AI Responses in Detail
This article describes, step by step, what happens between the moment a customer sends a WhatsApp message and the moment they receive your assistant's reply. No magic formulas: just a well-briefed artificial intelligence whose decision triggers the right response and the right action on the Reepli side.
Reepli relies on a fast conversational artificial intelligence (around one second of latency per response), calibrated for short conversations in multiple languages. The model choice is centralized: you have nothing to configure.
The four-step pipeline
The sequence is simple: message received, context assembled, assistant decision, action.
1. Reception
When a customer sends you a WhatsApp message, it reaches Reepli within moments. The assistant identifies the contact (or creates a new one) and records the message in the conversation.
2. Building the context
Reepli then prepares a complete brief for the assistant. This brief contains:
| Block | Source | Role |
|---|---|---|
| Role | System | "You are the WhatsApp virtual assistant of a business. Reply in the chosen language." |
| Catalog | Your services | List of services, prices, durations. |
| FAQ | Your FAQ | Frequently asked questions and approved answers. |
| Rules | Your rules | Strict business constraints. |
| Free-form instructions | Your free-form instructions | Anything you have added. |
| Active behaviors | Scenarios page | List of situations the assistant can handle. |
| Availability | Your opening hours and appointments | For appointment booking behaviors. |
To this brief, Reepli adds the most recent exchanges in the conversation (usually the last 8 messages, alternating between customer and operator/assistant), then the current message.
3. The assistant's decision
The assistant reads everything and decides on the best possible response. Three cases:
- Direct response — a tailored text, sent as-is to your customer.
- Pre-approved Meta template — used when the 24-hour window is closed (reminders, follow-ups).
- No response — the assistant judges that it is better to send nothing (off-topic message, human intervention in progress).
In parallel, the assistant can trigger a concrete action: create an appointment, record a quote, advance a lead, hand off to a human.
4. Sending
Reepli sends the response to your customer via WhatsApp and displays it in your dashboard, in the same conversation. If an action was triggered, it is executed in parallel (an appointment appears in your calendar, a lead advances a step, etc.).
What the assistant sees — a concrete example
Imagine a Parisian hair salon. Sophie sends: "Hi, would you have any availability tomorrow morning?"
The brief sent to the assistant looks (in simplified form) like this:
You are the WhatsApp virtual assistant of Salon Émilie. Reply in French,
in a warm and professional manner.
Catalog:
- Women's cut: €45, 1 h
- Color: €80, 2 h
- Highlights: €120, 3 h
Address: 12 rue des Lilas, 75011 Paris.
Rules:
- Open Tuesday to Saturday, 9 a.m. to 7 p.m. Closed Sunday and Monday.
- Do not offer an appointment less than 24 h away without calling the salon.
- For a color or highlights, ask for a photo of the desired result.
Active behaviors:
- Welcome & qualification, Appointment booking, D-1 appointment reminder, Review request
Availability tomorrow morning:
- 9 a.m., 10 a.m., 11 a.m., 11:30 a.m. (free)
History:
[customer] Hello, I'd like some information about your services.
[bot] Hello! Gladly, tell me which service interests you.
[customer] Hi, would you have any availability tomorrow morning?
Seeing this context, the assistant will probably reply:
Hello Sophie! Tomorrow morning I have these slots left: 9 a.m., 10 a.m., 11 a.m. or 11:30 a.m. Which service and which time work for you?
Notice that the assistant:
- Chose the Appointment booking behavior (and not a simple welcome).
- Reused the exact availability that Reepli provided.
- Asked for the missing service before confirming — it knows one is needed to book an appointment.
- Has not yet created an appointment because the customer has not confirmed yet.
Self-evaluation
The assistant evaluates the relevance of its own response internally. Reepli uses this to:
- Trigger an automatic escalation when the assistant is unsure of itself on a sensitive behavior.
- Identify borderline cases to make troubleshooting easier.
You have nothing to configure: these checks run in the background.
Latency and cost
| Metric | Typical value |
|---|---|
| Response time (median) | around 1.2 seconds |
| Response time (slower cases) | around 3 seconds |
| Cost per response for you | included in your plan |
Your plan's monthly quota is expressed in messages (not in opaque technical units), which is more predictable.
Why no fine-tuning?
Some competing chatbots offer the option to "fine-tune a model on your data." Reepli made the opposite choice:
- No fine-tuning, but a rich context. Updating your context is instant; a fine-tune takes hours and costs a lot.
- No model trained per customer. All accounts use the same artificial intelligence, briefed differently.
- Automatic updates. When better technology becomes available, every Reepli account benefits from it immediately.
For a WhatsApp assistant that needs to answer questions about appointments, pricing and qualification, a good context is more effective than a specifically trained model. You keep control of the behavior by editing text; no need to wait for a re-training.
What about data security?
- Messages are stored on our infrastructure with encryption at rest.
- No personal data is shared with third parties beyond what is strictly necessary to generate the response.
- Our artificial intelligence provider does not retain your data for training (a "zero retention" policy).
For more details, see the Privacy page in your settings.
Advanced use cases (on request)
The Business plan allows:
- Custom behaviors built by the Reepli team for your industry.
- Connections to your third-party tools when an assistant decision is made.
- Multi-user for teams that share the Inbox.
Contact support to discuss these options.
Going further
- Chatbot overview
- Configuring scenarios
- AI logic and rules
- Variables and personalization
- Testing your chatbot
If you want to dig deeper (abnormal latency, third-party integration), contact support from the chat at the bottom right of the dashboard. The team can share additional details.