The builder lab
Outcomes
What you will build
This lab brings the pieces together into one working AI Employee: a custom tool that checks real open times on a calendar and books one, a second call that adds the person to the CRM as a contact, and the capability that tells the AI when to reach for each. One employee, two real connections, doing the thing instead of just talking about it.
You will build it in a sandbox account, where nothing is at stake and you can break things on purpose.
Model before you build
The build is the easy part. The modeling is where projects go sideways, so you do it first, on paper, before you connect anything.
Start with the hardest version of the client, not the simplest: a business with several locations and several people who each work across more than one of them. Map who books into which calendar, and where each new contact belongs. Then check that structure against the simplest case, a single location with one person, to be sure it holds at both ends. Solve the hardest case first and the easy ones are already handled; do it the other way around and you rebuild when the hard case arrives.
Settle the data contract
One decision shapes everything downstream: who keeps each record current. Is the calendar the source of truth for availability, or the CRM? Is a roster of people maintained by an API sync, or by hand? Name that contract out loud before you build, because changing it later means rebuilding on top of it. A data contract is not paperwork; it is the decision that keeps two systems from quietly disagreeing.
Build it in chunks
Do not wire the whole thing and hope. Bring it up one piece at a time:
- Get the availability call working on its own, outside the platform, until it returns real open times.
- Get the booking call working on its own, so a test booking actually lands on the calendar.
- Get the contact upsert working, matching on email the way you practiced.
- Turn each working call into a custom tool, then attach them to a capability that says when to use each.
- Run the whole flow as the AI Employee, and spot-check the first few live bookings yourself.
Confirm each piece alone before you chain them. Measure twice, cut once: on a build with this many moving parts, small steps are what keep the last one from collapsing the rest. When something misbehaves, the Explanation view shows you what the tool actually sent, so you fix the real cause instead of guessing.
Work the five steps above in a sandbox account. When the AI Employee books a real meeting and the contact appears in the CRM without you touching either system, you have built an agent that acts in the real world, which is the whole point of this path.
Knowledge Check
Three quick questions on the judgment behind the build: modeling, the data contract, and working in chunks.