Customer service
AI knowledge-base assistant for your policies
Collect your real policies first: cancellations, check-in, payments, pets, accessibility. Put them in one current source, then connect an assistant that answers only from that source and says when it does not know. Start internally with staff, then open it to guests with a clear handover to a person. Review wrong answers weekly and fix the source, not just the bot.
Effort:
Medium
Time saved:
Roughly 20–40% fewer repeat policy questions; measure it in a pilot
The problem today
Guests and customers ask the same policy questions by email, chat and phone, and each staff member answers a little differently. Policies sit in old PDFs, emails and people's heads, so mistakes and disputes follow.
How it works, in plain words
Think of a new receptionist who has read every policy you have, answers only from those pages, and fetches a colleague when the answer is not there.
What goes in: your current, approved policies and FAQs, written as short questions and answers in one place.
What the AI does: it searches those documents first for the passage that fits the question, then writes a short answer and shows which document it used.
What a person checks: staff review wrong or incomplete answers each week and fix the source document, not just the bot.
What comes out: consistent answers for staff, and later for customers, with a clear handover to a person when the assistant is unsure.
It knows nothing beyond your documents. If a policy is outdated or missing, the answer will be outdated or missing too.
How it could work for businesses like yours
Illustrative examples, not client case studies.
A 12-person branding agency in Chalandri
Designers keep asking the account team about each client's brand rules, approval steps and file formats. The agency would collect every client's brand guidelines and its own working policies into one source and set up an internal assistant that cites the page it used. They would track how many of these questions reach account managers before and after.
A 14-person sportswear e-shop in Peristeri
Customer service gets daily questions about returns, exchanges and delivery times, and answers vary by person. The shop would rewrite its policies as short Q&As, test the assistant with about 100 past questions, and use it internally first. A realistic first target could be the same answer to the same question from every agent.
A short-term rental manager in Rhodes with 35 apartments
Guests message about check-in, parking, pets and cancellations, often late at night. The company would build an assistant on one policy document per property and open it to guests only after weeks of internal testing, with a clear route to a person. They would measure how many messages still need a manual reply.
How to do it, step by step
Gather every policy and FAQ in one place. Remove duplicates and outdated versions, and give each topic an owner.
Rewrite policies as short questions and answers. Include the exceptions and who is allowed to approve them.
Set up an assistant that answers only from these documents and shows its source. Tell it to say "I don't know, let me connect you with a colleague" when the answer is not there. This approach is often called RAG.
Test it with 50–100 real past questions in Greek and English. Track wrong, incomplete and invented answers, and fix the documents behind them.
Launch internally first, for reception or customer service staff. Open it to guests on your website only when wrong answers are rare and handover to a person works smoothly.
Label it clearly as an AI assistant, review conversations every week and update the source whenever a policy changes.
Tools to consider
Custom GPTs in ChatGPT: a quick way to build an internal assistant on your uploaded documents. Caveat: fine for staff use, but a public website needs more control over answers and logging.
Claude Projects: keeps your policy documents as shared context for the team. Caveat: it is for internal use; it is not a website chatbot by itself.
Microsoft Copilot Studio: builds assistants grounded in SharePoint or website content and publishes them to Teams or web chat. Caveat: setup takes more time and usually some technical help.
Intercom Fin: a customer-facing AI agent that answers from your help content and hands over to staff. Caveat: it makes most sense if you already use Intercom for customer support.
Data protection and GDPR
The policy documents are usually not personal data, but guest conversations are. Do not connect the assistant to booking systems or guest profiles at the start. Tell users they are talking to an AI and how messages are stored; the AI Act's transparency duties for chatbots phase in through 2026–2027. Set a retention period for chat logs and never ask for passport, card or health details in the chat. You may still be responsible for what the assistant tells customers about refunds or prices, so keep people in charge of exceptions. Check with your DPO or legal advisor before going public.
Mistakes to avoid
Connecting the assistant to outdated or contradictory documents.
Launching it publicly without a clear way to reach a person.
Patching wrong answers in the bot's instructions instead of fixing the source policy.