twimco.Let’s talk ↗︎
Menu
Field notes / AI in practice

One AI agent for shopping, support and retention needs one rule for who may change it

Delight.ai's new Customer Agent handles shopping, booking, support and retention in one place, and its companion tool lets business teams change it in plain English. The single agent fixes a real problem. It also means marketing's new offer and support's refund rule now live in the same brain.

Editorial engraving of a bright open office where four colleagues from different teams gather around one shared logbook as one of them writes an entry.

Delight.ai, Sendbird's customer-agent business, launched Delight Customer Agent on November 9, 2026: one AI agent for shopping, booking, support and retention, plus Midas, which lets teams describe changes in natural language, test them against simulated customers and apply approved updates. One agent ends the start-over problem between departments. Before adopting any shared customer agent, decide who approves a change that touches another team's promise, test it on cases that cross departments, and keep one log of every change.

What Delight.ai launched on November 9

Delight.ai, the customer-agent business owned by messaging company Sendbird, announced two products on November 9. The first is Delight Customer Agent, a single AI agent meant to handle shopping, booking, support and retention. Instead of one bot for sales, another for support and a third for the loyalty team, the same agent carries the customer's identity, preferences and history across all of them, on chat, voice, email and messaging.

The second is Midas, a tool for changing that agent without filing an engineering ticket. A team describes what it wants in natural language. Midas turns the request into a process the agent can follow, tests the change against simulated customers, puts approved updates live, including new policies and offers, and reviews conversations to suggest what to improve next. Delight.ai's head of product says it means changes made "in hours" instead of days.

The release cites a Gartner survey, published November 4, finding that just 4% of customer service leaders have achieved the return they expected from AI initiatives. BJ's Wholesale Club is quoted as a customer: whether a member is browsing for a grill or asking about a renewal, "they're talking to the same agent."

What one agent fixes, and what it creates

The problem Delight.ai is describing is real. Customers hate repeating themselves. A shopper who asks the sales bot about a product, then the support bot about delivery, then the retention bot about cancelling, meets three strangers. One agent with one memory is a better experience, and it lets the business see the whole relationship.

It also moves a problem. When each department ran its own bot, each department's rules stayed in its own box. Marketing's offers lived in the marketing bot. Support's refund rules lived in the support bot. With one agent, they share a home. And with a tool that lets any team change that agent quickly, a change made for one purpose can show up in a conversation it was never meant for.

Picture a retention team adding a discount for members who say they want to cancel. Sensible on its own. But the same agent also answers a member asking, with some frustration, about a late refund, who mentions cancelling in passing. Now the discount appears there too. Nobody did anything wrong. The change simply crossed a line that used to be a wall.

When every team can change the agent in an afternoon, the question stops being how fast you can change it and becomes who said yes.

How the work moves

One change, three departments

  1. Marketing asks

    Offer 20% off to members who mention cancelling.

  2. The agent learns

    The retention flow now offers the discount on request.

  3. Support inherits

    A member asking about a late refund mentions cancelling and gets the offer too.

  4. Finance notices

    Discounts appear on accounts that were never at risk.

Illustrative scenario showing how one agent change can cross team lines; not a Delight.ai screenshot.
When every team can change the agent in an afternoon, the question stops being how fast you can change it and becomes who said yes.

Decide who approves which change

This isn't an argument against shared agents or self-service tools. It's an argument for deciding, before launch, which changes a team can make alone and which need a second pair of eyes from another department.

A simple split works for most businesses. Wording and tone within a team's own flows can be approved by that team. Anything that states a price, a deadline or a guarantee, changes a refund, return or cancellation rule, or lets the agent take an action with money attached, needs sign-off from the team that owns that promise. Our note on why a macro may not invent a promise sets out the same idea for support teams.

Delight.ai says its agent operates "within the company's boundaries, approvals and oversight," and Midas applies "approved" updates. That's the right design. The approval rules themselves still have to come from you, and they need to reflect how your departments actually share responsibility, not just your org chart.

Name a tiebreaker, too. Sooner or later marketing will want an offer that support thinks will cause complaints, or sales will want a promise operations can't keep. Decide now who settles that, so a disagreement doesn't turn into two teams quietly editing the same flow back and forth. In a smaller business it might be the founder or head of customer experience. In a larger one it's often whoever owns the customer relationship as a whole.

Test the changes that cross departments

Midas tests changes against simulated customers before real ones see them. Use that well. The most useful simulated customers aren't the ones the change was designed for. They're the ones in other departments' flows who might stumble into it.

  • The frustrated support case who mentions cancelling.
  • The new shopper who asks about a member price they aren't entitled to.
  • The returning customer whose history includes an old complaint.
  • The booking customer who asks something the agent was never trained on.

Ask every team that might be affected to contribute two or three test cases. Keep them in a shared set that every future change runs against, so the library grows with the agent. It takes an hour to start, and it catches the changes that look harmless in one department and expensive in another.

Decision guide

Who approves which change

ConsiderationTeam can approve aloneNeeds a second team to sign off
WordingTone, greetings, how a product is describedAnything that states a price, deadline or guarantee
OffersA discount limited to its own flow, with an end dateAn offer that can appear in support or account conversations
PoliciesClarifying an existing ruleChanging a refund, return or cancellation rule
ActionsLooking things up, booking an existing slotIssuing credits, refunds or account changes
Decision framework for a shared customer agent; adjust the lines to your own teams.

Keep one change log for one agent

The last piece is the simplest. If one agent serves every department, keep one record of every change to it: what changed, who asked, who approved, when it went live and how to undo it. When finance asks why discounts jumped last Tuesday, or support asks why the agent started promising next-day refunds, the answer should take minutes to find.

Make undoing a change as easy as making one. If a new offer misfires on a Friday evening, the person on call should be able to switch it off without waiting for the team that asked for it.

Review the log weekly at first, with someone from each department in the room. It's also the best way to see whether the shared agent is paying off. Gartner's 4% figure suggests most organizations can't yet show the return they hoped for. A clear record of what changed, and what happened after, is how you find out whether you're in the 4%. Our earlier look at how support agent vendors price conversations and resolutions is a useful companion when the cost question comes up.

If you're planning a customer-facing agent that spans teams, and want help with the approval rules, test cases and change records that keep it trustworthy, see our AI applications and agents work or get in touch.

Related work: Spin:Market

twimco.

We build alongside the people doing the work.
Bring us the problem you’re thinking about.

Related work

Spin:Market ↗︎