What is the practical answer?
A useful AI chatbot is a business system with a defined job, approved source information and a reliable handoff. Compare what the company will connect, test and maintain before comparing model names or a ‘best agency’ list.
Decide whether you need a chatbot, agent or workflow
A chatbot can answer questions and guide a visitor. A workflow can move predictable fields between systems. An AI agent can interpret a request and use approved tools to complete a bounded task. The right company should help you choose the smallest system that solves the job.
Ask what happens when information is missing. A clear handoff to a person is part of a reliable design; an assistant that invents a confident answer is not a successful customer experience.
Review the knowledge and integration plan
The proposal should identify the pages, FAQs and documents the assistant can use, who approves updates and how conflicting information is resolved. It should also distinguish a link to a booking page from reading live availability, creating an appointment or changing a CRM record.
Bring the actual systems you use. The supported APIs, authentication, rate limits and data permissions determine what can be built safely. A model name alone does not answer those implementation questions.
- Approved source pages and documents
- Booking, CRM, help-desk or payment integrations
- Allowed actions and human approval points
- Failure, retry and escalation behavior
- Ongoing ownership and model or hosting costs
Ask for an evaluation set
A serious company should test ordinary questions, unclear wording, outdated information, requests outside scope and a handoff. Ask to see how links, prices and service details are checked against the approved source. If the business needs a connected action, test an interrupted API and a duplicate request too.
Measure answer usefulness, correct handoffs and completed next steps. Do not call the system 100% accurate without a defined test set, date, data source and error definition.
Separate build cost from running cost
Interface design, knowledge preparation, evaluation, website integration and connected tools affect the initial build. Model usage, hosting, storage, third-party subscriptions, monitoring and content maintenance affect ongoing cost. A quote should show which assumptions change when conversation volume grows.
Ask who owns the account, source code and prompts, and how you can export the approved knowledge. These details matter when the chatbot becomes part of customer support.
| Question | Good evidence | Warning sign |
|---|---|---|
| What does it know? | Named pages and an update owner | ‘Trained on everything’ without a source list |
| What can it do? | Documented tools and approval boundaries | Unrestricted autonomous actions |
| How is it tested? | Representative questions and failure cases | A demo with only ideal prompts |
| What will it cost? | Build, usage and support separated | One unexplained monthly figure |
| What happens next? | Relevant links and human handoff | A chat that ends without an actionable route |
Where Melos fits
Melos builds website assistants around a company’s approved information and can scope a separate integration when a chatbot needs to connect to a supported system. The AI chatbot page, cost guide and published Webora chatbot project show the intended boundary: useful service guidance, clear next steps and no claim of perfect or unrestricted automation.
Send the website, recurring customer questions, source documents, systems to connect and the person who approves answers. Melos can turn that into a bounded first release with tests, fallback behavior and a handover plan.
Review the AI chatbot service ↗Read what affects chatbot cost ↗