A website chat assistant for common customer enquiries

What shipped

Chat tools are no longer only for large support teams with a full helpdesk stack. The useful change for an Irish SME is simpler: a website or portal chat assistant that answers common questions from approved company material, asks a few qualifying questions, and creates a handover note when it cannot answer safely.

This is not a replacement for customer service. It is a first-line filter for the questions your team already answers every day. Opening hours, booking rules, returns, service areas, delivery options, warranty steps, document requirements and basic pricing guidance are all good candidates, provided the answers are already agreed inside the business.

The important part is grounding. The assistant should not make up policy. It should retrieve answers from a controlled knowledge base, such as selected web pages, service documents, booking rules, FAQs, email templates and internal notes that have been checked. If the answer is not in that approved material, it should say so and offer a handover.

For an SME, the most useful shipped version is usually small. Pick one enquiry area, set the boundaries, test privately, then decide whether it is good enough for live use. That is a good fit for an Artellis AI prototyping engagement, because the goal is to prove the workflow in days or a few weeks before anyone commits to a broader rollout.

The human review step should be named from the start. I would call it the Customer Handover Review. A named staff member checks all handover notes, corrects any weak classification, and approves any proposed new answer before it is added to the live knowledge base. Until that review happens, the output is decision support only.

Who should care

This is worth a look if your customer service, sales or reception team gets the same basic questions all day, but still needs a person to judge what the customer actually wants.

It suits businesses where the first exchange matters, but is often repetitive. Examples include clinics handling appointment questions, trades and service firms qualifying jobs, training providers dealing with course enquiries, retailers answering return questions, venues handling booking queries, and B2B suppliers receiving product availability or account queries.

It is especially useful where enquiries arrive outside office hours. The assistant can give a safe answer at 8 pm, collect the customer’s name, contact details, order or booking reference, and ask what outcome they need. If it cannot answer, it creates a clear handover note for the next working morning.

The value is not only fewer replies typed by staff. The better measure is the quality of the first pass. Did the assistant identify the enquiry type? Did it gather the missing details? Did it avoid answering outside policy? Did the staff member get enough context to respond quickly?

This also helps owner-managers see the real shape of customer demand. After a few weeks, the chat log will show which questions keep recurring, where website content is unclear, and which enquiries should be routed to sales, support, accounts or operations.

It is not a good fit where every enquiry is high risk, heavily regulated, emotionally sensitive or dependent on judgement that cannot be written down. In those cases, a simpler triage form or internal drafting assistant may be safer. For more on practical SME use cases, see the Artellis AI insights library.

A worked example

Take a Dublin-based service business with 35 staff. It takes bookings for home visits, maintenance calls and small projects. The office team receives phone calls, website forms and emails asking the same questions: Do you cover my area? How soon can someone come out? What information do you need? Is there a call-out charge? Can I reschedule? What happens if the job is urgent?

At the moment, staff copy and paste parts of old replies, then ask follow-up questions manually. Some enquiries arrive overnight and sit until the next morning. Others are vague, so the first reply is only a request for more information.

A private prototype could start with one enquiry area: booking a standard service visit. The approved material might include the service area list, booking hours, call-out policy, required photos or documents, cancellation rules and escalation conditions. The assistant is instructed not to quote a final price unless the pricing rule is explicit. It can explain what usually happens, collect details and mark the enquiry for review.

A customer asks: “Can you send someone to Bray tomorrow to look at a leaking unit?”

The assistant checks the approved service area material. It asks for the Eircode, contact number, whether the leak is active, and whether there is any visible electrical issue. If the customer mentions a safety risk, the assistant does not diagnose. It gives the approved safety wording and routes the handover as urgent.

The handover note might read: “Urgent service enquiry. Location: Bray, Eircode provided. Issue: active leak from unit. Customer requests tomorrow. Safety concern: customer mentioned water near socket. Recommended route: call customer before 10 am, do not book without staff review.”

That note is far more useful than a raw chat transcript. A staff member can act quickly, and the business can measure whether the assistant gathered the right details.

After 30 real customer questions are tested privately, the team reviews failures. Maybe the assistant was too vague on cancellation rules. Maybe it tried to answer a question that should always go to accounts. Maybe the service area document had an old county list. These are workflow findings, not just bot findings.

Before live use, the business agrees three lists: questions the assistant may answer, questions it must hand over, and questions it must refuse or redirect. That boundary document is as important as the software.

Where I'd start

Start narrow. Choose one enquiry area where the answers are stable and the team already has agreed wording. Bookings, returns, service eligibility and document requirements are usually better starting points than complaints or complex technical advice.

Gather the material your best staff member already trusts. Do not upload everything. Use the current FAQ, policy snippets, service pages, approved email replies and any internal notes that describe the real process. Remove outdated documents and mark anything uncertain.

Then write the safe answer boundaries in plain English. For example: “The assistant may explain our standard returns window, but must hand over damaged goods, missing refunds and legal threats.” Or: “The assistant may describe appointment types, but must not confirm clinical suitability.”

Test privately on 30 real questions from email, forms or call notes. Score each chat on four points: correct answer, useful qualification, safe escalation and clear handover. If a chat fails, decide whether the fix is better source material, a clearer boundary, or a different route to a person.

Name the reviewer before launch. The Customer Handover Review should have an owner, a daily rhythm and a simple rule: no new answer goes live until a person has approved it. That keeps the assistant useful without letting it quietly drift away from company policy.

For measurement, keep it practical. Track the percentage of chats resolved without staff typing, time to first response, handover completeness, and the number of new or corrected knowledge base entries approved each week. If those numbers improve and staff trust the handovers, you have a useful workflow. If not, stay in prototype mode and tighten the scope.