Back to blog
Getting started & practical4 min readAbdirahman Hersi, Founder

AI Agent Integrations: What to Connect First and What to Ask Before You Do

Calendar, customer records, online store and ticketing. How to prioritise connections, why reading should come before writing and which questions separate a real connection from a logo.

A booking flow connected to a calendar system
Getting started & practical

Start with the AI agent answering from your knowledge, and add connections only once you know which questions they should solve. Connect reading before writing, ask what is actually available for your platform and test each connection with real cases before customers meet it.

An AI agent can do a lot without being connected to anything. It can answer questions about opening hours, services, terms and delivery straight from the knowledge you have approved. But when a customer asks "can I book Thursday?" or "where is my order?", knowledge is not enough. The agent then needs to read or change data in another system. That is what an integration does.

This guide helps you prioritise connections, avoid the most common mistakes and ask the right questions before you promise customers anything.

Start with knowledge, add connections for a reason

Every connection is one more thing that can fail, needs upkeep and requires permissions. So do not add a connection because it exists, but because it solves a recurring question that knowledge cannot. A good test: can you finish the sentence "The connection lets the AI agent answer [question] without a person having to look it up"? If not, wait.

Four groups of connections

GroupWhat it enablesRead or writeWhat to check
Calendar and bookingShow available times, book, change and remindRead first, then writeThat the confirmation matches what is saved in the calendar
Customer records and leadsSave enquiries and route them correctlyMostly writeWhich fields are filled in and who owns the follow-up
Online store and deliveryOrder status and delivery informationMostly readThat the status is current and the AI does not guess
Ticketing, email and messagingHandover to the team with contextWriteThat the case lands with the right person and the right details

Reading before writing

Reading data is relatively harmless. If the AI agent fetches the wrong time, you can see it and fix it before anyone is affected. Writing data is different: a wrong booking, a changed order number or a faulty lead can reach real customers. Introduce writing only once reading works, limit what may be written and log what is written. Ask your team to review the first written records by hand.

What is actually available

A long list of logos says little. Availability depends on your platform, your plan and how the connection is configured, and some connections are set up by the vendor's team during onboarding rather than by you. Read the integrations page for an overview, but always confirm with the vendor what applies to your specific environment before you build a process around it. The same goes for phone and messaging channels, which depend on configuration.

Questions to ask before you connect

  • Is the connection available for our platform today, or is it planned or set up by your team?
  • Does it read, write, or both?
  • Which permissions does it need, and can we limit them?
  • What happens when the connection is down or answers wrongly? What does the customer see?
  • How is what the AI agent reads and writes logged, and who can review it?
  • Who owns the connection on our side, and who is alerted if something breaks?
  • How is personal data that passes through the connection handled?

Build in this order

  1. Knowledge. Get answers to recurring questions right. See the knowledge base.
  2. Handoff. Decide where cases go and what travels with them. See human handoff.
  3. The first read connection. For example available times or order status.
  4. The first write connection. For example creating a booking, with limits and review.
  5. More connections. One at a time, with a separate test for each.

Test each connection with real cases

Test the happy path, but also the awkward ones: a time that has just been taken, a customer who changes their mind, an order number that does not exist and a connection that responds slowly. Check that the AI agent says it cannot see something instead of guessing. How to test an AI agent before you launch it has a test plan, and how the booking flow works technically goes through booking step by step.

For online stores, handling delivery questions with an AI agent shows how to split order questions into what knowledge covers and what needs a connection or a person. Also read the vendor checklist if you are comparing several systems, and see solutions for how connections fit into different flows.

Frequently asked questions

Does an AI agent need integrations to work?

No. It can answer from your knowledge without any connections. Integrations are needed when it must read or change data, such as a booking, an order or a customer record.

Which connection should we start with?

The one that solves your most common question. For many it is the calendar or booking system, for online stores order data and for service businesses the customer register. Pick one, test it and build from there.

What is the difference between reading and writing data?

Reading means the AI agent fetches information, such as available times. Writing means it changes something, such as creating a booking. Writing has bigger consequences if something goes wrong and should come after reading works.

Is every integration on a vendor's list available straight away?

Not necessarily. Availability depends on platform, plan and configuration, and some connections are set up by the vendor's team. Always ask what applies to you before you build a process around a connection.

Ready to see it in action?

Book a demo