A customer support chatbot is useful when questions repeat and the answers already exist, written somewhere: in the terms of sale, in manuals, in FAQs, in closed tickets. shardana.ai develops custom AI chatbots that read those sources, query company systems, answer consistently with the brand's tone and state their own limits instead of making things up.
The work does not stop at the language model. It includes preparing the sources, connecting to CRMs, management systems and databases, also by developing custom connectors and APIs, integrating with the channels customers really use, escalation rules towards the team and monitoring of conversations. To see one of our chatbots at work, try AntasAI from the chat bubble at the bottom right.
What is a business chatbot for customer support?
A business chatbot for customer support is an assistant that answers customer questions using the company's knowledge and systems. It reads procedures, FAQs and contractual terms, but it can also query databases, CRMs, management systems and ticketing through dedicated tools, for example to read the status of an order or a case. With these elements it composes the answer and indicates where each piece of information comes from. When the request is delicate, such as a complaint, a payment or personal data, or when elements are missing, the chatbot does not leave the customer without an answer and does not improvise: it gathers and cross-checks all the relevant data and hands it to a team member in a ready summary, with a draft reply to approve. The agent does not waste time searching across several tools and can focus on the decision. The scope stays defined by you, and every step leaves a trace.
Want to see a chatbot at work? Try AntasAI
The fastest way to understand how we work is to try AntasAI, the Idea Engineer. It is a chatbot we developed ourselves that you can use right away, by clicking the chat bubble at the bottom right of this page. You describe a challenge that is still fuzzy, and it guides the conversation, explores the context and proposes three directions to evaluate: Quick Win, Scale and Disruptive. It relies on a knowledge base and keeps the dialogue within a precise scope, as we would do for your company's chatbot. It shows you in practice how we set the tone, the retrieval of content and the limit of what the system can claim. We do not publish client cases we cannot document: when a project has the client's consent we will tell it with verifiable timelines, stack and metrics. In the meantime, AntasAI is the demonstration you can touch.
Discover the details of AntasAIOpen AntasAI in a new tab
“A chatbot that makes up an answer costs more than one that admits it does not know: the first loses the customer's trust, the second earns it.” — Maurizio Brioschi, founder of shardana.ai
Rule-based chatbot, AI chatbot or SaaS platform: which one to choose?
The choice depends on how many different questions you receive, how much the answers change and how much you need to integrate. A rule-based chatbot follows hand-written decision trees: it is predictable and cheap, but it only handles standard questions and breaks as soon as the customer phrases the request differently. A custom AI chatbot understands natural language, answers from your documents and can be connected to your systems, with a higher initial investment and a usage cost tied to volumes. A SaaS platform is the quickest way to start, but customization, data ownership and integrations depend on the vendor. There is no answer valid for everyone: if the questions are few and fixed, rules are enough; if the knowledge is broad and changing, AI is better; if you only need a quick trial, a SaaS may be enough.
| Criterion | Rule-based chatbot | Custom AI chatbot | SaaS platform |
|---|---|---|---|
| Questions handled | Few and predictable | Broad, in natural language | Broad, within the limits of the product |
| Integration with your systems | Limited | Complete, defined in the project | Depends on the connectors available |
| Control over data and answers | Total but rigid | High: sources, permissions, logs | Partial, governed by the vendor |
| Initial cost | Low | Higher, tied to the scope | Low or none |
| Recurring costs | Maintenance of the decision trees | Models, infrastructure, improvement | Subscription, often per volume or user |
| When it fits | Fixed FAQs and low traffic | Broad knowledge and your own processes | Quick trial or standard needs |
Which channels does a chatbot work on: website and WhatsApp?
A chatbot can be published on the channels customers already use, and the two most requested are the website and WhatsApp. On the website it is embedded as a chat window, with the brand's tone and graphics, and it can know the page the customer starts from. On WhatsApp the conversation happens where customers are used to writing, but it requires the WhatsApp Business Platform, with a dedicated number, approved message templates and consent rules. The same core of sources and rules can serve both channels, so answers stay consistent. The work changes in the interface and the conditions of use: message length, handling of attachments and images, response times and handover to an agent. Other channels can be added, such as email or an internal app, when the use case justifies it. It is best to start from a single channel and expand after reading real conversations.
How do we stop the chatbot from making up answers?
A chatbot makes fewer mistakes when its task is narrow and verifiable. The system first retrieves the relevant information, from approved documents with a technique called RAG and from company systems through tools with precise permissions, and builds the answer only on those. Every answer reports the sources used, so whoever reads it, or checks it later, can verify it. When the data is not enough, the chatbot says so. For sensitive requests, such as complaints, payments or personal data, it prepares the case for an agent with the data already gathered and a draft reply to approve. After release, conversations are recorded and reviewed: unresolved cases show which content or integrations are missing, and they are corrected at the source. No system is infallible, but sources, confidence thresholds and escalation make mistakes rare, visible and correctable. The example below shows the system's output.
{
"question": "Can I change the delivery address after placing the order?",
"answer": "Yes, until it ships. After that, you need to contact the carrier.",
"sources": ["Terms of sale, art. 6", "Shipping FAQ, item 3"],
"confidence": "high",
"action": "answer"
}
{
"question": "I would like to dispute the March charge.",
"answer": null,
"confidence": "low",
"action": "hand_over_to_agent",
"summary_for_agent": {
"customer": "Customer record and contract read from the CRM",
"payments": "March charge found in the management system",
"history": "Two previous tickets on the same topic",
"draft_reply": "To be approved before sending"
}
}Illustrative example of the output format: the first answer cites its sources, the second hands the case to an agent.
How much does it cost to develop a business chatbot?
The cost of a business chatbot depends on the scope, so we do not publish a single price list. The main factors are the quantity and cleanliness of the sources, the channels to cover, the integrations with the CRM, management system or ticketing, the actions the chatbot can take besides answering and the privacy requirements. On top of the development spend come recurring costs that are best estimated from the start: the use of language models, which grows with the number of conversations, the infrastructure, the messages of the WhatsApp Business Platform and the time spent improving content. A first limited case, with a single source and a single channel, lets you measure results and volumes before widening the project. After the first discussion you receive a written proposal with phases, deliverables and costs, so you know what you are buying before committing to the whole.
| Item | Type | What it depends on |
|---|---|---|
| Analysis and scope | One-off | Number of processes and sources to map |
| Source preparation | One-off | Quality and format of existing documents |
| Development and integrations | One-off | Channels, systems to connect, actions allowed |
| Language models | Recurring | Number and length of conversations |
| Infrastructure and channels | Recurring | Hosting, WhatsApp Business Platform, monitoring |
| Continuous improvement | Recurring | Unresolved cases and content updates |
How long does it take and how is a chatbot developed?
A chatbot is developed in short steps, so that you see real results before investing more. It starts with a discussion of your current customer service, the most frequent questions and the tools in use. An analysis follows that defines the sources, the scope, the escalation rules and the success criteria, for example the share of requests resolved without an agent. Then we build a prototype on a limited case, which the team tries with real questions. If it works, the chatbot is connected to the channels and systems, refined on unresolved cases and released. After release the work continues, because customer questions change and content must be updated. The durations in the table are indicative and we confirm them after the analysis, because they depend mainly on the availability of the sources and on how quickly your team validates them.
| Phase | What happens | Indicative duration |
|---|---|---|
| First discussion | Current customer service, frequent questions, tools | One meeting |
| Analysis and scope | Sources, escalation rules, success criteria | About one week |
| Prototype | Working chatbot on a limited case, tried by the team | 2-3 weeks |
| Development and integration | Channels, connected systems, refinement on real cases | 2-4 weeks |
| Release and improvement | Production, monitoring, content updates | Ongoing |
Frequently asked questions
- Can the chatbot use only my company's information?
- Yes. The system reads only the documents and systems you authorize, with precise permissions, and states when it does not have enough context to answer, preparing the case for a person.
- Can the chatbot read data from the CRM or the management system?
- Yes. It can query databases, CRMs, management systems and other company tools through dedicated tools. If a system offers no APIs, we develop a connector or a custom API to bridge the data.
- Do we need to have a well-organized knowledge base already?
- No. You start from existing documents, FAQs, tickets and procedures, then organize and improve them as real customer questions emerge.
- Can a chatbot work on WhatsApp?
- Yes, through the WhatsApp Business Platform. It requires a dedicated number, approved message templates and consent rules. The sources and rules can be the same as the website chatbot.
- How are complex cases or complaints handled?
- With escalation rules: the chatbot gathers and cross-checks data from the CRM, the management system and the history, prepares a summary with a draft reply and passes the conversation to the team, who decide without having to search for information.
- How much does a business chatbot cost?
- It depends on the scope: sources, channels, integrations and volumes. After the first discussion you receive a written proposal with phases, deliverables and costs, including the recurring expenses of models and infrastructure.
- Does the chatbot need to be managed after release?
- Yes. Questions and content change, so it is worth reviewing unresolved conversations and updating the sources. Monitoring is part of the project.
Related services
- AI development for businesses: the pillar page with all AI solutions.
- MCP servers: the layer that connects AI agents to company data and tools.
- AntasAI: the Idea Engineer chatbot, to try online.
- AI automations for businesses: processes across CRMs, management systems, email and documents.
Tell us about your customer service
Describe the questions you receive most often, the channels you use, where the answers are written and in which systems the data lives. We will reply with a first assessment and, if it makes sense, with a proposal for a first concrete case.