MeasuresThe team tracks whether the customer contacts again within a week about the same issue, satisfaction after each conversation and how often a person is requested.
Agents spend their time on damage claims and scheduling problems, where they are needed. The repeat-contact measure exposes one weak answer, about deliveries to apartment buildings, which is rewritten in a day.
Roll out in stages and measure what matters
Measures that reflect service
- Issue resolved without a second contact
- Customer satisfaction after AI conversations, compared with agents
- Share of handovers that arrive with full context
- Wrong or unsupported answers found in weekly sampling
- Time agents now spend on complex cases
Measures that mislead
- Deflection rate taken alone
- Conversations ended, counted as resolved
- Average handling time as the main target
- Satisfaction surveys shown only after successful chats
- Cost per contact with no quality measure beside it
Add channels and topics one at a time: website chat before messaging apps, email before voice. After each addition, read a sample of real conversations every week. Nothing replaces a team leader reading what the assistant actually said to customers.
The implementation checklist
AI customer support implementation checklist
0 of 11 checkedWhat to do next
Sort last quarter's contacts by reason and pick the two largest groups that an agent resolves by pasting the same answer. Then ask your agents which help articles they do not trust. Those two lists are the start of the project.
Our catalogue of AI agent solutions shows support agents for specific industries, and our AI integration services page explains how we connect them to your systems. Tell us your top contact reasons.
Common questions
Will AI replace our support agents?
In most teams it changes what agents do more than how many are needed. Routine questions are answered automatically, and agents handle complaints, exceptions and complex cases, which take longer and matter more. Plan for that shift in skills and workload. Teams that cut staff on the strength of a deflection figure often find quality and satisfaction fall.
Should customers be told they are talking to an AI?
Yes. It is honest, it sets expectations and in some jurisdictions disclosure is required. Customers tolerate limits from a system that identifies itself and offers a person when needed. Pretending to be human damages trust the moment it is discovered, and it is usually discovered quickly.
How do we stop the AI from giving wrong answers?
Restrict it to answering from your reviewed content and connected systems, have it decline when the content does not cover the question, limit the topics it handles, and test it against real questions before launch. After launch, sample conversations weekly and correct the content behind any weak answer. Wrong answers usually trace to wrong or missing documentation.
Which channel should we start with?
Start where conversations are text, volumes are meaningful and your team already works, which for many businesses is website chat or the help centre. Email suits drafted replies reviewed by agents. Messaging apps come next. Voice is the most demanding channel and is best added after the content and handover have been proven in text.
How long before we see results?
Agent-assist tools can help as soon as the content is in reasonable shape, since agents remain in control. Direct answers to customers should follow once testing and a period of assisted use show the answers are sound. The pace is set mainly by the state of your help content and by how quickly it can be corrected.
