AI Concierge Across a Hotel Group — 3 Properties, 3 Channels, 4 Languages, 91.7% Handled by the Bot, Improved From Real Chats Every Day

The Client

A B2B distributor of industrial valves, pumps, and fittings, selling to contractors and factories — with the owner running it alone. No sales team, no back office. He personally answered every one of roughly 50 daily LINE chats: pricing questions, quotation requests, payment slips, shipment tracking.

What makes this business different from ordinary retail is that price is not attached to a product — it is attached to brand, model, size and standard together. One brand treats the standard as an option inside a single model; another splits the same standard into an entirely separate model.

And all of that knowledge lived in one person's head.

The Challenge: One Wrong Answer Costs More Than Ten Slow Ones

The difficulty stacked in four layers.

Layer 1 — everything landed on one person. Around 50 chats a day, on top of running the rest of the business. Every repeated question was time not spent closing deals.

Layer 2 — this is knowledge work, not lookup work. 3,260 items across 52 brands whose sizing standards are not one consistent system. The same question can have a different correct answer depending on the brand involved, so keyword matching against an FAQ table cannot answer it.

Layer 3 — what the bot says is a price. Most incoming questions are pricing questions, and customers act on the answer immediately. A bot that guesses one price wrong does more damage than a bot that says nothing — because it gets it wrong in the owner's own voice.

Layer 4 — customers do not type the spec. Many simply send a photo of a valve, with no brand and no model, and ask what it costs.

What it costs to leave alone

The owner stays the bottleneck on every deal. The business grows only as fast as one person can type. And the day he is ill or out on site is a day nobody answers customers at all.

The First Call That Changed the Whole Project

The deal closed against a simpler scope: a bot that answers basic questions and generates PDF quotations.

The first call — about 35 minutes — surfaced three layers of complexity that were not in the agreement:

1. •Linking brand, model, and spec-sheet files together, not just a price table
2. •Picking the correct sizing standard per brand, since brands do not follow the same logic
3. •The owner's real expectation — customers get an actual spec file to open, not a text description

Rather than stopping to rescope and rebuild from zero, the owner chose to keep testing the live bot himself, round after round, sending feedback continuously while the build adjusted incrementally — instead of waiting for a complete written spec before starting. A second call added shipment tracking and the handover manuals; a third, later, was about scope itself.

And this is the part worth being honest about: moving ahead while the scope was still soft shipped the work very fast, and ended with far more delivered than was sold — which leads to the conversation at the end of this study.

The Solution: Remove the Shortcut, Don't Forbid It

One idea holds the whole design together: if the bot still has a route to answer a price without querying the catalog, then no matter how strictly the rule forbids it, one day it will take that route. The answer is to take the route out of the system.

What the bot does

Key Technical Decisions

Built in 7 Days Against a 14-Day Plan

What Made It Sound Like the Owner, Not a Generic Script

Training on real chat history — 1,827 real conversations analysed in detail, down to noticing the owner used one particular sentence-ending particle over another by a ratio of roughly 19 to 1, which was the opposite of what the bot had been using.

Real past replies were then pulled into 15 of the most common question categories — pricing, quotations, stock availability, shipping cost, payment — as the bot's standard answers, along with a clean line for brands not carried and for invoice name changes.

There was also unglamorous data work nobody sees: re-categorising over 320 valve rows and rescuing another 398 rows stranded in an "other" bucket into their correct categories. A product in the wrong category is a product that cannot be found.

What Was Delivered

• •The LINE OA bot, searching the owner's own live price data
• •PDF quotation and invoice generation with document numbering and tax-ID validation
• •AI reading of product photos and payment slips
• •Spec-file delivery from a 71-brand library
• •Human handoff with automatic bot silencing
• •A daily pending-quotation reminder and an automatic old-file cleanup
• •A separate test channel from the live one, so changes can be tried without touching customers
• •Three Thai-language manuals — a 21-step testing guide, a guide to entering product data without breaking the bot, and a post-handover system-care guide

The Results

1. The bot handles 90% of routine questions

Around 50 chats a day used to land on one person who was also running everything else. The bot now takes 90% of that question volume; what remains are cases needing real judgment, which are handed over automatically.

What that actually means is not "time saved" — it is that the business no longer stops at one person's typing speed.

2. The catalog grew from 1,355 to 3,260 items and stayed live-searchable

The system started at 1,355 items across 14 brands and expanded in stages to 3,260 items across 52 brands.

Every price and spec the bot states is pulled from a live two-way sync with the owner's existing files — never guessed — so an answer is right even when the owner isn't watching the screen.

3. 1,389 spec files openable from a chat link with no login

The spec library holds 1,457 files across 71 brands. On the first import, 38% (549 files) saved successfully but with an empty model field, because the first-pass filename parser only recognised one naming pattern.

Three rounds of fixes brought the unmatched count down to 409 — mostly files that genuinely carry no model information, like cover pages and wiring diagrams. The result: 1,389 files a customer can open straight from chat. They get the real document, not a description of it.

Where the 90% Figure Actually Comes From

It was not measured by an automated tracking system after go-live — the project closed as a self-hosted handover before there was accumulated usage data. It comes from two sources:

One — analysis of real historical conversations. 1,827 conversations totalling 49,311 customer messages found that 12 of the 15 most common question categories, about 90% of message volume, already had answers the bot covered before handover. The remaining 3 (shipping status, spec requests, invoice name changes) were closed with ready-made answers before go-live.

Two — the test results. A prepared technical test suite passed 31 of 31, and the owner's own hands-on test on his phone passed 21 of 21 — covering direct model-name lookups, remembering repeat customers without re-asking, and keeping the discount ceiling confidential.

This is written out because a number with no stated origin does not deserve to be believed.

The Four Most Expensive Lessons

One — telling an AI to "check first" is not the same as removing the shortcut. The system originally let the AI decide for itself when a product "looked unfamiliar" and skip the real catalog lookup. Even with an explicit rule against it, it still slipped through occasionally. What worked was removing that shortcut completely. A rule in the instructions isn't enough — the system has to be built so the AI can't make that mistake, not just told not to.

Two — a bot's memory repeated a wrong answer that had already been fixed. During testing, the owner caught the bot repeating an answer that had been tested and corrected. Its memory had cached a "no data found" reply from an earlier round and repeated it without a fresh lookup. Fixed by banning memory for product answers and cutting the retained message count from 10 to 4. The wider lesson: mid-testing fixes need a fresh conversation each time, or you're debugging against the ghost of an old exchange.

Three — only a test on the real channel can be trusted. Testing through a simulator — sending messages straight into the system from a computer rather than a real phone — corrupted Thai characters into unreadable garbage, making the bot look broken when it was answering correctly on an actual phone. Whenever Thai text is involved, verify on the real channel.

Four — files imported doesn't mean files findable. "1,457 files imported successfully" was both true and useless, because 38% of them could not be found. Always check what percentage is actually findable, not just how many files were processed.

The Scope That Was Over-Delivered, and Then Sold

By closing time there was visibly more running than had been quoted — a handful of quoted modules had become 15 working features.

There were three options: cut the extra features, give them all away, or have the conversation directly.

The third was chosen — a written comparison showing exactly what sat inside the original price and what went beyond it, with three paths offered for the client to choose from.

The client chose to buy the whole bundle on monthly installments, and all 15 additional features stayed on. Nothing was cut.

Talking about scope plainly earned more work rather than costing a client.

Being Honest About What's Still Open

• •No long-term usage data. The owner chose to install and run the system himself rather than continue a monthly support contract, so there's no real monthly chat volume or measured cost saving after go-live. The 90% figure comes from the delivery-testing report, not 30-day-plus post-launch tracking.
• •Document numbering is not yet synced. The bot's quotation numbers and the numbers the owner had been issuing by hand have not been reconciled; that was deferred to launch day, and the client was told.
• •The unanswered-question log is not reconnected. After one architectural change, the component that records questions the bot could not answer was removed and has not been wired back — which means those questions are not currently being collected for review.