How Bookable added an AI agent side panel to their product in 20 hours that was used by 45% of their customers in two and a half months

Summary

Challenge: Bookable’s engineering team spent 3 months and at least 140 hours building an AI agent side panel in-house with LangGraph, a custom MCP server, and CopilotKit. The agent was slow, expensive, and frequently failed to complete tasks, so was effectively unused by customers.

Solution: With A2A Net, an initial version of the Bookable agent was created from Bookable’s OpenAPI spec in 1 hour, integrated into Bookable’s web app in 2 days, and they spent at least 20 hours integrating it, or 14% of the in-house engineering effort. A2A Net’s agent harness and optimiser made it faster, cheaper, and successfully complete more tasks. A single 55-minute, $9 optimisation run improved its average rating from around 2/5 to 4/5.

Results: In the two and a half months since launch, the share of customers using the agent every week has grown from 5% to 45%, and customers that use the agent send an average of 10 messages a week, up from 3. Whilst it’s too early to say what effect this will have on revenue and other metrics, an AI agent side panel typically increases activation, conversion to paid, and retention by about 10% (e.g. Shopify Sidekick and Wix’s AI onboarding).

Challenge

Bookable’s software engineering team spent 3 months building an AI agent side panel for their product. Although it’s difficult to know the exact number of engineering hours spent on building it, GitHub commits estimate that at least 140 hours were spent on generating code for the MCP server, agent, and side panel. However, GitHub commits don’t include time spent on generating code that wasn’t committed, meetings, learning, etc.

The agent that Bookable built used LangGraph and a multi-agent architecture with specialist agents. For the agent’s tools, it used the Bookable MCP server which was made for the project. Bookable used CopilotKit for the side panel.

Despite the considerable amount of time and effort put into the agent and MCP server, the agent was slow, expensive, and frequently failed to complete tasks. Due to this, the agent was effectively unused by customers.

Solution

With A2A Net, an initial version of the Bookable agent was created in 1 hour, and integrating it into Bookable’s web app took 2 days. GitHub commits estimate at least 20 hours were spent on generating code for the side panel integration. That’s 14% of the initial engineering effort compared to building it in-house.

This was possible because: 1. A2A Net is a fully-managed platform, and 2. Bookable had an OpenAPI spec. Without an OpenAPI spec it would’ve required a small amount of engineering effort (1-40 hours) on Bookable’s behalf. Although an OpenAPI spec isn’t required to build an agent in A2A Net, it’s strongly recommended because: 1. updates to the API can be made available to the agent without having to update the agent’s code or an MCP server, 2. an MCP server can be automatically created from an OpenAPI spec.

What’s more, the Bookable agent that was created in A2A Net was faster, cheaper, and successfully completed more tasks. This was in large part due to the powerful agent harness that every agent built in A2A Net gets out of the box. It was also due to A2A Net’s optimiser. Once the initial version of the Bookable agent was created, it was tested and rated on a range of tasks. Those ratings were then used to optimise the prompt. The initial optimisation took 55 minutes, cost $9 in tokens, and improved the agent’s average rating from around 2/5 to 4/5. Without A2A Net, it would take the average AI Engineer days to make a similar improvement.

Results

Before A2A Net, the Bookable agent was effectively unused by customers. The Bookable agent built with A2A Net was made available to Bookable’s customers two and a half months from the time of writing. In that time, the number of customers that use the agent every week has gone from 5% to 45%. What’s more, customers that use the agent send an average of 10 messages a week, up from 3.

Whilst it’s too early to say what effect this will have on revenue and other metrics, an AI agent side panel typically increases activation, conversion to paid, and retention by about 10% (e.g. Shopify Sidekick and Wix’s AI onboarding) by helping customers get to value, and by making the product more valuable.

What’s Next

Since Bookable became a customer in July, A2A Net has shipped 21 major features like adding support for Excel and PDF files, scheduled chats, email as a channel, giving customers the ability to give feedback to chats, and more. That’s about two a week on average!

The next major features planned are improvements to the optimiser, including having it automatically optimise the agent’s prompt from customer feedback. Bookable have also built a Bookable Admin agent to help Bookable staff support their customers.