Featuring Delight.ai CEO John Kim.
The assumption sitting underneath most AI roadmaps in customer experience is that trust follows capability. Make the model accurate enough, fast enough, natural-sounding enough, and consumer hesitation takes care of itself. Budgets get built on that premise. So do the deployment timelines many of us are defending in front of a board right now.
The 2026 Delight AI Index, released in May, puts real pressure on it. Asked what builds confidence in an AI agent, consumers put the ability to reverse a mistake at the top — ahead of accuracy, ahead of transparency. 57% said being able to undo an AI’s action makes them trust it more. Separately, 83% said the brand is accountable when an AI interaction goes wrong, not the technology vendor behind it. John Kim, CEO of Delight.ai and co-founder and CEO of Sendbird, the company that powers the platform, reads those numbers as a correction to how the industry has framed the problem. What consumers are asking for isn’t a smarter agent. It’s a way out.
Sequence the Autonomy, Don’t Slow It Down
Only 34% of consumers in the same study said they’re comfortable with an AI agent taking action on their behalf. That’s a jarring number to sit next to a roadmap full of autonomous deployments, and the obvious reading is that brands should pump the brakes. Kim doesn’t read it that way.
“I don’t believe that leaders should see the gap as a reason to slow AI adoption, but they should be more intentional about where they introduce autonomy. Our Delight AI Index shows that consumers are comfortable with AI handling routine, low-risk tasks, but they still want people involved when the stakes are higher. The limiting factor isn’t the technology, but it’s the perceived consequence of getting it wrong. That should change the sequence of AI adoption, not the pace.”
Sequence rather than pace is a useful distinction for anyone building a 12-month plan, because it turns a go/no-go argument into a routing question. Which interactions carry low perceived consequence for the customer, and which ones carry high? A shipping status lookup and a disputed charge sit in completely different places on that scale, even though both land in the same queue. Kim’s framing suggests the roadmap should be ordered by consumer-perceived risk, not by ticket volume or by whichever integration was easiest to build. Every clean resolution, he argues, buys permission for the next one. Autonomy gets earned in that order or it doesn’t get earned at all.
The Cost Was Never in “What Time Do You Open?”
Most agentic CX deployments today are tuned for a single shape of question — one request, one system, one answer, resolved inside the session. Kim’s argument is that the industry optimized for the cheapest part of the workload and then congratulated itself on the containment rate.
“Now take a real one. ‘My flight got cancelled. I need a refund, and can you rebook me on the next available? If the layover’s overnight, I need my hotel extended too.’ That’s three requests, four systems (booking, airline, hotel, finance), one conditional that can’t be resolved until the rebooking lands, and at least one approval gate, because a refund crosses a money threshold. It also won’t finish inside the session. Some of it resolves over the next two hours, after the customer has closed the app.”
He calls this long-horizon work, and it breaks nearly every assumption baked into single-turn Q&A systems. The case has to survive past the session. It has to reach parties who don’t share a database. It has to know which decisions it isn’t allowed to make on its own. Kim’s sharpest point is the one about where money actually goes: “the cost has always been in the exceptions, and exceptions are long, multi-party and unpredictable by definition.” For a CMO, that reframes what the AI line item is supposed to buy. If your program is measured on deflection of routine questions, you’re measuring the part that was already cheap. The expensive, brand-damaging cases — the cancelled flight, the botched order, the refund nobody can find — are the ones that were left out precisely because they’re hard. That’s also where the reputational exposure lives.
Self-Correction Isn’t the Same as Self-Grading
There’s a reasonable objection to agents that catch and fix their own mistakes, and it’s the one a skeptical marketing leader raises first: an AI grading its own homework is not oversight. Kim agrees with the objection, actually — he just doesn’t think it describes what a well-designed system does.
“I wouldn’t trust an AI simply because it says it made the right decision. I would trust a system that’s designed to surface potential mistakes and continuously improve based on evidence. That’s a very different thing from AI acting as judge and jury… The question is how quickly it detects errors, how transparently it explains them and how reliably it learns from them. That is where trust is built.”
Detection speed, explanation quality, and correction rate are all things you can put on a dashboard and hold a vendor to. That matters more than it sounds, because it converts a philosophical argument about AI trustworthiness into a procurement conversation with actual criteria. On the customer side, Kim’s read of the Index data lands somewhere similar: trust comes from control rather than intelligence. Consumers want to know a decision can be corrected and that a person can step in when it counts. His warning to anyone treating that as a phase-two problem is blunt — accountability, he says, has to be built into the system from the start. Human-in-the-loop checkpoints, reversal paths, and clear decision rights are architecture, and architecture is expensive to retrofit after the agent is already live and talking to your customers.
The Next Trust Gap Is Between Machines
Something else is arriving at the same time, and it’s barely on most marketing agendas. Consumers are starting to run their own assistants — tools that book, dispute, and negotiate on their behalf. Those assistants will increasingly talk to yours.
“The next trust gap isn’t between people and people, it’s between AI and AI. As agents increasingly act on our behalf, transparency has to expand beyond the customer interface. AI agents need to be able to communicate what they’re doing, why they’re doing it, what data they’re accessing and where human approval is required.”
Standards like the Model Context Protocol are pushing in that direction, but Kim’s point is about design intent rather than plumbing. Brands have mostly built agents for efficiency and measured them on tickets closed. An agent that represents your brand to another company’s agent is doing something closer to negotiation, and it needs to be legible — able to state what it’s doing, on whose authority, and where it stops. That’s a different specification than “resolve more, faster,” and it probably belongs to marketing as much as to the service organization.
Which brings the whole thing back around to the measurement problem underneath it. Kim’s one prescription for the next twelve months is to “stop measuring AI by efficiency alone” and start measuring “whether customers understand, trust and are willing to hand more responsibility to it.” The Index found that only 16% of consumers are comfortable with full AI autonomy today, while 54% expect to feel differently within a year. That’s a genuinely narrow window, and it’s the kind of number that should reshape a scorecard.
Most AI dashboards in marketing organizations today can tell you containment rate and handle time. Very few can tell you whether customers are handing over more responsibility than they did last quarter. If Kim is right that autonomy is something consumers grant rather than something brands deploy, that second number is the one worth building. Start by asking your team what would even go into it.
