Back
29 Min
A 90% deflection rate sounds impressive until you ask what actually happened to the customer. Giulio Castiglione runs support for Playtomic across 36 countries and nine languages, and he refuses to treat that number as a win.
Giulio Castiglione, Customer Care Director at Playtomic, has a problem with the metric most support teams lead with. Deflection tells you how many tickets you kept away from a human. It says nothing about whether those people got their problem solved. On this episode of the Fini Podcast, he explained how he thinks about deflection vs resolution, and where he draws the line on what AI should and should not do.
Meet Giulio Castiglione
Giulio leads a 34-person support team at Playtomic, the booking platform for racket sports, covering 36 countries in nine languages. His starting point is unusual. Most leaders reach for AI to cut costs. Giulio started automating in 2022 for the opposite reason: to grow without adding headcount. Because he moved early, he treats AI as a way to handle more volume well, not a reason to shrink the team.
Deflection vs resolution: the number that lies
"If a vendor comes to me saying we have a deflection rate of 90%, for me there is no value in that message," Giulio said. Deflection counts how many tickets you kept away from a human, not how many people you actually helped. Make it harder to reach an agent and the rate climbs while nothing improves.
He gives a sharp example. When a real bug hits the product, raising deflection on purpose, by posting a clear message in the bot, is the right move to take pressure off agents. The rate goes up, but it tells you nothing about quality. Resolution is the harder, more honest target. The catch is to never confuse "deflected" with "resolved." When a case is hard, an abusive customer or an unusually complex request, the right move is a clean escalation to a human who can fix it. Asked which single metric he would keep, Giulio did not hesitate: resolution.
Knowing when to take the speed limit off
Giulio compares a powerful model to a Ferrari. The skill is no longer raw speed. It is knowing when to set the limit and when to take it off. A capable model can solve almost any request, including ones it should refuse, like a refund backed by a reason that is not true. So the human's job is shifting. Instead of AI supporting the human, the human increasingly supports the AI by setting the guardrails that make power safe to use.
The same instinct shows up in how his team measures quality. They moved off pure speed metrics like first response time and average handling time after noticing that fast answers were not better answers. Now they watch customer effort, such as how many questions it takes to resolve a ticket, and they keep humans on the cases that need judgment: vulnerable users, abuse, churn risk, and conversations that turn negative.
What support leaders should take from this
Strip away the analogies and a short playbook is left.
Check what "resolved" really means. Press vendors, and your own dashboards, on whether a deflected ticket actually solved the customer's problem. Do not reward a bot for hiding work.
Set the guardrails before you scale the power. Decide what the AI must never do on its own, like approving a refund on a weak reason, and keep a human on those calls.
Measure effort, not just speed. A high number of back-and-forth questions signals friction even when response times look good. The goal is scaling support without sacrificing quality, where volume rises but the bar on each interaction holds.
Treat languages as culture, not translation. Giulio's team hires a few native speakers per market to encode local tone, like the formal register German customers expect, then trains the model on it. That is how you cover many languages without a large local team everywhere.
Turn closed tickets into product fuel. Every resolved ticket gets an AI summary. Giulio mines those for patterns and hands product an impact document every two months, which is how support finally gets heard.
Pick the metric you want to move before you launch. His biggest warning: deploying AI just to cut people, with no quality target, is how CSAT slips. As he puts it, AI will not steal customer service jobs. It will replace the people who never learn to use it well.
Listen to the full episode
Giulio goes deeper in the full episode of the Fini Podcast, including how his team resets expectations now that customers compare support to ChatGPT. You can reach him on LinkedIn, where, fittingly, he replies with a low first response time.
Resolution over deflection is the principle Fini is built on. Book a demo to see it on your own tickets.
Leo: Welcome back to the Fini podcast. I'm Leo. My guest today is Giulio Castiglione, Customer Care Director at Playtomic. He leads a global support team of 34 people across 36 countries and nine languages. What's interesting about Giulio is how he thinks about AI and customer support. He's not chasing deflection rates or automation for its own sake. He's focused on the harder question: how do you scale support with AI without destroying the customer experience? Giulio, welcome to the show.
Giulio: Thanks, Leo, for the invitation, and welcome, everybody.
Leo: Let's start with a challenge every support leader is facing. Inbound volumes are growing, but headcount isn't keeping pace. How have you handled that gap without burning out your team?
Giulio: In our case the situation has been a bit different. We started looking at automated solutions a few years ago, around 2022, because we wanted to support our growth without needing to hire people. So we were in the opposite situation. We started with the first bots, the typical decision-tree ones, and moved to today's agentic solutions, which are much more efficient, always with the idea that hiring more people when you're growing doesn't make sense, you need a more scalable way. Doing that from the beginning made things easier, and we're not in a position where we need to cut people now, so we're in a safer place.
Leo: You started years ago, around when ChatGPT was emerging. How have you seen customer support AI evolve, and what big shifts have stood out?
Giulio: It's a huge change that's allowing customer experience to be more efficient. When we started, there was no way to have something help you write a message. There were macros, but you had to spend time choosing the right one. Then tools like ChatGPT started helping us create better messages. After that, AI started helping us analyze tickets. QA used to mean manually reviewing thousands of tickets a month; now an AI can read an entire conversation against your guidelines and provide a score automatically. So AI is really helping customer experience get better. It forces you to rethink some old behaviors, but if you control them, you can bring a lot of value to customers and make your life easier.
Leo: You've said there's a big difference between deflection and resolution. A lot of companies track deflection, but that doesn't tell you if the customer's problem was solved. How do you think about that distinction, and why does it matter when evaluating AI?
Giulio: It means a lot. If a vendor comes to me saying we have a 90% deflection rate, for me there's no value in that message, because deflection is like opening a bottle of water or not, you can choose your deflection rate. If you make escalation to an agent more difficult, your deflection rate grows, but that doesn't mean you're giving a great experience. Deflection mainly measures how many tickets pass by, regardless of the reason. For example, if you discover a big bug in your product, it's better to raise deflection by showing a clear message in the bot, because otherwise your agents can't handle the situation well and customers wait too long. That reduces pressure and raises deflection, not the best experience, but it's because of the bug. On the other side, the efficiency and quality of conversations is far more important. If the bot really solves more tickets, that matters. But it's not always easy to teach a bot how to solve certain situations, like an abusive customer, harassment, or generally complicated tasks. So the goal must be the highest resolution rate possible, but if you confuse deflected with resolved, it's better to allow an easy handover to an agent who can provide the best experience for those cases.
Leo: That relates to your Hall of Fame quote, that AI is like a Ferrari now, and the real differentiator is knowing when to set the speed limit and when to take it off. Can you explain what you mean by setting guardrails and letting AI run?
Giulio: My analogy related to a recent announcement about a new, very powerful model. It made me think about what dangerous means: it's dangerous because it can do things a human couldn't even think of. In the future an AI could reach a 100% resolution rate, but do we really want that? Say a customer bought a product online and tells the bot they want a refund, giving a reason that isn't real. A powerful enough model could solve that, but that's not good, we didn't want to refund that customer. So AI can be powerful, but the right guidelines and limitations make the difference, and there a human has the role. In the evolution of AI, it's not AI supporting the human anymore, it's increasingly the human supporting the AI, giving it the tools to become powerful by limiting it. That's why I made the Ferrari analogy: when the AI is a true Ferrari, the human needs to be the best pilot ever.
Leo: If a model comes out that's even better, do you think there's a day AI won't need a human in the loop, knowing these customers don't deserve the refund even though it could give it? Or will it always be the pilot for the driver?
Giulio: We should distinguish AI in the customer experience environment from AI in general. In customer support, probably yes it could, but there should always be a human to provide new guidelines, because products keep evolving and people keep evolving, so you have to let the AI evolve with them. Speaking generally, a fully independent AI scenario would scare me, and that would be the biggest problem, but I'm sure the people working on AI want to avoid it. In general I don't think it's going to be possible to have AI completely independent from humans.
Leo: A lot of teams optimize for speed first and add quality later. You've spoken about making quality a core principle. How did your team make that shift in practice?
Giulio: Early on we spent a lot of time on speed metrics, mainly first response time and average handle time. But we saw there's no real connection between a conversation answered quickly and one answered hours later, same for average handle time. Users want their problem solved in the easiest way possible, so we measure quality on that now. For example, how many questions are we asking the customer? If it's very high, that's a bad signal. We want to reduce the number of questions and detect the real problem using our internal tools. Now we're in a new phase: customers want a full resolution and they want it fast, because they're used to ChatGPT, which gives an answer in seconds without asking many questions. So we're learning again how to set and manage expectations. The journey starts with a ticket, then the bot, which needs to be quick and provide the best resolution, or quickly hand over to an agent who answers fast because we've managed ticket volume, using a copilot to generate an even better answer with a human checking it. AI supports both customers and agents at many touchpoints, and the more you synchronize each one, the better the quality and times you deliver.
Leo: You run support in nine languages across dozens of countries, a lot of complexity. How do you maintain consistency with so many moving parts?
Giulio: First, finding the right people. We have native people from each of the nine languages, a colleague from Portugal, one from Finland, and so on. What they bring is mainly cultural guidance, because it's not just about translating a ticket. If I'm Italian and receive a ticket in Portuguese, I can get it translated to English, answer in English, and the user receives it translated, so the language gap is covered. But a customer in Portugal writing Portuguese is not the same as a customer in Brazil, it's a matter of culture, not just language. So those local people provide the cultural aspect, and we also train AI, thanks to them, to translate following certain guidelines. For example, in Germany people communicate very formally, using the third person even in customer service, and using the second person can come across as too familiar. So we taught our model to use a certain tone when translating to German. Managing many languages doesn't mean hiring only locals, it means hiring a few to train people and AI models to align with the culture, so anyone can do the job and effectively become a German speaker, even if they're from somewhere else.
Leo: You use both human and AI-based QA, so AI is checking work that's increasingly AI-powered. What does one catch that the other misses?
Giulio: It's a fun way to see it, AI writing and AI judging, though they're different agents. AI can be a great QA agent, a great support agent, and even a coach, I personally use Claude as a coach, sometimes recording a meeting and asking how I communicated. On your question: with a big volume of tickets, only human QA gives you a partial picture, like CSAT, which I don't love because you might miss the ticket that would really help. QA gives you much more visibility, and you can train it to check many things accurately, even pointing to the exact sentence that wasn't good. With nine languages, human-only QA would need at least nine people per language, which isn't affordable. A human touch is always needed, so we do both: if QA flags a ticket as very bad, a human reviews it too, so the AI gets better analysis and the agent knows a human also read and evaluated it.
Leo: Where do you see humans in the loop, and where are you not comfortable with AI making the final call?
Giulio: Specific cases with vulnerable users, abuse, or harassment, where a human is better at catching certain behaviors. Also more complicated tickets, a customer mentioning churn risk, or one whose sentiment turns negative. In those cases a human is better at giving the right feedback. AI can still manage a lot of tickets, and honestly those sensitive ones aren't the most frequent, otherwise you'd have other problems. So for the rest, AI can be much better, and for those cases a human is better.
Leo: On Claude as a mentor, you're not alone, I use it day to day too. Now, on collaboration with other teams: many support teams say product doesn't listen to them, but you work directly with product and UX. How do you get product to listen?
Giulio: It's a battle for many colleagues. My first suggestion is to understand the language of the product team. I tell my team to imagine product speaks Chinese and no one on our team speaks it, so let's learn to speak Chinese. I spent a lot of time close to our product team to understand how they want to receive things. I learned that total volume isn't a measure they prioritize on, impact is. Between 1,000 people mentioning a problem and 20 people whose lives are made completely impossible by a problem, they prefer to fix the second. Another company might be the opposite. So we started talking about impact. On the other side, every time a ticket closes, an AI generates certain fields, one being a summary, like a customer who couldn't pay with a certain method, plus the interactions. Then I use Claude to find common patterns, not common issues, because common issues don't always show the real pain. The AI helps me group these into work streams of things making customers' lives harder. We decide which are most impactful, ask the AI follow-ups like which customers are affected, and create a document I share with product every two months. They receive it like a customer interview, with impact and description, and use it to manage the roadmap. When we agree on a point, it's win-win: fewer tickets for us, fewer problems for customers, and less time for product building the roadmap because we come with everything ready.
Leo: I love the point on product speaking Chinese. One of the most advanced things I've heard is a support leader who automated a workflow that fed product changes from support into a coding tool that built a mock-up of the new feature, reducing the gap to execution. One final question before rapid fire: if a support leader is told to implement fast, cut costs, and still maintain quality, what's your advice, and what should they do first?
Giulio: First, analyze your ticket volume and understand what part comes from repetitive tasks. That's the first step, and repetitive tasks often don't need much knowledge, sometimes not even AI. Once you've covered those, understand your biggest need: is volume still too high, is it the complexity of the tickets, where is the problem? Based on that, you start solving, improving the agents' work, improving the bot, or finding easier ways for customers to reach you. Once you understand the problem, that's your roadmap. On cost, it depends where the costs are, but if improving tickets drops your agents' occupancy rate very low, it gets complicated to keep the whole team, being honest. Luckily that's not our case, but people who are late to AI will face that.
Leo: Rapid fire. Where do you see AI and support heading in the next two to three years?
Giulio: Becoming the best helper ever and making life easier for people using AI. AI isn't going to steal customer service jobs. It's going to take the place of the people who aren't able to use AI as their best assistant.
Leo: If you could keep only one metric, deflection or resolution, which and why?
Giulio: It depends on the company, but if I can say anything, of course resolution. It's more fun to work on resolution.
Leo: The biggest mistake leaders make implementing AI in customer experience?
Giulio: Launching it without knowing which metric they want to improve. If you add AI just to cut people, you can forget about CSAT, and the experience gets much worse.
Leo: Giulio, this was great. Where's the best place for people to connect with you?
Giulio: LinkedIn, whenever you want, I always answer with a low first response time.
Leo: Perfect, Giulio. Thanks so much for joining us. And for everyone listening, if you want more honest conversations about implementing AI in customer support, make sure you subscribe to the Fini Podcast, and we'll see you next time.
Giulio: Thank you.
What is the difference between deflection and resolution?
Deflection counts how many tickets were kept away from a human agent. Resolution counts how many customers actually got their problem solved. Because deflection can be inflated by making it harder to reach a human, a high deflection rate does not prove a good experience. Resolution is the more honest measure of whether support worked.
Is a high deflection rate a bad thing?
Not always, but it is easy to game. Giulio Castiglione notes that during a product bug it can be right to raise deflection on purpose, by posting a clear message in the bot, to take pressure off agents. On its own the figure reveals nothing about quality, which is why he prefers to track resolution.
How do you scale AI support across many languages?
Treat it as a culture problem, not a translation problem. Hire a few native speakers per market to encode cultural rules, such as the formal tone German customers expect, then train your AI models on those rules so they can write like a local in each market.
