Chatbots that answer from your own content and hand off to a person when they should
We build website, helpdesk and customer-service chatbots that answer only from documents you approve, show where each answer came from, and pass the conversation to your team when they are unsure. Three of them are live below: ask them anything.
AI chatbot development services design, build and run chatbots that answer questions for customers or staff. A well-built chatbot answers only from content you approve, such as help articles, policies and product documents, shows where each answer came from, and passes the conversation to a person when it is not sure. Metwaves builds them for websites, Microsoft Teams and customer service, and you can try three of our chatbots live before you talk to us.
Models and platforms we build chatbots with
If two or more of these sound familiar, a chatbot built on your own content will usually pay for itself in time saved.
Six kinds of chatbot, one way of keeping them accurate: approved sources, citations and a clear handoff.
Answers product, order and policy questions on your website from your help centre and documents, and opens a ticket with the full conversation when it cannot help. Customers get an answer at any hour; your team gets fewer repeat questions and better tickets.
Runs in Microsoft Teams, answers from SharePoint policies and respects who may see which document. Built with Copilot Studio when your firm runs on Microsoft 365.
Ask contracts, manuals and policy libraries questions in plain English and get the paragraph that answers it.
Answers pre-sales questions, collects what your team needs to know and books the call.
Guides customers through reservations, plan details or course choices using your own data, like our car rental, healthcare plan and university demos.
Already have a bot that gives poor answers? We test it against your real questions, show where it fails and fix the content, retrieval or model behind it.
Most chatbots fail because they guess. Ours check what the person may ask, find the passages that answer it in your approved content, and only then let the language model write the reply. If the answer is not in your content, the bot says so and hands off.
The chatbot reads only from the sources you approve and writes only to the systems you allow, such as your helpdesk for tickets or your CRM for leads.
Each demo answers from its own content. Ask the sample questions, then try your own. Demos accept messages up to 300 characters and 40 messages a day.
Explains plan benefits, deductibles and coverage, and helps find an appointment slot.
Try asking
Gemini Pro on Google Cloud
Open the demoAnswers prospective and current students on courses, fees, deadlines and visas.
Try asking
Gemini Pro on Google Cloud, trained on a public university website
Open the demoHelps customers choose and reserve the right vehicle for a trip.
Try asking
Gemini Pro on Google Cloud
Open the demoof staff requests resolved without a person by our Teams helpdesk agent
from start to live, including content clean-up and a pilot group
chatbot demos you can try on this site today
Source: our Copilot Studio helpdesk case study for a UK professional services firm.
Agree these numbers before launch, so the bot is judged on results rather than on how clever it sounds.
| Measure | What it tells you | Why it matters |
|---|---|---|
| Resolution rate | Share of conversations the bot finishes without a person | The main return on the project. Our Teams helpdesk agent resolves 30–50% of staff requests. |
| Answer accuracy | Share of test questions answered correctly, from the right source | Measured on a fixed question set before launch and after every content change. |
| Handoff quality | Whether the ticket a person receives has the question, context and transcript | A bad handoff makes the customer repeat themselves, which costs more than no bot. |
| Unanswered questions | Questions the bot could not answer, grouped by topic | Each month this list tells you which content to write or fix. |
| Customer satisfaction | Thumbs up or down, or a one-question survey at the end of a chat | Read alongside resolution: a bot that closes chats but annoys people is not working. |
A professional services firm in the United Kingdom wanted staff to get answers to everyday IT, HR and policy questions inside Microsoft Teams. The answers already existed, spread across SharePoint sites where old and current versions sat side by side.
We cleaned up the content with the firm's owners, built a Copilot Studio agent that answers from SharePoint through an Azure AI Search index, cites the policy it used, respects document permissions, and opens a ticket with the conversation attached when it cannot help.
Answer three questions to see what usually fits. We will tell you if a ready-made tool is enough.
Must it change records in your systems?
Refunds, bookings, updates in your CRM
Do answers come from many documents?
Policies, manuals, contracts, help articles
Do your staff work in Microsoft 365?
Teams, SharePoint, Outlook
It needs to take actions, so it needs permissions, guardrails and human review. See AI agent development.
Your content and people are already in Microsoft 365, so build where your staff work.
Retrieval from approved content, with citations and a handoff to your team.
For a few FAQs on one website, a ready-made tool such as Intercom or Tidio is cheaper. We will say so.
The job changes by industry, and so do the rules about what a bot may see and say.
| Industry | Common chatbot jobs | What to watch |
|---|---|---|
| Healthcare | Plan benefits, appointments, clinic information | Patient data rules such as HIPAA in the US: the bot answers general questions without personal data unless it is built and hosted for it. Healthcare demo |
| Education | Courses, fees, admissions, student services | Student records stay out of a public bot; account questions need sign-in. University demo |
| Professional services | Staff IT, HR and policy questions | Document permissions must carry through, so people only see what they may open. Case study |
| Travel and rental | Bookings, changes, vehicle and trip options | Live availability comes from your booking system, never from the model. Car rental demo |
| Retail and eCommerce | Order status, returns, product questions | Order lookups need the customer to identify themselves before the bot reads the order. |
| Financial services | Product and account FAQs, document requests | No account actions without strong sign-in, and every answer traceable to an approved source. |
Each step ends with something you can test, so you never wait months to see the bot answer a real question.
We agree what the bot may answer, which sources it may use and where it hands off. We also collect a set of real questions your customers or staff ask; it becomes the test set.
A working bot on your real documents, not a demo dataset, so you judge it on the questions that matter to you.
We score every answer in the test set for accuracy and source, fix the content or retrieval behind the failures, and repeat until the scores hold.
Website, Teams or another channel, with a clear route to a person or ticket. Our Teams helpdesk agent went from start to live in 8 to 12 weeks.
We review unanswered and low-rated questions, add or fix content, and re-run the test set before each change.
We quote after we have seen your content and channels. These are the factors that move the price most.
One clean help centre is quick to connect. Ten SharePoint sites with old and new versions side by side need clean-up first, as they did in our Teams helpdesk project.
Answering questions is one job. Looking up an order, booking a slot or opening a ticket in your system means building and testing each connection.
Each channel (website, Teams, a messaging app) has its own setup, sign-in and testing.
A bot that answers policy or health questions needs a larger test set and more review rounds than a product FAQ bot.
Each extra language needs its own test questions, and sometimes its own content.
There is also a monthly running cost: model usage per conversation, hosting, and the monthly review that keeps answers current. Ask any vendor for both numbers.
Every engagement starts with a call about your questions, content and channels.
| Model | Best for | How it is priced |
|---|---|---|
| Fixed-scope chatbot build | A defined set of sources, channels and handoffs | Fixed quote after scoping |
| Monthly improvement and support | A live bot that needs content updates, monitoring and new topics | Monthly |
| Audit of an existing bot | A bot that is live but gives poor answers | Fixed quote |
The models and platforms below run in our live demos, our Teams helpdesk agent and our AI agent projects. We choose per project and can change the model later without rebuilding the bot.
Azure OpenAI
Dialogflow CX · Google Cloud · LangGraph
Azure AI Search · SharePoint · Retrieval-augmented generation (RAG)
Questions that separate a team that has run chatbots in production from one that has only built demos.
When the assistant must take actions in your systems, not only answer.
Connect the bot to your CRM, helpdesk, booking or accounting tools.
Prove an AI assistant works on your own data before you fund the full build.
Our guide to choosing between building and buying a chatbot.
We will show you how a chatbot built on your content would answer them, and tell you honestly whether a ready-made tool would do the job.
Book a Chatbot CallIf you don't see an answer to your question, you can send us a message from our contact form.
It depends mostly on how many content sources the bot reads, which systems it must connect to and how many channels it runs in. A product FAQ bot on one website is a small project; a helpdesk agent that reads several SharePoint sites and opens tickets is a larger one. There is also a monthly running cost for model usage, hosting and improvement. We quote after we have seen your content and channels, so the price reflects your project rather than a package.
A first working version on your own content usually comes early in the project, so you can judge it on real questions. Our Microsoft Teams helpdesk agent went from start to live in 8 to 12 weeks, including content clean-up, testing and a pilot group.
The bot answers only from content you approve, shows the source for each answer, and says it does not know when the answer is not there. Before launch we test it against a set of real questions and fix the content or retrieval behind every wrong answer. After launch, unanswered and low-rated questions are reviewed every month.
Our live bots run on websites and in Microsoft Teams. Messaging apps such as WhatsApp connect through their business APIs, and we confirm the channel, sign-in and handoff during scoping.
A chatbot answers questions. An AI agent also takes actions in your systems, such as updating a record, booking a slot or posting an invoice, within the permissions you give it. If you need actions, see our AI agent development service.
You do. The content, configuration, conversation logs and any code we write belong to you and live in your accounts where the platform allows it.
Yes. We test your current bot against your real questions, show which answers fail and why, and fix the content, retrieval or model settings behind them, or recommend a rebuild if that is cheaper.
Retrieval-augmented generation means the chatbot first searches your approved content for the passages that answer a question, then has the language model write the reply from those passages only. It is what lets a chatbot answer from your policies and products instead of guessing, and it is how each answer can link to its source.
With the enterprise model services we build on, such as Azure OpenAI, your documents and conversations are not used to train the provider's public models. We confirm the data terms for the model you choose during scoping, and we keep your content in your own cloud account where the platform allows it.
Ask to try a bot they have built, ask how it handles a question it cannot answer, and ask for the monthly running cost as well as the build price. The checklist above lists the signs we would look for.
Discover IT Industry Insights and Technology Trends
Nine AI agent development firms a small or mid-sized business can shortlist, with Clutch figures, minimum budgets and the weaknesses their own clients name. Published by Metwaves, which is not ranked.
Learn More >An architecture review of the four big enterprise AI agent platforms: where each one reasons, acts, finds its data and connects to outside agents, what it costs, and when a custom build is the better choice.
Learn More >Ten custom software firms a small or mid-sized business can shortlist, with Clutch figures, minimum budgets and the weaknesses their own clients name. Published by Metwaves, which is not ranked.
Learn More >
Let us help you with free analysis & estimate.
Our product expert will contact you within 12 hours.