AI in customer service, and the tradeoff nobody says out loud: Doug Hanna of Kustomer

The honest question about AI in customer service is not whether the technology works. It is what you are willing to trade. Doug Hanna's framing: if satisfaction drops five points and cost drops 50%, most companies take that deal. If satisfaction drops five points and cost drops 8%, they do not.

Hanna is president and COO of Kustomer, an AI-native customer service platform. He is a returning guest, having previously appeared as COO of Grafana Labs, which he helped scale from 75 to over 1,000 people. Before that he was at Zendesk, and before that at Automattic, where he and Michael answered support tickets together.

Why he chose the application layer

Hanna knew he wanted to be in AI after Grafana, and looked across the stack: infrastructure, models, and applications.

He kept returning to the application layer, and his reasoning is worth noting for anyone making a similar bet. Combining AI capability with actual use cases and business processes was where he could see real differentiation and genuine impact on business results.

Customer service was a space he had worked in before, which made it a natural landing point.

The company that went into Meta at 300 and came out at 90

Kustomer's history is unusual: private company, acquired by Meta, then divested roughly two years before this recording. Internally the current iteration is called K2.

The number Hanna gives is the striking one. About 300 people going into Meta, about 90 coming out.

Which meant the first year was rebuilding to what he calls fighting weight. They went from around 90 to 140 or 150 quickly, and were around 200 at the time of recording.

His description of the difference between those two modes is a useful distinction for any operator staffing after a contraction. Getting to fighting weight means hiring enough to keep the lights on, keep customers happy, and make sure sales inquiries get answered. Past it, you are hiring ahead of the work: deciding which initiatives to staff and what new capabilities to add, and scaling intentionally.

He notes the space itself changed underneath the company. At the time of the acquisition, AI was not on the mainstream radar.

What an AI agent actually is

Hanna is careful to say he does not claim to be an AI expert, and is most steeped in what is happening in customer service specifically.

His working definition: AI agents are specialized programs a customer configures for particular tasks, as opposed to one generic model handling everything.

And the reason for the distinction is empirical rather than theoretical. Kustomer's customers found that specialized agents execute their tasks with higher quality, higher reliability, more predictability, and in ways that fit their business.

His examples: an agent set up to handle returns. One to handle travel bookings. One to handle exchanges for a specific product line.

The contrast he draws is with the generic question-and-answer bot, which he says is much of what customer service AI looks like today.

Do not boil the ocean

Michael's question is the hard one: moving from deflection, meaning a bot that pushes you toward an answer, to actual problem resolution, across an enormous range of industries.

Hanna's answer is to refuse the general problem. They encourage customers to work use case by use case, because there are a million things you could do and the technology, while good, is not perfect.

The numbers he cites give a sense of the range. He had recently spoken with someone whose company was seeing around 80% resolution, which he calls remarkable and not unheard of. His estimate of the average is closer to 40%.

And his framing of why even 40% matters is the business case in one sentence: you can deploy a technology that does not require a full-time person to maintain and it resolves 40% of your contacts.

The argument for specialization is about durability rather than volume. Better results, higher satisfaction, and resolutions that stick, rather than the customer returning ten minutes later asking for a human.

Humans are not always right either

Michael describes a badly behaved agent in a different domain: one that did not identify itself, fired off detailed questions, then cut people off mid-answer.

Hanna's response reframes the quality question in a way more operators should hear.

Customer service organizations have always run quality assurance, listening to calls and reading emails to check whether standards were followed and the right knowledge was applied. That exists, he points out, because of the inconvenient truth that humans are not always accurate. Everyone has been told something by a support representative that turned out not to be true.

So his caution is against overambition rather than against the technology. Push too far and satisfaction suffers.

Which produces the tradeoff at the heart of the episode. You have two levers, satisfaction and cost. If satisfaction drops five points and cost drops 50%, most companies take it. If satisfaction drops five points and cost drops 8%, they do not.

And the answer varies by how customer-centric the business is and what its model is. His illustration: a luxury hotel chain may decline a trade a budget chain would accept.

What he finds genuinely interesting about this moment is that companies are having candid discussions about quality and reliability against cost, measured honestly against the reality that people are also imperfect and need constant training and coaching.

On the technical side, the products are adding observability: what is the AI actually telling people, can you give it feedback that a different answer would have been better, and the ability to tune the hidden prompts covering company policies and tone.

The gradient between human and AI

Rather than a binary, Hanna describes a spectrum.

AI handles everything. Simple knowledge requests: the return policy, shipping speed, whether a date is available. Things a consumer could find themselves with research.

Humans handle everything. Due to system limitations, AI limitations, or customer preference. His example is booking his honeymoon, where a complex multi-country itinerary involved real conversation with a person who came to understand their preferences.

The mix, which he finds most interesting. Humans handle part, AI handles part, with handoffs both ways. Continuing the travel example: the humans work out where to go and roughly when, then hand off to AI to build the sightseeing itinerary for a particular city. Or an airline agent explains a change policy, then transfers you to the AI that can execute it.

Where it gets more interesting still is across channels. You call, get your policy clarified, and then receive an email with a link to an AI chat you can deal with in your own time.

His honest note: he knows customers doing this, and has not yet experienced a human-to-AI-and-back handoff himself as a consumer, and is curious how it will feel.

Friction, volume and the limits of the argument

Hanna's observation about what happens when service gets easier is one operators should plan for. As friction falls, people contact you more, because it is easy, in the same way that easy search increased how often people look things up. If the incremental cost is low, companies mostly do not mind.

But his most useful qualification is the one that runs against his own product category.

The last ten to fifteen years of online customer service focused on deflection and self-service, and some of what that produced is better than any conversational interface. His example is returning something to Amazon. It once required contacting a person. Now it takes seconds, with a choice of drop-off options.

He would take that over speaking to a human every time, and, he says, probably over speaking to an AI too, because it is purpose-built for exactly that job and works. He does not need an unstructured conversation to say he wants to return a book.

So AI is not the end of the story. It is a different approach that makes good service accessible to more companies. Amazon has teams working on the returns experience. A smaller retailer cannot.

What has changed about expectations

Asked what he has carried from Zendesk to Grafana to Kustomer, Hanna points at reliability.

Customer expectations around uptime and stability have risen substantially. People do not tolerate their internet going down or websites being unavailable. When a major cloud provider has an outage and part of the internet breaks for ten minutes, it is remarkable enough to be news, and otherwise things simply work.

The 5 things I took away from this conversation

1. Say the tradeoff out loud. Five satisfaction points for 50% of cost is a deal most companies take. Five points for 8% is not. Naming the exchange rate turns an argument about whether AI is good enough into a decision the business can actually make.

2. Specialized agents beat one general bot. Not for philosophical reasons, but because Kustomer's customers get higher quality and more predictable results from an agent built for returns than from a model asked to handle everything. Pick use cases rather than boiling the ocean.

3. Forty percent resolution with no dedicated headcount is already a strong result. Doug's point about the average being well below the best case is a useful corrective to the demos. The business case does not require the ceiling.

4. Sometimes the button is better than the conversation. His Amazon returns example is the one I keep thinking about. A purpose-built self-service flow beats both a human and a chatbot, which is worth remembering before conversational interfaces get applied to problems that were already solved.

5. Fighting weight and hiring ahead are different modes. Coming out of Meta at 90 people, the first job was keeping the lights on. Past that, staffing becomes about which initiatives to fund and what capabilities to add. Knowing which mode you are in changes what a hiring plan is for.

FAQ

How is AI used in customer service? Primarily through agents configured for specific tasks, such as processing returns, handling bookings or managing exchanges, rather than one general question-and-answer bot. Hanna's argument is that specialization produces higher quality, more predictable resolutions that customers do not have to escalate afterward.

What is agentic AI in customer service? Specialized programs that a company configures and tasks with completing something, as distinct from a model that answers questions. The distinction Hanna draws is between deflection, pointing a customer toward an answer, and resolution, actually completing the request.

What resolution rate can AI achieve in customer service? Hanna reports hearing figures around 80% from the strongest implementations and estimates the average is closer to 40%. He notes that even 40%, achieved without a dedicated person maintaining the system, is a substantial result.

How do you decide how far to automate customer service? By naming the tradeoff between customer satisfaction and cost explicitly. Hanna's framing is that most companies would accept a small satisfaction decline for a large cost reduction, and would not accept the same decline for a small one, with the answer depending on how customer-centric the business is.

Should AI replace self-service flows? Not necessarily. Hanna's own preference for well-built self-service, using Amazon returns as the example, is that a purpose-built flow can beat both a human and a conversational interface. The greater opportunity is companies that could never afford to build those flows themselves.

Also mentioned

• Kustomer, its acquisition by Meta and subsequent divestiture
• Grafana Labs, which Hanna helped scale from 75 to over 1,000 people
• Zendesk and Automattic, earlier stops in Hanna's customer service career
• Deflection versus resolution, the shift defining the current generation of support tools
• Observability and prompt tuning, the controls being built into AI support products

Listen to the full episode

Doug Hanna on Between Two COO's

Between Two COO's is hosted by Michael Koenig. Subscribe on Apple Podcasts, Spotify, or wherever you listen.