Midas is an AI agent in Delight.ai that builds and improves your Delight Customer Agent through conversation. Tell Midas what you want your agent to do. It writes the change into an Actionbook, shows it as a flowchart, tests it with simulated customers, and prepares it for your team’s approval.

Can Midas build a new agent, or only improve an existing one?

Both. Midas can build a new agent from the policies, procedures, and tools you already have, and it can improve the agent you already run by updating its Actionbooks.

Does Midas change my agent on its own?

No. Midas proposes changes, including what to improve next based on your customers’ conversations. You decide who can make and approve changes, and nothing goes live until someone with approval rights says yes.

What is an Actionbook?

An Actionbook is the step-by-step process your Delight agent follows for one kind of conversation, such as a return, a booking, or a promotion. It holds the rules, the tools the agent can use, and the points where it hands off to a person in Delight Desk or your current helpdesk. Midas writes and updates Actionbooks for you.

Midas is part of Delight.ai and is rolling out to customers in stages. Get a demo to see it on one of your own processes, or ask your account team about access.

Introducing Midas: Build and Improve Your AI Agent

Your team’s understanding of customers grows with every conversation. Turning that knowledge into a better AI agent should be part of their everyday work.

Today we’re introducing Midas, an AI agent in Delight.ai that builds and improves your Delight Customer Agent. Your team describes what it needs. Midas turns that direction into instructions, tests how the agent responds, and prepares the changes for approval.

Give the people who know your business a way to shape its AI

Your CX, operations, and commercial teams know how your business should serve customers. Midas lets them turn that knowledge into how your agent behaves, through conversation. With Midas, your team can:

Build an agent from your policies, procedures, and tools.
Test how it responds before a change reaches customers.
Improve its behavior by describing what you need.
Analyze results and customer conversations to decide what to improve next.

Your team decides what goes live.

Build: Create an AI agent through conversation

Describe what your agent should do and give Midas the policies, procedures, and tools it needs. Midas proposes a plan, and your team approves it before anything is built.

Where your guidance leaves room for interpretation, Midas asks questions and recommends an approach, so your team makes the decisions that shape how the agent serves customers.

Midas turns those decisions into Actionbooks, the instructions your agent follows, and maps each process as a flowchart. Your team can see how the agent will work, refine it, and get it ready for testing.

Three steps in Midas: add your files (a returns policy, a delivery FAQ and an order lookup tool), a four-step plan for you to approve, and Midas mapping the new delivery and returns agent as a flowchart that is ready to test.

Test: Know how your agent will respond before it goes live

Giving more people the ability to change an agent also means giving them a way to judge the work.

Midas tests proposed changes through simulated customer conversations, checking the agent’s responses against your rules. When a test fails, it investigates the cause, updates the instructions, and runs the tests again.

Your team can also step into the conversation. Open the tester, write as a customer would, and see how the agent answers before anything goes live.

Midas runs five simulated customers against the new returns agent and all five pass. In the tester, you play a customer whose sofa was delivered 40 days ago, and the agent explains the 30-day return window and offers to connect them with the team.

Improve: Make improvement part of running the business

An agent’s launch is the start of an ongoing responsibility. As the business changes and your team learns more from customers, the agent needs to develop with it.

Midas connects that work. Your team can describe a change, review the updated instructions, and test the result in the same conversation. Ask Midas to move the holiday delivery cutoff to December 15, and it updates the Actionbook, tests it again, and shows the new version next to the one that’s live.

Every version is kept, and your live agent keeps running as it is until the new version is approved.

You ask Midas to move the holiday delivery cutoff to Dec 15. Midas updates the Delivery questions Actionbook to version 4, retests it, and a side-by-side comparison highlights Dec 20 changed to Dec 15.

Analyze: Find the next improvement in your customers’ conversations

Once a change is live, ask Midas how it went. Midas reads results and conversations together, so every number comes with the reason behind it.

It points to the questions and moments where customers stop, ranked by how often they come up, and turns each finding into a proposal for the next round. When you’re ready, Midas drafts the change, and it goes through the same tests and review as every other change.

Midas shows results for the Delivery questions Actionbook: 3,240 conversations, 412 orders, 12.7% conversion. “Is delivery included?” came up in 36% of conversations, and those shoppers bought at 7% against 16% for everyone else, so Midas proposes including delivery in the next promotion offer.

Govern: Control who can act, what changes, and what goes live

Opening this work to more of your team calls for clear controls. Midas gives you control at three levels.

Who can act. Decide who can build and edit, who can suggest changes, and who can approve them.
What changes. Every change starts in Development, away from your live agent. Every version is kept, so you can see exactly what changed and compare it with the one before.
What goes live. A change goes live only after the plan is approved, the tests pass, and someone with approval rights says yes.

Once a change is live, Trust OS guardrails keep the agent within the limits you set in every conversation. For how we think about giving an agent more autonomy, read The Trust Stack.

Three controls in Midas. Who can act: the support lead can edit, the ops manager can edit and approve, and marketing can suggest. What changes: version 4 is a draft and version 3 is live, with a compare link. What goes live: plan approved and 5 of 5 tests passed, waiting for approval.

What changes for the people who build agents

Midas also changes the work for the people who set up agents today. Brendan Corr builds agents with our customers as a forward deployed engineer.

“We used to translate every request into an Actionbook, configure it, write test cases, and go back and forth on the details. Now Midas drafts the Actionbook and the tests from the customer’s SOPs, and I review them. A new use case that used to take two weeks now goes live in a day.”

Brendan Corr, Forward Deployed Engineer, Delight.ai

A small step toward a larger ambition

We want improving an AI agent to become a normal part of running a business, as routine as updating a policy or planning a promotion. That means an agent that changes as quickly as the business does, shaped every day by the people who talk to your customers, with every change tested and every release in your hands.

Midas is a step in that direction. It comes from a simple belief: if you understand the business, you should be able to shape its AI.