Bản chép lời đầy đủ của ClearGrid's Mohammad Al-Khalili on Building AI for a Trillion-Dollar Debt Market

purely from those two channels. Jamie: How complicated was the branching there? Mohammad: [07:00] Super complicated. It doesn't happen through templates, it's not "Dear first name." It's all AI-generated today, very personalised, context-aware. But even doing that, we couldn't reach what the legacy collection agencies were achieving. So we knew voice was a channel, and IVR wasn't enough, we needed to figure something out. I really didn't want to hire humans, not because I have anything against it, but because I wanted to build a tech company. My co-founder and I kept having this discussion, and we thought, okay, maybe we should do a voice AI. [08:00] Our first version of the voice AI was an absolute nightmare. Back then we didn't build our own, we used companies like VAPI to put something together. Those were great for sales, for pre-sales, small use cases. But they couldn't handle concurrency the way we wanted, and compliance became an issue, because you're technically handing that data to an AI somewhere in the US. So we realised we couldn't use all this amazing technology that OpenAI, Gemini and ElevenLabs built. It was simply non-compliant for banking. [09:00] So we had to figure out something on our own. We did a hybrid, we borrowed some, we built some, we partnered on others. Things like text-to-speech need a huge research team, that's not my problem to solve, so we partnered with someone who does it well and can host it locally, anywhere, so residency is sorted. Jamie: Why did you feel the compulsion to go the voice route at all? Was it just an expectation, because the legacy agencies do it? Mohammad: It was a lot of that. No matter how techy we are, we're not going to outsmart people who've been doing this for 40 years. There's a reason they do what they do, a lot we shouldn't adopt, but a lot they must be doing for good reason. Voice is a very powerful channel for debt collection. [10:00] And down the line we did start hiring human agents, because we realised AI can't do everything. ----------------------------------------------- [10:00] "WE GET PAID WHEN YOU GET PAID" Mohammad: That's exactly why I went with, we're not SaaS, we're not just an agency, what you care about is the outcome. I get paid once you get paid. It's not a subscription or a consumption fee. Whether I use AI, humans or aliens, that's my problem, I'll just make sure you collect your money in an ethical, compliant and dignified way. Jamie: And is that percentage-of-the-amount model the one that already existed in the agency world? Mohammad: Yes, and that's why we went with it, because it's what the institutions are already used to. It's already different enough for them to wrap their heads around, especially banks, where changing a policy that's been in place for a while is a nightmare. So we just went with what they have. ----------------------------------------------- [11:00] A SHORT HISTORY OF DEBT Jamie: How entrenched was that? Has debt collection changed much, or has it been the same forever? Mohammad: If you talk about debt itself, it actually allowed us as humans to innovate enormously across history. We started writing systematically because of it, some of the first scrolls in ancient Mesopotamia were records on clay tablets, and a lot of the maths we have today was thought up because of it. [12:00] The politics of slavery changed because of it at some point too. Jamie: Let's skip ahead a few thousand years. Mohammad: The beauty of what's happening in this region is that people are coming to the conclusion that collecting debt doesn't need to be harsh. It needs a good understanding of what's happening with the customer, and giving the borrower a dignified experience does not reduce your conversion rates. That epiphany came shortly before we started, thankfully, so it laid the right ground for us. [13:00] If we'd started 10 or 15 years ago, we wouldn't have succeeded, because you can't ask AI to make judgments it's not supposed to make. A lot is changing in the region around debt, how it's handled, how the borrower is viewed. ----------------------------------------------- [13:00] SOLVING FOR THE BORROWER Jamie: It's a good point, because historically the optimisation has only been for the collector, not the social impact on the borrower. Mohammad: Which brings me back to the lender-versus-borrower question. Disruption happens once you start thinking about solving for the borrower too. And the borrower has a lot of problems, payment is a problem, dispute handling is a problem, getting all their options upfront wasn't easy. Jamie: It nearly seemed like a UX problem, the extent to which optionality just didn't exist. There was no sense of agency for the borrower to say, this is how I can pay it back. Mohammad: [14:00] Exactly. So instead of acting only on behalf of the lender, we started giving the borrower the best options they can get upfront. Previously, even for a settlement that was already pre-approved, the borrower would have had to make numerous calls and talk to many people to get what was already approved for them. Now they can do it at the click of a button. Do what the borrower deserves from the beginning, help them, and by helping them you help the lender, rather than the idea that the only way to get something is to pressure them, which doesn't work. ----------------------------------------------- [14:30] WHO THE CUSTOMERS ARE Jamie: On the lender side, who's the customer for something like ClearGrid? Mohammad: [15:00] Right now we serve banks, fintechs like BNPLs, non-bank financial institutions, telcos, car rental companies. On the B2C side, anyone lending to a consumer. On the B2B side, anyone lending to a business. Basically anyone who's owed anything, as long as it's institutionalised and recorded, with enough proof. We don't do peer-to-peer, which you see in other markets. Jamie: Does the amount of context you have on the borrower vary by the type of

lender? A telco versus a BNPL, say. Mohammad: [16:00] We don't get a lot of visibility or context, we just get enough to be able to collect. And it depends on the service. In the full agency model, we get more. As a technology partner, it stays the bank's data and tools, someone on their side operates the machine. So in the agency model, it's not a lot, we just get what we need. ----------------------------------------------- [16:30] WHY AI IS THE FAIRER COLLECTOR Jamie: How much context do you need, and does it vary between lending products? Mohammad: [17:00] It does, and that's something that's changed. There used to be this thought that you'd treat people differently based on what they bought. There's still some truth to it, you can't treat someone who went into debt buying potatoes and tomatoes to survive the same way as someone who bought a designer bag. That context helps. But at the end of the day, it's not for private companies to decide who deserves empathy and who deserves to be treated with dignity. So we keep it standard. The nice thing about AI is that it really cannot make those judgments. It treats everyone as fairly as possible, because everyone has their own situation. Jamie: Can you talk through how it works, maybe on the agency side? Mohammad: [17:30] The process is simple. We receive the accounts, the outstanding balances, and we start reaching out to customers, understanding their problem, and if there's a solution we offer it. In many cases there isn't much we can offer. Jamie: And the initial outreach, is that an AI call, a message? Mohammad: [18:00] It depends. That's where it doesn't treat someone in debt over tomatoes and potatoes the same way as someone who bought the latest iPhone, the system doesn't speak to them with the same intensity. Even their entry points differ. Everyone gets a welcoming SMS and email. With some it continues on soft channels, with some it goes to AI calls, with some it goes straight to a human, though rarely. Usually the AI does the heavy lifting at the beginning and hands off to humans when needed. Jamie: So the AI's the hunter, and the humans are the gatherers. Mohammad: [19:00] Exactly. Debt is taboo, there's a lot of shame around it, it's very personal. The nice thing about AI is that it doesn't have an ego. If someone tries to insult the AI, it's not going to care, it just goes back to the topic, optimising for its end goal. It does have exit points, there's a certain number of times it can push, and there are times it just hands off to a human. A frustrated customer isn't necessarily going to pay, so why prolong it, just end the call or transfer it. Jamie: With customer support, the cases that end up with a human tend to be the hardest. Does that change who you hire when you do hire humans? Mohammad: [20:00] There are definitely cases that shouldn't be handed to AI yet, and I say yet, because I have a lot of faith in the technology. A vulnerable customer, for example, AI shouldn't handle that, and AI is actually better than humans at recognising when it's happening. As soon as we know, it shouldn't stay with the AI, you don't want it to say the wrong thing. Complicated settlements outside the guardrails should go to humans, and sometimes there are disputes we've never thought of, my roommate stole my credit card, that sort of thing. ----------------------------------------------- [20:40] COMMAND: A CRM BUILT FOR AI Jamie: Where's the margin better, the agency or the tech? Mohammad: Obviously the tech. We have Command, which is a CRM-like platform that can trigger email, SMS and AI calls, and manage the journeys for human calls. [22:00] The difference between it and a traditional CRM is that we built it with AI workflows in mind, we built it for AI. The number one user of the internet is going to be agents, so we made sure AI can operate it without clicking unnecessary buttons. It has its built-in dialler, built-in everything. We don't charge for it in the traditional way, it's consumption-based, not seats. You can buy it in full, or in parts, some clients just take the voice-AI component, some just want the CRM for their human agents. [23:00] We built Command because the problem we were solving internally was that the AI does something, the human does something, and neither has context on what the other did. So we wanted a platform built with that in mind first. Jamie: Fragmentation is one of the biggest issues right now, even with traditional CRMs from a sales perspective. Mohammad: [24:00] To do what we do, you'd have to destroy everything and build from scratch. It would have been ten times harder to take a CRM built for humans and bolt AI on, or vice versa. The reason we don't charge for the CRM is that the barrier to building one today is low. At one point we actually had to do that destruction ourselves, and the team rebuilt it in almost no time, with all the prior thinking we'd done. You could put university grads in a hackathon and they'd build you a great version, I built one myself with Replit and Claude Code. There's still scale and certain things you need an engineering team for, but the barriers to build are low. So where do CRMs go? They'll have to pivot towards something more vertical rather than horizontal. That's exactly what we did. Yes, we have a CRM, yes we have a voice-AI component, but I don't frame us as super-horizontal. ----------------------------------------------- [25:00] WHY HE WON'T BUILD A VOICE-AI COMPANY Mohammad: Building that voice layer, connecting speech-to-text, text-to-speech and the LLM, isn't much work anymore, it's become commoditised. Any company will be able to build it for themselves. The reason we don't position ourselves as a voice-AI company is, one, it's a race to the bottom, and two, more importantly, pickup rates have been dropping historically. I'm not interested in building for a channel that won't be where it is today. Back in the '50s and '60s,

pickup rates were 50 to 70 per cent, people felt honoured receiving a voice call, and if they missed one they'd apologise. [26:00] Then businesses started calling, then IVR bots, and now Gen Z don't really use it as a channel. Jamie: I'd rather do anything than pick up the phone. Mohammad: As voice AI gets cheaper and more commoditised, more calls will be AI calls, so people answering numbers they don't know won't be a thing. I think it'll drop to single digits. On top of that, Apple and others are bringing in call screening. [27:00] So voice as a channel is going to change, I'm not saying it'll die, but pickup rates from unsolicited, unrecognised numbers will drop to single digits. ----------------------------------------------- [27:00] AI DISCLOSURE Jamie: Tied to that is whether you need to disclose that something is voice AI. What's your thinking, and what's the current landscape? Mohammad: Things are changing, we're seeing more adoption of disclosure. For certain industries it's not voluntary, it's a must. [28:00] If you ask an AI whether it's a human or an AI, it's obligated to tell you. The question is whether it discloses at the beginning, the middle or the end, and we're seeing different types of adoption. I think in future AI will have to disclose, and after that it won't even have to, because everyone will just assume there's AI involved, it'll become the norm. And again, that's why I didn't want to build for voice as a channel, I'd have a hell of an anxiety every time there's an update or announcement. You can't build a business over a channel, especially going as vertical as we are. [29:00] Some horizontal players are doing amazing work and moving towards building foundational models, and others are simply building a wrapper. Jamie: What's the most effective channel then, email, SMS? Mohammad: As we stand today, voice is still the best. We don't know what it'll be in future, who would have predicted WhatsApp 70 years ago? ----------------------------------------------- [29:00] THE FUTURE OF CHANNELS Jamie: Give me a prediction. Mohammad: A lot of it will be augmented with personal assistants. Our cloud bots will handle our affairs, and the way they notify you will be, hey, you should know about this important thing in your life. [30:00] The objective will be about informing you, not barging you. I don't think any channel dies anytime soon, people just build tolerance, like with email, once upon a time people were honoured to receive an email and would forward it to all their friends. Those channels are still alive, just less effective. ----------------------------------------------- [30:00] CODE-SWITCHING AND BUILDING MULTILINGUAL Jamie: Something we didn't touch on is code-switching, the variety of dialects, switching between English and Arabic mid-sentence. How has the voice AI been designed for that? Mohammad: The region that tackled this before us was India, and they handled it pretty amazingly. That's where building something foundational helps. We're seeing beautiful models come out of the region. [31:00] There's interesting research, I think by Munsit, who trained a speech-to-text model on unstructured Arabic data, and even if it doesn't fully crack it, it's an interesting path. It's refreshing to see. Jamie: Does the AI hallucinate sometimes? Mohammad: Not hallucinate, it just says, I don't understand, can you please repeat. We do about 2.5 million calls a day. [32:00] Sometimes even humans can't understand what someone's saying. It's about how the LLM handles that, by asking the person to repeat or to elaborate. You might ask, why not just build a bilingual one? In our case it's a compliance thing, we like to keep the call to one language. You can move between languages, if you start in English and say you're more comfortable in Arabic, we can switch the agent. We allow filler words, like "ya'ani," but full switching between Arabic and English sentences gets messy, so we ask for elaboration to make sure we're not capturing the wrong thing. Jamie: In sales and marketing you could use that. Mohammad: [33:00] It's a different category. Those guys can do a lot with voice AI, we're always jealous of them, they're not held to the same standards. ----------------------------------------------- [33:00] THE PIVOT FROM EMPATHY TO DIGNITY Jamie: One thing I found really interesting when we chatted back in December was your shift in brand language. It was very much "empathy" originally, and over time you pivoted to a word you've used a few times today, "dignity." What precipitated that? Mohammad: A lot of things. When I was discussing it with my co-founders, we felt "empathy" sounded fancy when you're vibe-coding the first version of your startup. Jamie: It's very ChatGPT, washy-washy. Mohammad: You're not wrong. If I wanted to be an empathetic debt collector, I'd wave off the debt for everyone, and that doesn't work. [34:00] What I can do is give the borrower a dignified experience, follow the compliance guardrails, and not insult anyone. I'm a private company talking to a consumer who has rights and deserves to be protected, and there's a framework to protect them. I'm not here to judge them, no matter what I think their situation is. And that shift in brand did a lot more, especially with our human agents. When we said "empathy, empathy, empathy," it didn't register, they didn't know the next step. When we moved to "dignified," they understood their rights and the borrower's. It's a way of carrying yourself, as opposed to an end result. [35:00] Our CSAT scores actually went up, we're at 4.6 right now. We've been between 4.3 and 4.8. It's better than some Michelin-star restaurants. ----------------------------------------------- [35:00] THE ECONOMICS OF 2.5 MILLION CALLS A DAY Jamie: We mentioned 2.5 million calls a day. How expensive is a call? Mohammad: We worked hard to optimise that, which is why we didn't always adopt the best models,

we chose what makes sense. As long as it does the job, we're good. [36:00] We had to train a lot, and use some open-source models. Word error rate makes a difference, the voice environment is heavy, someone's in an elevator, and so on. But instead we focused more on the LLM side, to understand that you're not receiving perfect information. Sometimes within the same call the models switch. [37:00] If it gets into a compliance matter, latency doesn't matter, what matters is accuracy, so we use higher reasoning. When it does mathematical calculations, the layer that handles that is different. It's about conversation design, so the borrower doesn't feel any of it. Jamie: On conversation design, what thought went into the voices, things like gender? Mohammad: Our data-science team found that gender matters at different stages, and accents matter in some instances. [38:00] A lot of that started to die out as we had to disclose it's an AI agent. The vibe at the beginning was to mimic a human as much as possible, that horrible ChatGPT experience where you think a company hired someone who's not very good. The moment you disclose it's an AI, the customer's forgiveness changes, if it glitches, it's okay, it's meant to glitch. So you're no longer racing towards the best word-error rates and latencies, you're racing towards the right solution. Jamie: And it's fundamentally different to sales, where someone's selling you something. The impersonality is almost helpful here, because you're being vulnerable, you can be a bit more open. Mohammad: [39:00] In so many cases we see that, especially with lower-value tickets, you can resolve it just by understanding your options. Customers tend to prefer doing it via AI, and AI is correct far more often than humans in many instances. My problem with what's happening in the industry is that it capitalises on the AI not disclosing that it's an AI. [40:00] The question is whether those companies can build something durable post-disclosure, because disclosure is coming. Otherwise, what are you trying to do, manipulate the customer into thinking they're speaking to a human? What kind of company has that strategy? It's a red flag. ----------------------------------------------- [40:00] THE ANTHROPOLOGIST IN RESIDENCE Jamie: You have an anthropologist in residence. Not many startups in the UAE do. Talk to me about that. Mohammad: It started socially, she's a friend, and she was studying debt collection, so it was a match made in heaven. That's why I have a good understanding of the historical background of debt. [41:00] Anthropologists tend to make amazing UX designers, there are many examples, IBM had one in situ for something like 18 years, more in the Valley than here. Our conversation design, and even that maturity of understanding that voice is a channel that will fade and not one to bet everything on, came partly out of those conversations. We were very lucky to have her. ----------------------------------------------- [41:30] IS THERE AN EXISTENTIAL THREAT? Jamie: What's the existential threat to ClearGrid? With AI underwriting and credit scoring getting better, could the volumes you handle decrease? Mohammad: [42:00] I think underwriting is getting better, but I don't think it impacts us adversely. Better underwriting actually helps us, because the people who enter due to bad underwriting are the ones who don't recover anyway. So better underwriting means higher recovery rates, it just filters who comes through. And there will always be defaults, no matter how good the underwriting, things happen in people's lives, someone loses their job, can't pay their mortgage or their car. [43:00] That's okay, and the question is what we do after. So I don't think there's an existential threat. Debt is very old, it's not going away, and it's not bad. The interesting change is that it'll be seen as a phase in someone's life rather than something they're branded with forever. ----------------------------------------------- [43:00] WHAT KEEPS HIM UP AT NIGHT Jamie: What keeps you up at night? Mohammad: Building. The work it takes to build nowadays is underrated. You're constantly trying to understand the latest model and what's happening in AI, while navigating a very difficult problem to solve. [44:00] Weeks feel like years. So there's a lot, but I'd say it's mostly just keeping up with everything, and it's exciting, I'm up because of excitement, not fear. There was a phase when we thought voice AI could be the thing, and what kept me up then was worrying about ElevenLabs' next update, or OpenAI, or Claude, even Microsoft launching a Copilot in Word. The foundation model eating your lunch. [45:00] The moment we decided to go vertical rather than horizontal, voice AI is a channel, it's a feature, not the product, that's when I could sleep again. ----------------------------------------------- [45:00] "NO BIGGER PROBLEM WORTH SOLVING" Jamie: What's the experience of building the team been like? It's a very different industry to what you or Mohammed operated in before. What is it about commerce and delivery founders going into fintech? Mohammad: We didn't think of it as getting into fintech. We both had our reasons. Mohammed started just before me, and he told me he was building this thing. As I said, debt was very personal to me, so at first I was like, no way. [46:00] Then I simmered on it and thought, there's really no bigger problem worth solving for me in my life. The idea started very differently, we were looking at building a borrower portal, self-serve, giving payment options, which we still use today. Then we thought about solving for payments, paying back debt has a lot of layers, if you're out of the country, or your bank accounts are frozen. [47:00] But we realised that would need three years of licensing, so we pivoted towards the agency model. Once we got into the agency, we understood what the market

needs, and from there we started building. Jamie: It's an interesting approach, starting the agency, getting down and dirty, rather than just interviewing a few agencies. Mohammad: We both reached that conclusion. For me it came from my experience right before ClearGrid, working at Livspace, alongside Anuj and Ramakant, which was deeply inspirational. [48:00] I wanted to learn what they did to build an almost-$2-billion company, the secret sauce. What I realised is that the eagerness to solve a problem matters more than how. I'd come from tech companies where all we wanted was to solve a problem our way. What I learned from them is that whether you do it with technology or with humans, your job is to get the output done in the best way possible. That translated directly into ClearGrid. The other thing is the power of having your customer as part of your team. The agency, Allocate, is a client of Command, so when we innovate something, my best product managers are actually my debt collection agents. [49:00] Our first debt collection agent went from agent to growth hacker to product person, and we have another in the making. ----------------------------------------------- [49:00] HOW AI IS CHANGING WHO GETS TO BUILD Jamie: That's increasingly a phenomenon, engineers now have so much more agency to just get into it and create the product, the barrier is so low. Mohammad: A lot is shifting. I get this question often, the number of people who walk into my office afraid AI will replace them or make them obsolete is insane. The first time it happened I was shocked, but as it increased I realised it's real. A lot of jobs will change and evolve. [50:00] But the interesting thing I'm seeing firsthand is how many people could have been a much better something, they just didn't have the right chance or skill at the right time. A lot of our debt collection agents never graduated from university, they have one or two years of call-centre experience, and they were able to vibe-code some of the best features we have today. Jamie: It's an amazing democratiser and leveller. Mohammad: It changed a lot, including how we hire designers. [51:00] Previously a UI/UX designer would receive a brief, build it, ship it, done. Now the focus is on building a very good design system that people might break when they're vibe-coding, so anyone vibe-coding goes back to the UI/UX person, these components didn't make sense, and it's constant iteration. Then it goes to engineering, to production. We're not vibe-coding and clicking publish, it goes to engineering and they take it from there. Jamie: Any idea how much of the code being written is AI-generated? Mohammad: [52:00] We have an AI measuring that. Right now 70 per cent of our code is written by AI. It's not impressive compared to globally, partly because some components are unmeasurable, and partly because for security reasons there are parts that have to be completely isolated. When I talk to friends in the Bay Area, they're looking at 90 to 95 per cent. We use a product called WorkWeave to track it, they have a leaderboard showing where you are globally on AI adoption and shipping velocity. ----------------------------------------------- [53:00] WHAT'S NEXT: BUYING THE DEBT AND GOING GLOBAL Jamie: What's the tool of choice for coding on the team, Codex, Claude Code, Cursor? Mohammad: The tool actually tells us who's using what. I thought Claude Code would be more popular among engineers, but somehow they still default to Cursor, so the majority are on Cursor and the rest on Claude Code. Jamie: How big is the team? Mohammad: Given we're technically kind of a call centre, and we're dealing with close to [a very large sum] of money, we're around 85 people, including the agents. Jamie: What's next on the roadmap? Mohammad: [54:00] Doubling down on securitisation, debt purchase is coming next. That's where we really solve the problem, and it opens an interesting market for both investors and lenders. Building and servicing that financial product will be interesting, and a lot of learnings will come out of it. The question is what happens when a tech company, with the way we think about giving the customer the right options, actually owns the debt and decides what to do with it. [55:00] Then geographical expansion. We're in four countries, mostly the UAE and Saudi. I think there's expansion beyond MENA too. We've built a category, we're not really digital collections, not really an agency, and there's something there with space to expand globally. Jamie: Was that the ambition originally, or did it develop? Mohammad: It developed over time. When you start as a founder in this region, something gets stuck in your mind that you have to build only for the region. [55:30] That unspoken constraint, that you couldn't possibly build a category-defining product globally. I do believe we've built something at ClearGrid that could go global. We've launched countries without setting foot in them, on the technology side, because the agency side is very complicated, you need central bank approvals, licences, NOCs from every country. [56:00] But with Command I see a global play, especially with AI. And unlike a lot of companies in the UK, US or Australia, building in Dubai forced us to think in all languages and accents. The way we solved for Urdu, some Hindi, was almost as much work as Arabic. I want to talk to you in your own language and solve your problem. [57:00] We've done that for both AI and humans, so even when a human agent doesn't speak a language, the co-pilot on the live transcript can translate. Jamie: Dubai forces that on you, it's such a melting pot, a crossroads of so many cultures. Mohammad: We were lucky to be in the right place at the right time. The amount of learning we have, we can apply across any country, region or problem. [58:00] Even solving for the move between the UAE and Saudi consumer, the delta in learning there is enormous. Jamie: That's always underestimated externally, that the two are the same. They're not. Mohammad: Not at all the same thing. ----------------------------------------------- [58:00] THE RAISE, AND FIRST PRINCIPLES Jamie: What are you most excited about over the next couple of months? You announced your raise about a year and a bit ago, the pre-seed and seed together. Was that because you were in stealth? Mohammad: Exactly, we were in stealth, so it made sense to make the most of the announcements. [59:00] We're going to keep building. We didn't expect this level of growth, so we're taking it one step at a time. The objective is simple, build a profitable, stable business, and solve this problem in the best way possible. The debt purchase is big, I'm very excited about that, there'll be tons of learnings and innovation, and some interesting funding structures behind it. It was first-principles thinking. Why is the lender trying to give all their debts to a collection agent? They're trying to resolve it, they want it off their books. So, what if I just take it off their books? When we started Command, it cannibalised the agency model, and this will probably do the same, so it'll be interesting to see how all three parties interact. Small fish eating the bigger fish, and eventually the big fish. Jamie: This was tons of fun. Mohammad: Thank you, I really enjoyed it. Jamie: We'll call it there.