DEMOName, address, hours, brands and the first-visit discount are published facts (dascenter.ae, Google Maps). Every per-model price, stock level, slot, warranty and fee is SAMPLE data (as of ), pending the owner's confirmation. Nothing here is a DAS Tuning quote.
DAS Tuning (demo)
A customer chat agent that answers the common questions about tuning, parts and booking, and hands everything else to a person. Step one finds the answer in DAS Center's tables with fixed rules. Step two lets a small AI model (Llama 3.1 8B on Cloudflare Workers AI) word it naturally, and a check throws the wording away if it contains any number that is not in the table.
How this agent avoids making things up
- Step one, rules. It matches your message to one question from an approved list. If nothing matches well, a person takes it. Earlier messages fill in the car ("and stage 2?").
- A lookup reads the answer from a table: prices, parts, calendar, hours, policies or dyno results. A fixed sentence holds only those values, and the source chip names the row.
- Step two, wording. An AI model sees only that fixed sentence and your question, with orders to add nothing. It never sees the rest of the data.
- The check. Every number in the model's reply must appear in the row. One that does not, a model error, or the daily spending limit, and you get the fixed sentence instead. The chip says which you got.
- It never rules on legality, promises warranty cover or power, or diagnoses a fault. Those questions, and anything with no table row, skip the model and go to a person with the facts on record.
Drawn from the agent's own intent and tool list each time the page loads.
The full approved list
OUR RECOMMENDATION, not a measured result. Your WhatsApp export decides it.
Should an agent answer your WhatsApp by itself?
Probably not on day one. The demo above shows an agent can answer common questions from your tables. Letting it send every reply in your name is a separate decision with real costs.
The risks of a fully autonomous agent
- A wrong price or promise in your name. A customer screenshots "AED 2,400" or "100 hp guaranteed" and holds you to it. Our check blocks invented numbers and promise or legal words, but a check covers only what it is written for.
- Legal and warranty exposure. Remap legality and manufacturer warranty are unsettled in the sources we read. A reply that sounds like a ruling is a liability. The demo refuses and hands these to a person.
- Customers who want a person. A tuning job is a large spend on someone's car. Many buyers want to talk to the technician before they book.
- WhatsApp policy and bans. WhatsApp's Business Messaging Policy says you may contact people only if "they have given you their mobile phone number", that automated chats "must also have available prompt, clear, and direct escalation paths", and that WhatsApp "may limit or remove your access". Automating the ordinary WhatsApp app with unofficial tools puts your number at risk. The supported route is the WhatsApp Business Platform.
- Losing the relationship that sells tuning. Repeat customers and referrals come from the person behind the counter. A bot that answers everything removes the moment where you build trust.
- Data and privacy. Chats carry names, phone numbers, plates and photos. Every tool in the loop must keep them in your storage, and UAE data law expects consent you can prove.
The alternatives, from least effort and risk to most
| Option | What it does | Who sends | Effort | Risk |
| A. Saved answers | Quick replies in WhatsApp Business for hours, location, payment, the offer | You | Hours | Very low |
| B. Draft and approve | The agent drafts each reply from your tables, with its sources. You approve with one tap or edit | You, one tap | Days | Low |
| C. Triage and alerts | Labels every chat: price seeker, booking, high-value lead (exotic car, Stage 2 or 3), repeat customer, complaint. Alerts you on the valuable ones. Information seekers get a template or wait | You, on what matters | Days | Low |
| D. Part-autonomous | Sends replies by itself only for safe questions (hours, location, free slots). Everything else is drafted or handed off | Agent for safe ones | Weeks | Medium |
| E. Fully autonomous | Answers every chat by itself and hands off only what it cannot answer | Agent | Weeks, then ongoing checks | High |
Which one fits what you want
| If you want to... | Start with | Why |
| Save time today | A, then B | Garage chats tend to repeat a short list of questions (your export will show how short). Saved answers cost nothing, and drafts cut typing without giving up control. |
| Get alerts on the prominent clients and leave information seekers for later | C | The Ferrari owner asking about Stage 2 gets your attention in minutes. "What are your hours" gets a template. |
| Grow, when chats will take too much time | B + C now, D when the numbers show it is safe | Your approvals become the record that proves which questions the agent answers right. Move a question to D once its drafts go out unedited. |
| Scale to many chats without hiring | D for safe questions, C for the rest | Full autonomy (E) only after weeks of D with no wrong answers, and never for price promises, legality or warranty. |
Our recommendation: B plus C
Drafts you approve with one tap, plus triage that alerts you on high-value chats. You keep every reply in your name and your relationship with the customer, and you still save most of the typing. Nothing goes out without you, so a wrong number never reaches a customer. Each approval or edit is recorded, so the agent earns more autonomy one question at a time, with evidence.
What decides it: your WhatsApp export. It shows how many chats you get a week, what share are simple questions (candidates for templates or D), how many are high-value leads (the case for C), and how fast those got a reply. If 80% of chats are hours, location and price, A and D do most of the work. If the money is in a few exotic-car leads, C matters most.
WhatsApp policy quotes: whatsappbusiness.com/policy, read 2026-11-11. This is not legal advice.
Printable one-pager: should an agent answer your WhatsApp?What the tests cover (last run)
Make it yours: the WhatsApp step
What do your customers actually ask? Your own chats tell the agent what to say. This demo is grounded in the typical questions of a tuning garage, not in your business yet.
- Why. An agent that makes things up costs more than it saves. The questions in this demo are educated guesses. Your chat history holds the real list, the real wording in Arabic, English and Russian, and the answers your staff gave.
- Export, under your control. WhatsApp's own export makes a text file per chat, without media by default, and leaves out disappearing and view-once messages. You export your busiest chats.
- Kept in your storage. Names, phone numbers, plates and addresses are replaced before anything else happens.
- Consent. UAE data law asks that consent can be proven and withdrawn. Whether past chats need fresh consent is a question for your adviser; this is not legal advice. Default: strip identifiers, keep only the derived answers, delete the raw export after review, and add a consent line to the live bot.
- Extract. Customer messages only: the question, the car, the language, what staff answered, and whether it ended in a booking.
- Group and label. Similar questions are grouped by a program, then a person merges and names the groups.
- Write the answers. For each question: how often it comes up, the answer you approve, the table row it depends on, and when a person must take over. Anything your staff answered differently each time stays with a person.
- Prove it. Replay a held-back fifth of real questions and report how often the agent picks the right route and how often a number is wrong (target: never).
- What you get. Your ranked question list, about twenty ready answers, ten handoff rules, and the list of questions you have no written answer for.
Limits: it covers only the chats you export, it carries seasonal bias, and prices are only as fresh as your table.
ILLUSTRATION: invented shares, not measured from any chats
After the analysis this chart shows your real question mix, and the order of the agent's answers follows it.
Sources for the public facts used (price range, legal penalties, export steps) are in the research report behind this demo. No customer reviews or personal data are used.