Valuable Read -> Multi-Service Client -- Accuracy Pressure Testing

I frequently observe agencies encountering significant challenges with multi-booking functionality in GHL, particularly when external CRMs are integrated into the workflow.

Recent discussions have highlighted recurring issues:

AI systems scheduling incorrect dates or times
System inconsistencies causing booking failures
Integration conflicts between platforms

Given these concerns, I conducted a comprehensive QA test with a client to stress-test our booking infrastructure.

The rationale is straightforward: While a failed email automation represents a minor setback, deploying AI into customer-facing booking operations demands absolute reliability. When your AI handles direct customer interactions, any failure becomes a critical revenue risk that inevitably leads to client churn.

---

Client Context: Nail Salon with over 35 services, their main system is Vagaro, they have 5 technicians and a varying number of chairs based on pedicure and manicures.

Booking Context: It's always important to understand how people are booking with a service. In this case, you have three types of booking;

Single person - single service: When someone simply wants to book a single service without add-ons.
Single person- multiple service: When someone wants to book more than one service back-to-back.
Multi-person (group booking): When someone wants to bring a girlfriend, daughter, the list goes on...

As part of the QA, we ran around 30 tests, 10 different conversations styles to pressure test the system. The results are below:

Group booking: Bot identifies multi-person requests and directs callers to self-book via SMS link (not handled over the call). - Worked each time.

Single person – single service: Works smoothly, tested successfully 10/10 times. Correct slot and calendar every time, straight through GHL, Google and Vagaro (as Vagaro only has read only API).

Single person – multiple services:

Tested ~26 times in total.
22 successful bookings, 4 failed (lag/delay issues).
Works well for up to 3 services.
Beyond 3 services, bot becomes sluggish and response delays increase. At 7 services, lag was significant. --- Note - It's unlikely 7 would ever occur, but 2-3 services is typical when considering add-ons.

Findings

System is particularly stable on single, simple booking on a multi-calendar service setup within the AI assistant.
Multi-service booking functionality IS consistent but degrades significantly past 3 services, so you may want to consider moving to a booking funnel via SMS if this is a use case for your clients.
Group bookings are handled using a nice funnel with a client welcome video, and instructions on how to book. The downside is that it requires manual user input. In contrast, you could do a handover, but if it's high rates of group booking, the AI assistant is essentially an unnecessary middle man.

For a more visible breakdown of the pressure tests, you can below the google doc below:

https://docs.google.com/spreadsheets/d/12AtL4eFXGV43qkQXqOH2H0PEJD7-JI0pEdWvOx2yp0U/edit?usp=sharing

General AI prompting tip:

Build the functionality first before anything else. Don't faff around with personality etc till you have the fundamental objective fulfilled.
Test individual item at a time, it's easier to identify the point of failure than building the entire car first than piece by piece.
Map the user journey first to handle expectation: For example I know group booking will always be an issue to handle without using a calendar manually from the user.
Ensure you list at least 10-15 variations of a potential use case: I like to understand the risk of failure before putting it out in the wild, as if it's accuracy is not there, I will need to simply use the AI to passover or instruct via SMS -> booking funnel.

Time Consumption:

When taking on service clients with a significant service size, you may want to factor this into your pricing given it DOES require more testing to ensure accuracy and whether you need to add alternate communication styles into the workflow (such as an SMS -> Funnel).

Assistable Improvement Areas: Albeit more a magic wand, rather than something not being done well...

Ways to handle a large number of services without destroying character limits; I did try via custom values. However, this seems to work sometimes in certain areas of the prompt and not others.

Guidance on a better solution for group booking. Although, I think this is more limited due to GHL and singular contact being assigned to a number. I did consider if there was a way to create a tool that could trigger a new contact being created in GHL but have not yet explored this.

Peace out,

Nate

0 comments