Mistake 1: launching without a defined customer task
A chatbot added because every competitor has one can become a distracting interface with no clear job. The business may expect sales, support, bookings and specialist advice without defining which tasks the assistant can actually handle. Start with one visitor need and the next step your team can provide. A service question followed by an optional callback request is concrete enough to test. Automate the part you can explain and keep other responsibilities explicit.
Write a short acceptance criterion: the visitor receives an approved answer, understands whether confirmation is needed and can choose a relevant contact route. Give the receiving staff member an example record and ask what they would do next. If different people describe different promises, resolve that before launch. A setting labelled automation does not establish a working end-to-end customer process.
Mistake 2: training on everything without reviewing it
A large source collection can include conflicting policies, expired prices and private documents. More text is not automatically more useful. Choose current information connected to the customer task and remove withdrawn material. If a service requires a quote or staff approval, preserve that condition in the source. Someone in the business should be responsible for verifying the knowledge rather than assuming an import process also performs factual review.
Test the facts most likely to affect a decision: availability, location, price basis, timing and important exclusions. Ask a question whose answer is absent and check that the assistant acknowledges the gap. A fluent invented answer can be more harmful than a clear limitation because the visitor may rely on it. Update the relevant website content and chatbot knowledge together so they do not make different promises.
Mistake 3: demanding contact before giving an ordinary answer
An opening-hours question should not become a long sales questionnaire. Early contact demands can make the assistant feel like an obstacle rather than help. Answer the supported question first and offer a next step only when it is relevant. If the visitor wants a callback, explain the information needed for that response. Do not assume every conversation should end in a phone number simply because contact count is easy to display.
Keep optional fields optional. If a visitor declines contact, acknowledge the choice and continue providing appropriate public information. Repeated prompts after a refusal undermine that choice. A qualification question should clarify the requested service or prepare a human response, not collect information for an unrelated marketing purpose. Inspect abandoned conversations to see whether the prompt was relevant before concluding that visitors were not serious.
Mistake 4: confusing requests with completed actions
A preferred appointment time is not a confirmed booking. A question about stock does not prove live inventory access. A visitor asking for a refund has not triggered an approved account change merely because the assistant understands the message. Write the confirmation boundary into the reply and the saved record. State who reviews the request and avoid generating fictional confirmation numbers or guarantees when no connected system has performed the action.
During a trial, ask for an unavailable slot, a policy exception and an unsupported operational change. Check that the assistant preserves the limitation when the visitor pushes for certainty. The website promise should match the actual process. If a real integration is later implemented, demonstrate the action, permission and failure handling before changing the wording from request collection to confirmed execution.
Mistake 5: advertising a contact link as native integration
A WhatsApp link opens another app; it does not establish an API inbox, campaigns or transcript synchronisation. A separate enquiry workspace does not prove that HubSpot, Zoho CRM or a spreadsheet receives the record automatically. These distinctions are material to a buyer who needs a complete operational workflow. Describe the verified capability plainly and mark a connector unconfirmed until there is a working implementation and a clear support scope.
Ask what happens after the visitor clicks or submits. Who receives the message? Does context follow? Which fields are mapped and who handles a failed transfer? Manual copying can be a deliberate process, but it should not be marketed as real-time synchronisation. Keep the required existing system available until any proposed replacement has demonstrated the full journey your business depends on.
Mistake 6: testing only polished happy-path questions
Demonstrations often use a complete question and current information. Customers use abbreviations, change their minds and combine requests. Include those cases in the test set. Ask a short ambiguous question, an unsupported service request and a question in the language customers actually use. Then inspect whether the answer preserves source facts and whether the final summary reflects the latest requirement rather than incompatible earlier answers.
Test the contact refusal and the unavailable-answer branch as carefully as the ideal enquiry. Have another reviewer compare the replies with the approved source. Keep a labelled set of repeatable tests so a later instruction change can be checked against the same failures. Do not conclude that the assistant has a guaranteed accuracy rate from a handful of convenient examples or from fluency alone.
Mistake 7: overlooking mobile and alternative contact routes
A widget can look tidy on a desktop screenshot while covering a phone screen action. Open it at a narrow width, use the on-screen keyboard and inspect long replies. The visitor should be able to close the window, read the answer and reach the final contact request without losing access to essential page controls. Check product drawers, fixed navigation and consent interfaces that may share the same corner.
Keep a form, email or phone route available for people who prefer it or encounter a temporary problem. Test those links as part of launch. A chatbot should simplify the customer task rather than trap the visitor in one channel. When changing appearance, repeat the functional checks; a smaller margin or larger header can alter the usable conversation area even if the page still looks attractive.
Mistake 8: collecting sensitive details without a clear need
General lead conversations should not request passwords, card details, identity documents or extensive health and financial histories. Ask for the minimum information needed for the requested next step and provide the applicable privacy information. A clinic appointment enquiry can be administrative without gathering diagnostic information. An insurance callback request can identify a broad category without attempting underwriting in public chat.
Confirm access and retention responsibilities with the operator. A privacy page alone does not establish every legal or contractual arrangement for a deployment. Avoid using customer transcripts as public marketing examples without an appropriate basis and permission. Contact for a requested service and consent to repeated promotion are different purposes; keep them separate in wording and business processes.
Mistake 9: measuring conversation volume as business success
A busy widget may answer useful questions, collect poor-quality contacts or create unattended records. Count the intended stages separately: eligible visits, conversations, useful enquiries and staff-confirmed outcomes. Define the unit and exclude labelled tests. A contact record is not automatically a sale, and a request is not a confirmed appointment. Keep observation dates and the follow-up window visible beside any rate you eventually publish.
Record other changes such as traffic campaigns, prices and staffing when comparing periods. Without an appropriate design, a before-and-after increase does not establish that the chatbot caused it. A calculator can illustrate assumptions but is not measured evidence. Review actual records and limitations before publishing a claim. Never add fake leads to make analytics appear active or invent a customer result for a comparison page.
Mistake 10: leaving follow-up and maintenance unowned
The assistant can collect enquiries outside staffed hours, but someone still needs to respond. Assign responsibility for reviewing records, handling missing information and confirming requests. Test that process by sending a labelled enquiry and asking the recipient to act on the brief. If nobody knows who replies or what was promised, improving the greeting will not solve the failure. Keep realistic response expectations in the public wording.
Schedule knowledge review when services, hours, policies and prices change. Use a small sample of appropriately handled real conversations to identify one recurring issue, then make and retest a targeted change. Preserve a rollback method for the website snippet and keep another contact route available. A dependable chatbot is maintained as part of the business process, with accurate sources, clear boundaries and a person responsible for the next step.
Turn the review into a short correction plan
Pick the mistake that currently prevents the customer next step and record a specific change. For missing knowledge, add the approved fact and test the unanswered question again. For contact friction, move the request later or remove an unnecessary field. For unclear ownership, name the receiving person and verify a real handoff. Avoid changing several unrelated settings at once, because it becomes harder to tell which edit resolved the problem.
Repeat the same checks after the correction and inspect the resulting enquiry. The aim is a conversation a visitor can trust and a next step the business can provide, not a larger number of prompts or an impressive feature label.