AI in Customer Service Goes Beyond Chatbots (and What Works Instead)

When someone says “AI in customer service”, most people picture a chatbot: the widget in the corner of a website that answers FAQs. Chatbots have their place, but in our SAP Service Cloud V2 projects they are rarely where AI changes the numbers. The bigger effect comes from the features that help the service agent handle the difficult cases faster.

This post covers which of those features work today, which are overpromised in demos, and in which order to introduce them.

TL;DR: The most useful AI in customer service is agent-facing: automatic case classification, case summaries, reply drafts and knowledge suggestions. Start with classification (lowest risk), add agent assistance (fastest adoption), then work on routing (biggest structural change). Customer-facing automation such as SAP’s Digital Service Agent pays off for simple, predictable requests, not for complex cases.

Why do chatbots underdeliver?

Most chatbots handle simple requests well enough: order status, password resets, opening hours. Anything that needs judgement or context goes to a person. So companies run a chatbot project, measure the deflection rate and call it AI. Meanwhile the agents handling the complex cases, the ones that decide customer satisfaction and retention, get no support at all.

That is the wrong way round. The complex cases are where time, cost and customer loyalty are decided, and that is where AI support pays off first.

Where does AI move service metrics?

These are the capabilities that deliver results in SAP Service Cloud V2. Which of them you can use depends on your edition and licence, since several come with SAP’s CX AI Toolkit.

Automatic case classification

Every incoming case needs a category, a priority and often a product. Doing that by hand costs agents time on every ticket. SAP’s case categorisation reads the case content and proposes these fields. It learns from your historical data: the more consistently cases were classified in the past, the better the proposals. Accuracy therefore varies from customer to customer, so measure it in a pilot rather than trusting a figure from a slide.

Case summaries

Long cases with many emails and calls are hard to pick up. The case summary condenses the interactions so an agent who takes over, or a team lead who escalates, understands the situation without scrolling through the whole history.

Reply drafts

When an agent opens a case, the system drafts a reply based on the case content and context. The agent edits and sends. Starting from a draft instead of a blank reply is a clear time saver, and in our experience it is the feature agents adopt fastest: sceptics usually come round once they have used it for a few days.

Knowledge suggestions

Relevant knowledge articles appear next to the case, matched to its content, so agents do not have to search. This helps new agents most, when they handle issues they have not seen before. The reverse direction also exists: generating a draft knowledge article from a resolved case. It keeps the knowledge base from going stale, provided someone reviews the drafts.

Sentiment detection

Not every urgent case looks urgent. Sentiment detection flags negative signals in incoming messages so an upset customer can be prioritised early. Treat it as a coarse signal: a clear positive or negative is reliable, irony and subtle frustration are not.

Routing

Traditional routing is rule-based: product category to team A, region to team B. It struggles with edge cases, unbalanced queues and complex issues landing with junior agents. Combining AI classification with well-maintained skill profiles and routing rules sends cases to the right team from the start and reduces reassignments. It is also the capability that needs the most preparation.

What is overpromised in demos?

Works well today: case classification (with consistent historical data), case summaries, reply drafts, knowledge suggestions, coarse sentiment flags.

Works, but needs investment: routing (requires careful configuration and good skill profiles), knowledge article generation (needs a review process to keep quality up).

Overpromised: nuanced sentiment analysis, fully automated resolution without human involvement for anything but simple, predictable requests, and real-time analysis of phone calls in most environments. If a vendor shows you these in a demo and says “it just works”, ask for production references.

For customer-facing automation, SAP’s Digital Service Agent has been generally available since the Q4 2025 CX release. It answers routine questions and escalates to a person when needed. We look at what it does and where it stops in our post on the Digital Service Agent.

In which order should you introduce AI?

If you run SAP Service Cloud V2 and want AI to deliver value, the order matters. Do not switch everything on at once.

Weeks 1 to 2: audit your data. Check how consistently cases are classified, how complete the knowledge base is and how good the routing rules are. AI amplifies what is there, good or bad. This is usually where teams discover that their data is not in the shape they assumed.

Month 1: start with classification. Lowest risk, quickest payoff. Switch it on, monitor accuracy and let agents correct mistakes, which improves the proposals.

Months 2 to 3: add agent assistance. Enable summaries, reply drafts and knowledge suggestions. Train agents to treat drafts as a starting point, not a finished answer.

Month 4 onwards: tackle routing. It needs the most setup but brings the biggest structural improvement. Map agent skills, define criteria and pilot with one team before rolling out.

Many service cases also depend on ERP data: the order, the delivery, the material, the warranty. If Service Cloud V2 and S/4HANA are not connected, even the best AI proposal stops at that seam. We describe the connection in our post on integrating Service Cloud V2 with S/4HANA Public Cloud. For the wider picture of SAP’s CX AI features, see our CX AI Toolkit reality check. And if you want to go beyond the standard features, for example an AI agent that reads, classifies and updates cases in your own tenant, our MCP Server for SAP Sales & Service Cloud V2 exposes the V2 APIs as tools for any MCP-compatible AI.

FAQ

Which AI feature should I implement first?

Automatic case classification. It carries the lowest risk, pays off fastest and needs the least configuration. Switch it on, monitor accuracy and let agents correct exceptions; the proposals improve from those corrections.

How accurate is AI case classification?

It depends mainly on how consistently your past cases were classified. With clean, consistent history the proposals are good enough to save real triage time; with messy history the model learns messy patterns. Measure accuracy in a pilot on your own data before you rely on it.

Do AI features replace service agents?

No. The most useful features help agents work faster: classification shortens triage, summaries and reply drafts shorten handling, knowledge suggestions shorten searching. The Digital Service Agent handles only simple, predictable requests on its own. Anything that needs judgement still needs people.

What data do I need for AI features to work?

Classification needs a solid history of consistently classified cases; as a rule of thumb we look for at least a thousand. Knowledge suggestions need a maintained knowledge base. Routing needs defined agent skill profiles. Summaries and reply drafts work with the case content itself.

How do AI reply drafts affect quality?

Drafts are a starting point, not a final answer. Agents review, adjust tone, add specifics and handle exceptions before sending. Quality usually improves, because agents start from the relevant context instead of a blank reply, as long as nobody sends drafts unread.