A customer writes at ten in the evening to the company messaging number: the system installed the year before has stopped working. The operator who covers that number checks the phone at half past eight the next morning and finds eleven messages, three of them about the same request. She pieces it together, calls back and apologises for the wait. Almost every project on AI agents in customer service starts from scenes like this.
The question that keeps coming up during evaluation, however, is about volume: how many requests can be automated. It matters less than it seems. What decides the outcome are the edge cases, that is, what happens when the automation does not know the answer. A customer who gets a vague reply and cannot find a way to reach a person ends up more dissatisfied than one who waited twenty minutes on the line.
Three different technologies sold as AI agents in customer service
Anyone researching AI agents in customer service today meets three very different offers under the same name. Telling them apart matters, because they cover different needs and involve different commitments.
Traditional automation follows predefined paths: the phone tree, or the chatbot that answers only if the question arrives in the expected form. It routes reliably and stiffens on everything else.
Generative artificial intelligence understands context and writes adaptive answers. It summarises, translates and prepares drafts. The limit shows in the first week of use: it answers well, then an operator has to copy the information gathered into the company systems by hand. It moves the work instead of finishing it.
Agentic artificial intelligence adds the ability to act. An agent works towards a goal, coordinates several systems, carries out operations in sequence, handles exceptions and remembers context from one interaction to the next. This is where the jump in value lies, and the reason is practical: without a connection to the systems holding customer records, cases and deadlines, an agent produces text rather than completed work.
Which case to start an AI agents in customer service project with
The first case to hand to an agent is the one with high volume, recurring questions and answers already documented in the company. The most common mistake is starting from the most ambitious case instead of the most measurable one.
Good candidates are filtering support requests before a ticket is opened, covering messaging channels, answering questions on procedures and contract terms, giving the status of a case, and booking an appointment or a site visit. In areas this well defined an agent shows its value within a few weeks. A project that tries to cover the whole customer service from day one takes months and delivers theoretical results.
| Type of request | Who handles it |
|---|---|
| Status of a case or an order | The agent, reading the data from the ERP |
| Sending documents already issued | The agent, retrieving them automatically from the archive |
| Booking appointments and site visits | The agent, confirming on the shared calendar |
| Fault report | The agent collects the details, the operator decides on the intervention |
| Complaints and cases with a long history | The operator, with the conversation already reconstructed by the agent |
To widen the scope, the question to ask is whether one condition holds: when the agent cannot answer, the hand-off to a person is smooth and tracked. If it is, the project can grow; if it is not, extending it makes the service worse.
A thousand reports a month on a single number: what emerged
A company that installs and maintains systems at end customers’ premises brought us a situation that describes the problem well. All communication went through a single messaging number covered by a few people: quotes, reminders, installation dates, accounting documents and fault reports. Every month around four hundred documents were retrieved from the ERP and forwarded by hand, one by one. Support reports ran at about a thousand a month and arrived at any hour.
The first release did not cover technical support, even though it was the most visible part. It covered routing requests and sending documents, the most repetitive activity and the least interesting one for a skilled person. On support the agent stops earlier: it collects the details of the fault, retrieves that customer’s history from the CRM and prepares the report for the manager, who in the morning finds a request already prepared instead of twenty messages to read.
One configuration choice separates a well-designed project from an improvised one: if no operator is available, the conversation returns to the agent, which says so and offers a call-back. The customer is not left in front of a silent chat. The point deserves more attention than it gets in sales demos. In the evening, when nobody covers the channels, the agent keeps collecting reports complete with details and priority, so whoever starts the morning shift finds an orderly work queue instead of a list of messages to interpret. Over time the difference shows in loyalty: customers who get the call-back at the agreed time tend to stay, while those left hanging tell others about the wait.
Five questions to ask anyone selling AI agents in customer service
These answers say more than any demo and apply to any vendor of AI agents in customer service.
- How the hand-off to a person works, and what the customer sees at that moment.
- What happens if no operator is available: the conversation ends or returns to the agent.
- Which company systems the agent can read, which it can write to, and who builds the integration.
- Where conversation data is stored, for how long, and who can access it.
- Who changes the flows after release: the vendor on request, or an internal person after a few hours of training.
The last one weighs more than the others in the months that follow. An agent that nobody in the company can correct ages quickly, because price lists change, procedures change and customers’ questions follow.
Where the data lives and who answers for errors
In every first call about AI agents in customer service the question comes up of where the information handled by the system is stored. The concern is well founded and deserves verifiable answers rather than reassurance. Three points should be put in writing: the physical location of the data, whether it is used to train third-party models, and the list of who can access it. In sectors with sensitive information, such as healthcare and finance, the assessment also involves whoever is responsible for privacy in the company, and no technology vendor can replace that opinion.
Our position comes from a concrete, verifiable infrastructure choice. Aesir Srl hosts the language models on its own cloud platform, in its own Tier IV data centres in the European Union with replication, within a perimeter certified ISO 9001, 27001, 27017 and 27018. The practical consequence matters more than the certification: we state in the contract where the data is and who processes it, without referring to the terms of use of a third-party service. Whoever runs the infrastructure also covers security, because MSP and MSSP services fall within the same contract.
In the European regulatory framework the reference text remains the AI Act published on EUR-Lex, which classifies systems by risk level and sets transparency obligations towards people who interact with an automated system.
Four indicators that tell you whether AI agents in customer service work
The share of conversations handled automatically is the measure companies ask for first and the one that tells them least. Outcomes are what count, and four indicators are enough.
- Requests closed without human intervention and without the customer writing again in the following days.
- The time between the request and the first useful answer, not just any first answer.
- Conversations closed with no outcome, that is, with no solution and no hand-off to a person.
- Operators’ workload on requests that need expertise, measured before and after release.
The third weighs more than the other three and almost no company tracks it. A well-designed project provides a dashboard to read these figures; without that visibility, problems surface weeks late, usually through a review. For the same reason, in the projects we run we plan a period of observation and fine-tuning after the first release: the first weeks of real conversations show where to adjust the configuration.
AI agents in customer service deliver the most value when they reach the company systems and the channels customers already use. On the voice channel the topic links to the cloud phone system, because a badly routed call undoes any automation upstream.
Frequently asked questions
Do company systems need to be integrated?
An agent that answers general questions works without integrations. To report the status of a case or book an appointment, however, it must reach the ERP or the CRM: the technical check of these integrations is the first activity of the project.
Which channels can an agent work on?
Voice, website chat, SMS, messaging and social channels cover most scenarios. The scope depends on the channels customers use, not on those technically available.
Where is the data processed?
In projects we host on our platform, data stays in Tier IV data centres in the European Union. When a project relies on third-party services, the terms must be checked one by one and put in writing.
How long does the first release take?
A well-defined, well-documented case reaches its first release within a few weeks. Integrations with the ERP and CRM are the variable that shifts the timing, and they are assessed before starting.
Let’s talk
When assessing a project of this kind, the starting point lies in two things the company already knows: which three requests customer service receives most often, and where the answers to them are documented today. When those answers exist, the project has a solid base. If they do not, writing them becomes the first phase of the project: the material gathered feeds the agent’s knowledge base from the first release and stays with the company as an organised asset, useful on every contact channel.
If you would like to explore the topic or assess your company’s situation, fill in the form at the bottom of the page or write to support@aesir-tech.it: we will book a free consultation and start from your numbers.