An AI customer service agent understands questions in free language, answers only from sources the organization has approved, identifies the customer before revealing any personal data, and takes real actions: it opens a case, checks a status and updates a record in the CRM or ERP. When it has no source, or when a human judgment call is needed, it says so and hands over to an agent with the full context. The difference from a scripted chatbot: the agent is not limited to questions someone wrote in advance, and it moves the request forward instead of pointing to a form.
What is the difference between an AI customer service agent and a chatbot?
A scripted chatbot follows a decision tree: buttons, keywords and pre-written answers. It works well for expected questions and gets stuck when the customer asks differently. An AI agent understands the intent behind a free-form question, looks for the answer in approved sources, and can use tools: retrieve data, open a case or bring a person into the conversation.
Put simply: a chatbot answers. An agent understands, accesses information, uses tools and moves the process forward, inside the channel where the customer already is.
| Scripted chatbot | AI service agent | |
|---|---|---|
| Source of answers | Answers written in advance for each scenario | Approved sources: website, documents, knowledge base and system data |
| An unplanned question | Goes back to the menu or replies "I didn't understand" | Understands the intent, and if there is no source, says so and offers an agent |
| Actions in systems | Usually a link to a form | Opens a case, checks a status and updates a record, within permissions |
| Customer identification | Usually none, or a separate form | Identification inside the conversation, before any personal data |
| Handover to an agent | The customer starts over | The agent receives the conversation with its context |
| Maintenance | Manual update of every branch in the tree | Updating the sources, and regular testing on a set of real questions |
What can an AI customer service agent do?
1. Answers from approved sources
The agent does not answer from the model's general knowledge. It searches the organization's website, documents and knowledge base, and writes an answer only from what it found. This method is called RAG, and there is a guide: what is RAG. In KESHER, for example, an answer without a citation from a source, or with weak grounding, is blocked before it is sent.
2. Identifies the customer
General information is open to everyone. Personal information, such as an account, an order or a case status, is shown only after identification, for example with a one-time code checked against the CRM. In Ashtrom Group's WhatsApp agent, the agent identifies the customer and the order from the conversation and retrieves the data from Priority ERP.
3. Opens cases and tracks them
A resident writes "there's a broken streetlight near my house", and the agent turns it into a classified case with an address, shows it for confirmation and routes it to the contact center. Later, the resident can ask for the status in the same channel.
4. Acts in core systems
The agent connects to the CRM and ERP through a controlled layer: Dynamics 365, Salesforce, SAP or Priority. Read actions run on their own. Write actions require approval from the customer or an employee before they run.
5. Hands over to a human agent with context
At any point, the customer can ask for an agent. The conversation is passed on with what has already been said, who the customer is and what was checked, so nobody starts over. In AICare at Israel Aerospace Industries (IAI), the bot and the agent work in the same conversation, with translation for the agent when the customer writes in another language.
6. Helps the agent too
Part of the value is on the contact center side. At Allot, a Copilot inside the Salesforce Case screen shows the agent similar cases and solutions that worked, and fully inherits Salesforce permissions.
Which channels does a service agent work in?
The rule: one brain, many channels. The knowledge, identification and permissions are the same in every channel. Only the form of the answer changes.
- Website: chat with cards, forms and buttons built according to the answer (Generative UI).
- WhatsApp: the same capabilities in short lines and buttons, including voice messages that are transcribed. For connecting WhatsApp to your customer system, there is a guide to WhatsApp and CRM integration.
- Voice and phone: voice conversation on the website, and a phone bot for inbound and outbound calls. The details are in the guide to voice AI agents.
- Microsoft Teams: for internal employee service, such as an IT or HR help desk. An agent built in Copilot Studio publishes directly to Teams.
- Inside the agent's system: a Copilot in the CRM screen, as at Allot.
How do you make sure the agent does not make things up or expose data?
- Answers only from sources. The agent cites the source, and an answer without grounding is not sent.
- An honest answer when there is no source. "I don't have authoritative information on that" is better than a guess. The agent offers a human agent, a callback or another channel.
- Identification before personal data. Without identification, the agent answers only general questions.
- Permissions. The agent sees and does only what the user in front of it is allowed to, according to the permissions of the source system.
- Approval before action. A case, update or cancellation is shown to the customer for confirmation before it is sent.
- Transparency. The agent introduces itself as an automated system and says when the conversation is recorded.
An agent that touches personal data needs a design that complies with Amendment 13 to Israel's Privacy Protection Law: where the data is stored, what is sent to the model and who sees what.
How do you measure whether a service agent works?
Do not promise a number before you measure. Set a baseline from the existing contact center, and measure the same metrics after go-live:
- Containment rate. How many conversations ended without a human agent. Check that they ended in a resolution, not in a customer who gave up.
- Handover rate, and the reasons. A handover is not a failure. What matters is why: a missing source, a missing action, or a customer who preferred a person.
- Customer satisfaction (CSAT). A short question at the end of the conversation, compared with conversations handled by agents.
- Unanswered questions. The questions the agent could not answer are a task list for the content team. In KESHER this is a standing improvement loop.
- Answer quality. A fixed set of real questions with correct answers, run after every change to the sources or the model.
- On the agent side. When there is a Copilot for agents, measure how many suggested replies were sent without editing.
How do you roll out an AI customer service agent?
- Pick one domain. A topic with many repeat requests and well-organized sources, such as order status or service information.
- Collect real questions. From the contact center, email and existing chat, including awkward wording and questions that must not be answered.
- Sort out the sources. What is approved, what is outdated and who owns the updates. An agent is only as good as its sources.
- Run a POC on the real questions and sources. Test answers, identification and handover before adding channels.
- Connect one action. For example, opening a case, with approval before it runs.
- Pilot in one channel, then expand. Agents monitor the conversations, and you improve based on unanswered questions and handover reasons.
For the cost of development and ongoing operation, see how much an AI agent costs an organization.
Examples from our projects
KESHER · Rishon LeZion Municipality: an active pilot of a bot on the website, on WhatsApp and in the contact center console, connected to the municipality's Dynamics 365. The bot answers only from approved sources and identifies the resident with a one-time code before any personal data.
Ashtrom Group · AI agent on WhatsApp: "Where is my truck?" is answered inside the conversation. The agent identifies the order, retrieves data from Priority ERP and calculates the arrival time based on traffic. A person joins only when there is a delay, a route change or missing information.
Israel Aerospace Industries · AICare: an AI layer for the iCare customer portal, with search across technical literature citing chapter and page, opening a service request from a photo of a nameplate, and the bot and agent in the same conversation.
Allot · Copilot in Salesforce: a knowledge assistant for service teams inside the Case screen, which shows similar cases and solutions that worked and inherits Salesforce permissions.
Frequently asked questions
What is the difference between a chatbot and an AI customer service agent?
A scripted chatbot replies with answers written in advance. An AI agent understands free-form questions, answers from approved sources, takes actions in systems such as opening a case or checking a status, and hands over to a human agent with the context.
What happens when the agent does not know the answer?
It says it has no authoritative source and does not guess. Instead, it offers a human agent, a callback or another channel, and the question is added to the list the content team fills in.
Can the agent access a customer's personal data?
Only after identification, for example with a one-time code checked against the CRM, and only according to the permissions of the source system. General information is open to everyone.
Do we need to replace our CRM or website?
No. The agent is built on top of existing systems. KESHER at Rishon LeZion Municipality is connected to the municipality's Dynamics 365, and Ashtrom Group's agent works with Priority ERP.
How do we get started?
With one domain and real questions from the contact center. Build a POC on the real sources, test answers, identification and handover, and only then add channels and actions.
CTO at Zooz-Tech. Leads development and architecture of AI agents, Generative UI and Azure infrastructure.