Contents
In customer service, AI works well when it assists agents: sorting requests, summaries, approved draft replies, searching the knowledge base. A customer-facing chatbot is only useful within a narrow scope, with a quick handoff to a human. A first use case starts at $2,300 (€2,000)excl. VAT.
AI in customer service generates as many expectations as bad experiences: everyone has run into a chatbot that understands nothing and keeps you from reaching a human. Yet, used in the right place, AI genuinely helps teams. The difference comes down to one simple choice: support agents rather than replace them, and limit what the machine is allowed to do on its own. Here is what works, what disappoints, and how to get started.
What works: AI that supports agents
The most reliable use cases are the ones where an agent stays in control and checks the result in seconds.
- Sorting and summarizing incoming requests. Category, urgency, summary, routing to the right team. This is often the best starting point; the page on sorting customer requests with AI goes into detail.
- Preparing a draft reply. Based on the request, the customer’s history and your standard answers, AI suggests a text that the agent edits and sends. It never goes out on its own when it commits the company.
- Finding the right information. An assistant connected to your procedures, terms and product sheets answers the agent and cites its sources. See an AI assistant connected to company documents.
- Summarizing a long history before picking up a case: previous exchanges, orders, past complaints.
- Analyzing contact reasons over a month: which topics keep coming back, which are growing, and what could be fixed upstream (a help page, a delivery tracking email, a product fix).
What disappoints: the unbounded chatbot
Disappointment almost always comes from a chatbot let loose on customers with no guardrails:
- it improvises a plausible but wrong answer about a lead time, a price or a return policy;
- it promises a goodwill gesture nobody authorized;
- it blocks access to a human and turns a simple question into a complaint;
- it does not have access to the right data (order status, customer contract) and answers beside the point.
A chatbot can be useful, but within a narrow, controlled scope: frequent questions with stable answers, order tracking from an order number, booking appointments. Always with a visible way out to a human.
Use cases by level of risk
| Use case | Facing | Risk | Recommendation |
|---|---|---|---|
| Sorting and summarizing requests | Agents | Low | Ideal starting point |
| Document search assistant | Agents | Low to medium | Sources always cited |
| Draft reply | Agents, then customer | Medium | Approved before sending |
| Contact reason analysis | Management | Low | Monthly review |
| FAQ chatbot with limited scope | Customers | Medium | Controlled answers, handoff to a human |
| Open chatbot on any topic | Customers | High | Avoid |
What it costs
These amounts exclude VAT and come from my pricing.
| Situation | What is delivered | Indicative budget | Timeline |
|---|---|---|---|
| First use case on your current tools | Sorting and summaries, or draft replies, or an agent assistant | From $2,300 (€2,000)excl. VAT | 2 to 8 weeks |
| Built into a custom client portal | Structured requests, tracking, AI, dashboard, notifications | $6,800 to $22,400 (€6,000–€20,000)excl. VAT | 1 to 3 months |
In the estimator, built-in AI counts for $2,800 (€2,500)excl. VAT, a client portal for $2,800 (€2,500)excl. VAT and notifications for $1,200 (€1,000)excl. VAT, on top of the application’s core. AI API usage comes on top, billed by the provider. I do not put a number on reduced volume or response times upfront: the metrics are measured before and after, on your own requests.
The rules I set
- Human approval for anything that commits the company: refunds, goodwill gestures, lead times.
- Transparency: customers know when they are dealing with an AI. The EU Artificial Intelligence Act (Regulation (EU) 2024/1689) sets out transparency obligations for certain systems; your lawyer can tell you what applies in your case.
- Data limited to what is necessary, with a provider whose business terms exclude training on your data; see what data to share with AI.
- Continuous measurement: corrections to drafts, handoffs to humans, uncovered topics.
When AI is not the answer
- Volume is low and the team knows every customer: a direct relationship is worth more than automation.
- Requests stem from a recurring problem (late deliveries, a poorly explained product): fixing the cause reduces contacts more reliably than AI.
- Your support tool already offers these features and they suit how you work.
- Your standard answers and procedures do not exist yet: write them first, because AI cannot rely on what is not documented.
Where to start
- Export a month of requests and sort them roughly: you will see where the team’s time goes.
- Pick a first agent-facing use case, usually sorting or draft replies.
- Run a one-week audit and a prototype tested by your agents on real requests.
If you are not sure which use case to prioritize, the page on where to start with AI helps you choose. The full approach is on the AI integration page.
Frequently asked questions
Can a chatbot replace part of the team?
I never frame it that way. A well-scoped chatbot handles some of the simple, repetitive questions, which frees up time for complex requests. But it does not settle disputes, handle special cases or manage relationships with key accounts.
Should customers be told they are talking to an AI?
Yes, I always recommend it: it is a matter of trust, and the EU’s AI regulation sets out transparency obligations for certain systems. Have your lawyer confirm exactly what applies in your case.
What happens when the chatbot does not know the answer?
It must say so and immediately offer another channel: a form, a callback, or a transfer to an agent along with a summary of the conversation. A chatbot that goes in circles or improvises an answer does more damage than having no chatbot at all.
How do we measure whether AI improves customer service?
With metrics chosen before rollout: time to first response, share of requests resolved without follow-up, rate of handoffs to a human, reported satisfaction, and corrections made to drafts. Compare before and after, over a long enough period.