AI chatbot development: custom chatbots for your website

AI chatbot development: what to automate first

Start with repeated questions whose answers already exist and whose mistakes your team can spot. Shipping policies, product compatibility and service coverage are useful candidates when the underlying content is current. Leave disputed refunds, negotiated prices and unusual account changes with staff.

• Store support

  Answer product and policy questions with links, then retrieve order status and tracking after verification. Collect a return request without promising that a refund has been approved.

• A guide to your expertise

  Help visitors find answers in your own articles, guides or course material. Show which approved page supports the explanation and acknowledge missing information.

• Service inquiries

  Explain services, collect the details staff need and connect a scheduling flow. Confirm availability through the calendar rather than letting the model invent an appointment.

In 100 recent Upwork job posts we reviewed in October 2026, 5 were for chatbots. Buyers described store support using approved FAQs and policies with links, verified order updates and tracking, return intake and escalation; other requests covered a guide grounded in an expert's material and chatbot scheduling with Google Workspace. These are requested capabilities, not measured project results.

If the assistant must also take actions in your tools, read AI agent vs chatbot before you scope it.

AI chatbot development starts with approved content

A chatbot trained on your own data: what that means

For many support projects, the useful approach is retrieval: search your approved material, give the relevant passages to the model and ask it to answer within that evidence. This is often called retrieval-augmented generation, or RAG. It usually means connecting a searchable knowledge library, rather than retraining the model whenever your return policy changes.

Retrieval helps ground an answer but does not guarantee correctness. The assistant should link the policy or product page behind its claim, avoid filling gaps and recognize when sources disagree. A link is useful only if the linked passage actually supports the answer.

Website, WhatsApp, app or email?

A custom AI chatbot for website visitors

Choose the channel where the questions already arrive. A website widget can begin with public answers, while an assistant inside your app can use the signed-in customer's permissions. Adding channels also adds inbox routing, conversation history and testing responsibilities.

WhatsApp chatbot for business: channel rules

In the same Upwork review, 2 posts were WhatsApp Cloud API work. We scope the business account, approved templates and staff inbox alongside the bot; see Meta's WhatsApp documentation.

Private order data needs verified access

An order number typed into chat is not sufficient proof of identity. Use an authenticated account or an agreed verification flow, then have the application check permission before it reads an order. A phone number or convincing message must not override those checks.

Customer messages and retrieved documents can contain instructions that try to redirect the assistant. Keep permissions and refund rules in the application, outside the model's discretion. Test whether a request to ignore its rules or expose another customer's information is blocked.

An AI customer support chatbot needs a human route

Handoff is a working support process: an assigned destination, the relevant context and a clear expectation for the customer. Offer it when someone asks for a person, verification fails, sources do not settle the question or the case needs an exception. Avoid trapping a frustrated customer in repeated clarification questions.

• Give staff a useful case

  Pass the question, verification status, relevant source links and a short summary. Include only the private details the receiving team is allowed to see.

• Explain what happens next

  Show whether staff are available and where the reply will arrive. Outside staffed hours, collect an agreed contact route without promising an immediate response.

• Pause automated replies

  Once a person takes over, stop the bot from talking over them. Define how the conversation returns to automated support after staff close the case.

English and Spanish support need the same factual boundaries, with language carried into the staff queue. Test product names, mixed-language questions and local date formats. A translated greeting alone does not make the entire support journey bilingual.

Conversation and customer messages in real products

WLTR.AI lets users ask about business performance through chat and receive answers about revenue, subscriptions and marketing activity. Our role was the product build, using Next.js, React, TypeScript, Supabase, Stripe and OpenAI. It shows how a mobile conversation can present answers about business data.

San Blas On Board brings boats with different owners and rules into a booking workflow with rule-based quotations, calendar coordination and guest check-in. The quote builder includes an email preview, a useful example of preparing customer messages from structured booking details. It shows how structured booking details feed a customer message preview.

How our AI chatbot development process works

We agree on the first channel, knowledge scope, integrations and acceptance examples before implementation. The written scope also identifies what staff will still handle and who maintains the answers. This makes the launch decision something your team can review.

1. Step 1

   Map the questions and the support route

   Review anonymized inquiries, current volumes and common exceptions. Choose the first useful job, define the human destination and record what a successful conversation should achieve.

2. Step 2

   Approve content and permissions

   Identify the authoritative FAQs, policies and product records. Separate public answers from private lookups and agree on verification, retention, staff access and allowed actions.

3. Step 3

   Build retrieval and agreed connections

   Connect approved content, source links and the selected channel. Add order, CRM or scheduling access only within scope, with controlled failures and protection against duplicate requests.

4. Step 4

   Test ordinary and difficult conversations

   Check supported answers, outdated documents, unknown questions, unauthorized access, unavailable systems and handoff. Review English and Spanish examples with the people who handle support.

5. Step 5

   Pilot, review and hand over

   Begin with a limited audience and review real outcomes before expanding. Deliver source code, account access, documentation, content update instructions and an agreed way to pause or roll back the assistant.

AI chatbot development cost: scope and running costs

AI chatbot development cost depends on channels, content readiness, private data access, integrations, languages and the support process. We provide a fixed written quote before work starts. Our AI automation cost guide explains how to compare build scope and ongoing expenses.

• Build and subscriptions

  Separate implementation from hosting, inbox tools and channel charges. Identify which accounts you need and check current vendor pricing before committing.

• Usage and limits

  Estimate AI use from expected conversations, answer length and lookups. Track actual spending, restrict abuse and define alerts plus a pause or handoff when limits are reached.

• Maintenance and staff time

  Budget for policy updates, integration changes and reviewing unanswered questions. Someone must own escalations, content approval and checks after changes.

Start with the smallest channel and knowledge scope that can resolve a useful task. Agree on support, maintenance and any later channels separately so the quote is clear. Payment is by card through Stripe after scope is agreed; source code, accounts and documentation belong to you.

Measure useful answers and keep control

Measure customer outcomes alongside costs. An assistant that avoids transfer by leaving someone stuck is not resolving support. Review the conversations behind the numbers before changing the prompt or expanding access.

• Deflection with a resolution check

  Track eligible support conversations resolved without staff, divided by eligible conversations. Define resolved explicitly and review repeat contacts so silence is not counted as success.

• Handoff rate and delivery

  Track conversations sent to staff and verify that the case reaches the right queue. Break down the reasons: requested person, missing evidence, verification issue or exception.

• Unanswered questions

  Keep a reviewed list of missing or conflicting answers. Assign content owners, update the approved source and repeat the relevant checks before publishing the change.

For agencies: white-label

We offer white-label development for agencies under your brand, with NDA and non-solicitation available. Agree on client communication, account ownership, support routing and who approves policies. Your agency can present the delivery while we build and document the agreed system.