The Blind Spot First: Fixing the Revenue Bottleneck Before Adding AI with Achilleas Kasimidis, GoStudent

John Emmerson: Hello, and welcome to another edition of the aiblLIVE Leadership Series. I am delighted to be joined today by Achilleas Kasimidis, who is Global Director of Rev Ops at GoStudent. Lovely to have you, Achilleas.

Achilleas Kasimidis: Lovely to be here with you as well.

John Emmerson: Thank you very much. Achilleas and I first met earlier this year, when he agreed to come and speak at one of our Leadership Series events over the summer. It was a fascinating talk, and your approach to adopting AI is really interesting. We are going to dig into that shortly. But before we do, why not introduce yourself and tell us a little about what GoStudent does.

Achilleas Kasimidis: Thanks a lot, John. I also enjoyed my time at the event. It was in July, right?

John Emmerson: I think so, yeah.

Achilleas Kasimidis: It was almost a full day, well spent with the teams there. I truly enjoyed it.

Now, about myself. I run the revenue team commercially at GoStudent, which is basically all things around revenue. Customer acquisition, retention, failed payments, anything that generates revenue for us.

What we do at GoStudent is an online tutoring platform. It is a managed marketplace, where we connect students and parents on the consumer side of the business with tutors, who are the ones providing the service. We are present right now in more than 10 countries all over Europe. The core markets are Austria, Germany, Italy, France, Spain, and the UK, alongside Poland, Turkey, Greece, and the Netherlands.

At the same time, we have been expanding quite aggressively lately in the offline space. We are setting up offline tutoring centres in Spain and Italy, and we are aiming to grow more through offline expansion, which we will probably double down on in the next few months.

John Emmerson: It sounds like it is going really well for you. I know you have been on an incredible growth trajectory over the past few years, so long may that continue.

One of the things I loved when you spoke at growthLIVE earlier in the year was that you are not from a coding background. Not engineering, not development. And yet you have built some incredible AI agents to help you do what you do. Your argument at growthLIVE was, do not start with AI, start with your blind spot. So what do you mean by that? And why do you think most businesses get it the wrong way around?

Achilleas Kasimidis: It has been what, two, three, four years now since the first LLM models came out. Since then they have become smarter, faster, and cheaper. A lot has changed. With those tools, anyone, including myself with no coding background, can essentially start coding via the agents.

I am of the opinion that in this new AI setup, whoever only knows how to use AI tools will not go too far. Those who know how to use AI tools and at the same time apply strategic thinking before building will probably win in the long run. I have seen that with my team, and with different people inside and outside GoStudent that we talk to about AI.

AI tends to amplify human traits. So if you have the right traits, and those traits are the ones that usually win in the marketplace, they will take you further along the way.

What I mean by finding the blind spot before you build is this. As a leader running revenue teams, every team and every business has blind spots. Every business has specific challenges that, regardless of AI, you have to fix and find solutions for. So first, try to map out the top five priorities that would generate revenue, improve your bottom line, and fix your P&L. Then start adding manpower, resources, and budget on AI tools.

If you start using AI tools without a clear destination in mind, you are just going to spend money. You will probably learn things along the way, but you will not be as efficient as someone with a plan and a goal.

John Emmerson: We hear this time and time again from the people we speak to in this interview series, and in the calls we have about our events. It still blows my mind how often people get this wrong. They buy an AI tool and expect it to solve all their problems, without having figured out what problems they are trying to solve. It is nuts. That is not an AI thing, that is just general business management, from my perspective.

Achilleas Kasimidis: True. There are types of personalities that tend to go deep. They are curious, they want to get their hands dirty, they want to move things from A to Z. Those people, with the right use of AI tools, will definitely get forward and will stay relevant in the future.

Because again, everyone these days talks about AI. Everyone talks about Claude, about ChatGPT, about all the different LLMs out there. But if you do not know why you are using those tools, you are just going to spend money. You will probably make something nice, a good to have, but nothing that will turn things around.

John Emmerson: I love your use of the word curiosity. Whenever I have been involved in shaping values for businesses I have worked at, curiosity has been right at the top of the list. If you build a business with enough curious people in it, that will by its very nature deliver good results, because they naturally go the extra mile to find out what certain processes and tools can do. So I completely agree with you there.

Changing tack slightly. My understanding of GoStudent is that you run at quite high pace. Something like 300 leads per account manager per month. You have an internal SLA, I believe, that your first call is in under 30 minutes. In that environment, where things happen really quickly, what goes wrong when you integrate AI into it? And for a mid-market business that maybe does not have your operational depth, what should they look out for?

Achilleas Kasimidis: Let me be clear first. I am a big fan of speed. I would rather not over-analyse, not overthink, and execute fast. Execute with pilots, with an MVP before the full-fledged product, learn, test again, and potentially roll out once we have the learnings in place. So from my standpoint, I do not want to slow down and try to have all the solutions and all the answers in place before I start launching a new AI agent.

Having said that, many things can go wrong when you first deploy an agent. Let me give you one clear use case. The first use case we had was a sales AI caller.

Let me also explain how our sales funnel works. We have quite a simple sales flow. We generate leads, so we have inbound lead generation, signups who express their interest to talk to us. Then our sales agents call back. As I said, we want to be super fast. Ideally we call within the first 10, even 15 minutes. Not even half an hour, that is a strict SLA we have.

During that first call we try to analyse your needs, match you with the ideal tutor, and book you a free trial session. So you get to know the platform, how it works, and whoever is on the other side as a tutor. Then there is a second call, where we follow up with the same human agent, trying to convert you from a prospect to a paying customer.

The first use case goes back three years, when LLMs were not as advanced as they are today, but we were still exploring how to use them. We wanted to automate the first part of the sales flow. So we wanted to remove the human factor completely, from the moment the lead is created in our CRM until we book a trial session.

We found out quite fast, because of the high lead volume the AI was handling, that within the first few weeks we saw an almost four times increase in the number of trial sessions the AI agent was booking, compared to the humans. We were A/B testing at the same time.

John Emmerson: Wow.

Achilleas Kasimidis: For us that was a very positive surprise, and we got quite excited. To be honest we did not fully know what to expect. It was a purely new thing for us, we had never had AI calling anyone. And we were quite pleased with that 4x volume increase.

However, when the closing call was happening, in this case from a human agent, we saw a steep decline in the actual last step, from trial to customer. Then we realised two things were happening differently. Our sales agents had a clear mandate to close a customer. That is also how we incentivise the teams. We pay based on the revenue we generate.

The sales AI caller had a clear mandate to book a trial session. So that mandate was in conflict with the end result, which is closing a customer. Essentially the AI was not pre-qualifying as strongly as the human would. They were letting more trials go through the funnel without the necessary pre-qualification during that call.

So on one hand we were happy with the volume. On the other hand, we had a higher cost with more trials that were not converting into paying customers. For us that was the first reaction to testing the AI sales caller. We thought we had everything in place, we thought we had trained the model to do what it had to do, which is booking a trial. But of course we missed things in the first rounds of testing.

John Emmerson: I am really interested in this example, because it is fascinating to use real numbers. I get that the mandates were different and therefore the results were different. But by driving four times higher volume, the conversion into customers was down. Was your overall volume in customers up, or was it the same? So those customers were costing you more on a cost per acquisition, because you were doing more trials.

Achilleas Kasimidis: It stayed the same. Yes.

John Emmerson: Got it. So what did you do to change that?

Achilleas Kasimidis: We experimented with different things. First of all we tried to educate and retrain the agent. So the mandate is not just booking a trial, the mandate is also pre-qualifying. Going through specific milestone steps before you get the yes and move the customer forward to the next step.

We also tried to introduce a set of further pre-qualification steps before the AI sales call even happened. We wanted to understand more about the main signals we were getting during the signup flow. So we made the signup process a bit more complex. We introduced more steps at that stage. That allowed us to process lower volume through the funnel, but at the end of the day the end result was higher, because we were converting more customers at the last mile.

So it was a mix of different changes, and it did not happen overnight. It took us a couple of months to optimise the flow. We did different A/B tests, we tested different flows, we even tested different agents, different voices, different genders. Many things can play into optimising the flow.

John Emmerson: Really interesting. That is almost against what most people tell you to do during a signup process. Most people tell you to make it as simple as possible, to remove the barriers. You have found it is more successful to get that level of detail about your potential clients during signup, even if it means more people drop out of the funnel, because you are converting more business as a result.

Achilleas Kasimidis: It depends on the actual service you are selling. For us, education is a big investment that comes from the family. We also found out in the process of testing and optimising that the buyer decision, someone deciding to trust us versus going offline or finding another tutor, takes time. You have to nurture them properly and be quite clear on what you are selling. Rather than get the volume in as much as you can and close 5% of that traffic.

So I think it depends on the type of consumer service you are selling. Not everything is the same, not all marketplaces sell the same product. In our case we also put a lot of effort into pre-filtering tutors. We invest heavily in tutor quality, and we invest heavily in the service itself. We would rather stay focused on what we do, than try to get everyone in, lose time, and spend sales CAC that will never convert into revenue.

John Emmerson: Really interesting. I find these real world examples fascinating, and I know our listeners do as well. That was one you said was three years ago. Have you done anything of a similar ilk more recently? Have you put an agent in place in the past six months or so that has delivered real results, and can you give us a lowdown on what that might be?

Achilleas Kasimidis: We did a different use case here. It is not a sales agent, not a sales caller agent, but it is still a commercial agent, a revenue agent. It is a co-pilot we deployed for the main tool our account managers use day to day, when it comes to interaction with the different marketplace elements. Tutors, prospects, customers.

It is WhatsApp. WhatsApp is the go-to communication channel for us for one to one individual communication. It gives us the edge on speed, reachability, and connectivity. This is a channel that helps us connect faster and reach our customers way more easily than any other touchpoint we have used in the past.

Of course we have our structured, centralised support channels for 90% of interactions. But for the last 10%, which are probably revenue-making touchpoints, we still rely on WhatsApp, and on connecting human to human.

Having said that, having a tool that allows you to process hundreds of interactions a day from a human creates a lot of admin work. It is a legacy debt for us. Because as we grow, and the more customers we convert, that growth creates a situation within WhatsApp that is not very well manageable. We also found out that during the eight-hour workday, our account managers were spending almost two hours every day on basic admin topics within WhatsApp.

Most of those touchpoints were customers asking for basic information. Payment details, access to their accounts, when the next session was booked. Or tutors coming to the same account managers to ask for a change to a session, or a cancellation because they could not join. So we still had agents being deflected from revenue-making activities into support topics.

So we built a co-pilot that took over almost all of that admin work. Instead of an agent having to check in the morning how many new inbox conversations they had, the co-pilot was doing that. It was running a full scan and handling most of what I would call the non-revenue related topics. That would free up two hours a day of the account manager’s time, to call prospects, book trials, retain customers, and go into things that help us with revenue growth and profitability growth.

John Emmerson: Two hours a day is massive. Assuming most people work an eight-hour day, that is a quarter of a day you are saving. So that must have had a pretty significant impact on your business.

Achilleas Kasimidis: Yes. One thing that led us to deploy this co-pilot was the time our agents were spending on outbound or inbound calls. We essentially run the business through communication that happens on the phone, between the customers and the account managers.

We found out that within this eight-hour workday, they spent less than two hours on a call. Cumulatively. Total call time on average was less than two hours, less than 120 minutes, which is quite on the low end. Because they had another two or three hours spent on admin, plus their breaks, plus other things we could not always cater for.

So having the ability to take that time back allowed us to increase economies of scale. First of all it allowed us to increase call time for the agents. Revenue per agent went up, customers per agent went up, and overall the model worked way healthier than what we had before.

John Emmerson: That is wonderful. I love this stuff, I think it is fascinating. I am curious, did you build this yourselves, or is it something you bought off the shelf? What are the tools you are using that sit underneath the tech driving this?

Achilleas Kasimidis: In fact, this is an agent I built myself. It was one of the first agents I built with my own hands, again without coding, but having AI coding for me. The tool I used was Antigravity, which is a Google powered tool.

It allowed me to build a system that connects through your WhatsApp, scans your inbox, takes over the communications, and gives you dashboards and reporting. So on my side I can see how many messages we get on a daily basis, I can see where we had bottlenecks, and I can understand which admin and revenue-making tasks happen within that specific channel. All of that was built by myself.

It needs to scale at some point. Right now we have 20 to 25 agents using this co-pilot, and we are planning to hand over what I built to our tech team, to make it more scalable, more stable, and expand it to the full force we have at GoStudent.

Just to give you perspective on time investment and cost investment. It took me roughly two to three weeks to roll out the first MVP version, which was a light version of what today’s version 3.4 looks like.

At the same time I built different feedback buttons within the tool, so the actual humans behind WhatsApp could tell me whether the agent was doing what it was meant to be doing. That goes back to the first AI use case that was not working as expected. So for me it was really important, while deploying the agent, to have a way to track what the agent is doing and get real time feedback from the actual humans who have been doing that work for the last few years.

So that is where we are today. Two to three weeks to roll out the MVP, and a couple of hours a week to optimise and roll out the second, third, and fourth versions of the co-pilot. In terms of cost, it is less than 200 euros a month. For 25 co-pilot agents running real time, 24/7, handling tickets and handling the touchpoints, all that admin work that has been outsourced to the co-pilot.

John Emmerson: That is absolutely unbelievable, 200 euros a month. And that is just token usage, right, through Antigravity? I have not heard of Antigravity. Is it a tool where you use natural language to explain what you want to build, and you use MCPs to connect it to your WhatsApp and your CRM and various other bits and pieces? I know I am simplifying it massively, but is it broadly that kind of tool?

Achilleas Kasimidis: Yeah. It will write code, so it can build apps. It can build more or less whatever you would like to design. It is quite powerful. It builds a solid co-pilot that can execute what I described before, and we have seen so far that it is also quite stable, compared to a solution you would host with a tech team.

John Emmerson: You have quite a distributed team, right? You have people in Spain, in Italy, and various countries around Europe. So these are different markets, different languages, people doing their jobs in different ways. First, how do you get the tool to adapt to those differences? And second, how do you get people to adopt it, use it, and trust it? Two very different questions, but both really interesting.

Achilleas Kasimidis: On the first point, making it sound native and work as a native agent. In Spain, in Italy, north of Italy, south of Italy, and so forth.

The first part is that whenever you roll out an MVP version, you start with a basic version that is doing the job. It is not at 100% where it should be, it is probably closer to 60 or 65%. Then, when we rolled out this MVP, we selected a handful of account managers in the markets we wanted to roll it out in, and they were closely involved from design to set up to execution.

So we made them part of the process, we made them part of the product. They were the ones giving us feedback on how well the agent was guiding and handling those touchpoints. Whether the language used was the right one, too formal or less formal, and whether it identified the use case and took the right action.

So on the first point, I think you still need the human to be part of the process. You still need someone who is your point of contact, who can help you localise and take whatever you build into the next phase.

On the second point, trust. From my experience, when you solve a problem for someone, you can easily win. I would not say trust yet, but you can easily get endorsement.

John Emmerson: Yeah, interest.

Achilleas Kasimidis: For those account managers handling hundreds of discussions every day, the fact that an agent was handling 80% of that admin and noise, the whole idea sounded exciting. So we never had an issue selling it over to them.

Trust is something that takes time to build. By making them part of the process, by asking them to co-design and co-optimise what we were building, that helped us build trust over time. At the beginning most people are quite sceptical, especially a few years ago. They were, I would not say hostile, but very sceptical about AI handling touchpoints, discussions, and chats. Of course they have seen that over time those models have improved a lot.

By doing all these things, by getting them part of the process, by co-building with them, and of course by the tools themselves advancing, over time you build trust. But for me the first point is solving someone’s problem. You are always going to get a yes, and at least a positive attitude at the beginning.

John Emmerson: It is very sensible to involve them in the ongoing product design and the iterations of the models. That level of engagement, and the fact they feel part of changing something to make it better for themselves, will always drive engagement.

Flipping slightly. I keep asking you questions about this because I think what you have built is brilliant. But you mentioned to me earlier that GoStudent is progressing from a primarily online business to having an offline presence and training centres as well. What is the interplay between AI and that physical footprint? Does it change how you think about using AI, or is the thinking basically the same?

Achilleas Kasimidis: It does not change. From what we have seen so far, and this is something we started about a year ago, we do not see a difference between using AI for an online marketplace and using it for an offline business. AI can help scale the online business as well as the offline business.

With the right AI tools, for example, when we have a new offline centre opening in Milan or Madrid, we can fully automate the entire marketing stack with AI agents. We can easily replicate what we built for one centre to 10 or 20 centres. So we see more or less the same needs and similar use cases for AI in an offline centre as we have seen for the online business.

John Emmerson: I think what you are doing with this stuff is fascinating, and I am so impressed. Every time I speak to you it gets more impressive. We are nearly at time, so I will ask you the same question I ask everyone at the end of these calls. The vast majority of people listening to this are trying to navigate the world of AI. What would be your single piece of advice for a leader in a business, in any function, on what they should be doing to start with?

Achilleas Kasimidis: If you have never been into this so far, step number one would be to choose your tool. Name it Claude, name it Gemini, name it ChatGPT. Choose whatever works best for you as an LLM, as the major go-to tool you are going to use going forward for AI. And upgrade to a paid version.

By the way, I am not sponsored by any of those LLMs. I am saying out loud what I have done myself, and why I am here today. But definitely, step number one, choose the tool you want to use and upgrade to a paid membership. Free tools usually discount heavily the output you can get out of them. It could be a 20 euro or 50 euro monthly subscription, it could be a 200 euro subscription. You will see as you go in, and the more you use the tool the more expensive it probably gets.

Then describe who you are and what you do, both as a person and as a business. So build the memory. Build as much memory as you can behind the system you are using, so it can be your helpful assistant on a daily basis without you having to repeat yourself every single day.

Then try to map out what I said at the beginning. Maybe not necessarily blind spots, but write down your top five biggest challenges as a business today. And see how AI can help you solve those challenges. It can give you ideas you have not thought about before, and answers to things you might not have the answer to today. But clearly, try to build a list of things you want to solve, and then work through them with AI. See how a tool like Claude, for example, can help you work through those challenges.

Do not be afraid to start, and do not be too budget conscious at the beginning, at least. Of course you need constraints in place. You do not want to overspend on AI, especially if you have not used it before. But do not think about budgets, do not think about the cost. Set it up.

There are a lot of things out there online. A lot of leaders, a lot of voices that talk about AI. Find those voices that work well for you, and follow what they are already doing, and how they are already using AI to solve their own business pain points.

John Emmerson: That is really good advice, and I would add another element to it. Earlier this year I flipped from ChatGPT to Claude. When I started using Claude, I told it how I like to be interacted with, and what my objectives and goals were.

But then I asked it to interview me. I specified that I wanted a minimum of 150 questions, all multiple choice, so it was easy to respond to. Interview me, find out everything you need to know about me to be the best version of my assistant you can be, given the information I have already given you. Then I asked it to update that every month. To re-interview me to a lesser extent every month, based on its capabilities and how my objectives have changed. That is now a scheduled task.

I find that incredibly valuable for the value I get out of an LLM like Claude. And again, I am not sponsored by Claude, it is just the one I choose to use at the moment. I am deeply unloyal to these things. I have everything on a monthly subscription, so if something suddenly becomes much better, I can flip very easily.

Achilleas Kasimidis: I like the monthly check-in. I have also done that once, when I switched over to Claude a couple of months ago. But I do not run the check-ins on that monthly cadence. I like this idea, I am probably going to do this as well.

John Emmerson: It is a two-way thing, right? Your own personal objectives and the way you do things change all the time, especially when you are running a smaller business. But the abilities of the LLM also change constantly. The amount of update notifications you get from something like Claude or ChatGPT is many times a day. So they can do more. It might not understand that at the beginning, but the fact it can do more means it will ask you things in a different way, and you will get more out of it. I think it is hugely valuable.

Great note to end on. Achilleas, it has been as always an absolute joy to chat to you. Thank you very much for joining us today.

Achilleas Kasimidis: Thanks for having me.

Why should you start with a business problem rather than an AI tool?

Kasimidis argues that anyone can build with AI now, so knowing the tools is not the advantage. His view is that AI amplifies the traits people already have, and the traits that win are judgement, curiosity, and the willingness to see something through. The practical instruction is to map the top five priorities that would generate revenue, improve the bottom line, or fix the P&L. Only then should you put money, people, and tools against them. Start with the tool and no destination in mind and you will spend money, learn a few things, and be far less efficient than someone working to a plan.

What went wrong with GoStudent’s first AI sales caller?

The mandates did not match. The AI caller was told to book trial sessions, and it did, booking around four times more than the human team in A/B testing. Human sales reps were paid on closed revenue, so they pre-qualified harder. The agent did not, and more unqualified trials moved through the funnel. Conversion from trial to customer fell steeply, and total customer numbers stayed flat. GoStudent ended up with the same revenue at a higher cost per acquisition. Kasimidis is clear that they thought the model had been trained correctly, and only found the gap once the closing calls happened.

How do you fix an AI agent that produces volume but not revenue?

GoStudent changed both the agent and the funnel around it. The agent was retrained so the mandate included pre-qualification, with specific milestone steps to clear before moving anyone forward. They also added qualification steps earlier, inside the signup flow, so stronger signals arrived before the AI call happened. Then they tested. Different flows, different agents, different voices, even different genders. None of it happened quickly. Kasimidis says it took a couple of months to optimise. The result was lower volume through the funnel and a higher end result, because more of those leads converted at the last mile.

Is a low-friction signup process always better?

Not for every product. The standard advice is to strip out steps, and GoStudent did the opposite. Kasimidis explains why it worked for them: education is a significant family investment, and the decision to trust a provider takes time. Buyers need nurturing and clarity about what they are buying. A more detailed signup produced fewer leads and more customers, because the extra information let them filter properly. His caveat matters. It depends on the service you sell, and not all marketplaces sell the same kind of product.

What does an AI co-pilot for account managers actually do?

GoStudent runs high-volume one-to-one contact over WhatsApp, which generated a lot of administrative work. Account managers were losing around two hours of an eight-hour day to it. The queries were basic: payment details, account access, when the next session was booked, plus tutors asking to change or cancel sessions. The co-pilot runs a full scan of new inbox conversations each morning and handles most of those non-revenue topics. Kasimidis built it to give that time back for calling prospects, booking trials, and retaining customers.

What did it cost to build an AI co-pilot in-house?

Less than most people assume. Kasimidis built it himself with no coding background, using Google’s Antigravity, which writes the code for him. The first MVP took roughly two to three weeks, and later versions needed a couple of hours a week. Running costs are under 200 euros a month for 25 co-pilot agents operating around the clock. He also built feedback buttons into the tool from the start, so the humans using it could report whether the agent was behaving as intended. The next step is handing it to the tech team to make it scalable and stable across the whole business.

How do you get teams in different countries to adopt an AI tool?

Put them in the build. An MVP will be around 60 to 65% of where it needs to be, so GoStudent selected a handful of account managers in each target market and involved them from design through to execution. They fed back on whether the language was too formal, and whether the agent identified the use case and took the right action. That solved localisation and adoption at once. Kasimidis is careful about the word trust, which he says takes time. Solving someone’s real problem gets you endorsement first, and co-designing the thing builds the trust afterwards.