Fost Singapore 2026 - From Models to Moments: Why Most AI Products Fail in Week One.

This talk explores the critical relationship between APIs and artificial intelligence (AI), emphasizing that APIs are essential for AI's success. The speaker discusses the high failure rate of AI products within their initial weeks and identifies the lack of impact as a primary reason for these failures. He introduces the concept of 'agentic APIs,' which are designed to understand user intent, execute actions, and learn from interactions to improve user experience. The speaker provides a prototype demonstration of these agentic APIs in action, showcasing their ability to analyze bank statements, categorize transactions, and deliver personalized insights quickly. This session highlights the evolution of APIs from simple data retrieval to proactive systems capable of executing tasks, ultimately advocating for a shift towards these more intelligent API frameworks.

about models to moments. And before that, uh just a small uh context here. Uh I'm one of the uh speakers of API Days events. Uh in API Days Singapore 2026, it was my uh third time being there uh physically. Uh there was an online event also happened in API Days Singapore, where also I was privileged to speak. It is always a good uh experience to be there,

talk to people uh who uh think alike. And uh initially we started with APIs and eventually with this, you know, uh revolution of AI and agent API and MCP, we now are, you know, shifting our uh conversation. I met many people over there, many good uh and old friends of mine, and we all were talking about uh AI and all, but one thing is for sure that

we all uh very commonly spoke is the API uh part, right? The role of API, the importance of APIs in the era of AI and agents, right? Uh >> [snorts] >> one thing unanimously we all pondered over and we conclude that without API, AI is nothing. Without [snorts] API, agents are nothing. So, it's a it's a very uh you symbiotic relation that APIs and AI have these

days. And the next most uh important or critical point that we all discussed is that the ROI or the success of any AI product. In last 2-3 years, there are so many uh innovations we have seen around AI and I I mean, as I myself have uh seen many things in the workplace that around AI being uh started. But most of the time we have found that

the initiatives or the products that is those are getting built in with the help of AI are mostly getting failed in next few weeks, right? Why so? And that's the topic that we going to discuss today. Before that, a small introduction about me myself, Abhijit Day. I'm one of the uh regular speakers in API Days events and professionally uh I'm the senior vice president in Axis Bank

India, which is the third largest private bank in India. I lead all the AI and API banking product initiatives. Uh my total ex is around 19 years and I uh I'm working around this API space for last 10 12 years. Uh that's a small brief about me, but uh I'm more curious to talk about the topic that I'm going to deliver today and which I have actually

delivered to the August crowd of I mean before the August crowd of API Days Singapore uh day before yesterday on uh 15th of April. Uh [snorts] so, the topic is that models to moments, why most AI products fail in week one and how agentic APIs can fix it. Please be noted here, I have introduced a new word called agentic API. Uh we need to uh we need

to delve into this. We'll see what agentic APIs are. And uh going forward, we'll we'll talk about that this product failure are not about the lack of intelligence because AI is all about intelligence, right? Uh intelligence is not missing here. What is missing is that impact, the ROI, When we talk about ROI, uh a very uh significantly uh interesting uh thing I heard recently is that uh

you have seven, eight, or 10 uh AI initiatives on or AI products, right? Uh what you uh actually see the the the retention reality of those products are you'll see that one or two will be successful, Uh similarly, when you are a VC and you have invested in multiple startups, you will see uh you have invested in 10 startups, hardly one or two will click or one

or two will get success. Rest all you have to average out, but the learning that you get from those seven or eight really failed startup investment, that will actually help you to uh gain more benefit for your for your next adventure, right? So, similarly, in AI also it's more like investing in startups, right? Now, what is a retention reality? This I have uh done a research. This

I have done a research with many of my peer uh colleagues, my friends who are into the same space like me. All of they are saying that on the whenever they launch any AI product, right? On the very first week with those campaigns, with those nudges, customers are very excited on the day one and they uh open the product. They see what's new here, what I mean

in fact any any of the uh you know new uh features get uh published by all these, you know, you know, cloud or ChatGPT or, you know, open cloud. So, whenever these all come, uh you'll see uh on the very initial week, they are people are very excited, right? Uh [snorts] normally in banking or in finance, we have seen that on by day uh most of these

products gets abandoned. I mean, why it it happens? That is a key question that we have in our mind and that's a key question that we'll try to uh try to you know, resolve today, right? Now, you'll see there are no crash in your app, there is no complaint on the products. Just just a silence. I mean, there is no response from the customer, nothing. No no

interaction, You do a complete log checks and all. You will see whatever questions have been raised or whatever problems been triggered, all been answered correctly by the AI, but there is no change. Why? So, here the intelligence is not the problem. The problem is the impact, the usefulness. So, whenever we are targeting any uh targeting to build any AI product, the things which we should keep in

our mind is that whether this is really impactful for customer or not. Whether this really needs AI to solve it or not, you will see the graph here, the retention from day one to day seven. It's gradually fading out. The deeper is the you know, more you know, user retention, but uh eventually you will see people are not using your app because it's not actually targeting the

right problem statement or customer base, right? So, it's again not about the intelligence, it's about the impact that you are targeting here, Now, what user expect from AI to do things, right? User expect that AI will do the thing, right? Most of the AI around the recommendation. Most of the AI products are about the suggestions, right? What user expect and this I have seen in many of

the banking or finance app, right? What user expect is that a assistant that can act on his behalf. It remember whatever been told to him. It will take action based on his behavior, based on his pattern, based on his nature, right? It it generate personalized context, right? Uh then the last one is that it will integrate with the other tools, right? Maybe your banking app can be

integrated with your Outlook or your banking app will be connected with your calendar. So, these type of MCP or this open kind of integration is very important here, right? Which user actually expect. But instead what they get is they get a chatbot. Most of the time we build a chatbot or bot-like engine and say that this is an AI, which is actually not bot or chatbot kind

of AI initiatives are, you know, 8-9 years old initiative which maybe are, to be very honest, are not relevant today. Especially when people are more into conversation, they're not into chat, right? And I do believe that in next couple of years, none of the app, right, will be a regular app. It will all be, you know, transformed into a conversational app. So, that's what user expect, right?

the gap. What user expect, what they get, and what are the key gaps that all these AI products are having and why these AI products are failing there, right? First of all, the memory. Whatever user is searching, whatever user is interacting with you, you should keep it in your memory, right? Because whenever the user is coming back, the context has to be there. The context has to

be provided over there. Otherwise, the generic output will be for everyone, right? For example, if let's say I'm I'm in a travel app. I'm I'm not taking consciously I'm not taking the example of any banking app So, let's say I'm in a travel app. And I booked it I have searched for a ticket Mumbai to Amsterdam, okay? Now, today I have searched after 7 days again again

when I'll be back. And this time maybe I'm looking for some hotel, right? It means within these 7 days I may have booked my air ticket from uh Mumbai to uh from Mumbai to Amsterdam from any uh other uh website or from any other app. But, when I'm back after 7 days, the nudge, the agent, there should be an agent experience that will identify that I have

already booked my ticket, and now I'm here for the hotel, and that agent will nudge me with the best hotel deals in Amsterdam. So, these type of contextuality is missing in most of the air product, and that's why the personalization touch is missing. People started thinking that oh, wow, I mean, again I have to redo the things. Oh, you don't know who am I. You don't know

I mean, you're running an app. I'm doing business with you for uh over over many years, but you still know about my preferences. So, these type of negative feelings are the gap and these are the systematic gap. These are not the problem with the uh service provider. It's not the uh booking agent or the problem with the bank. It's a problem with the system, the technology, the

architecture which you have built, right? So, this we we need to actually answer for. The third one is no execution. I mean, the same example. Uh let's say I come back. You showed me all these book uh hotels and all. Now, you want me to select the room, identify number of travelers, what are all my preferences, then you want me to book the room. Why not that

app itself says that hey, look, based on your preferences, we have identified this particular hotel room. Uh this much is the payment. I have identified that you have these many credit cards. Out of these, this credit card number one for bank A, if you use it, you can have a benefit of 5%. You just say, yes, I'll book it on your behalf. This type of execution we

need. And most of the time we found that it's not there, right? So, when we dig into this why this happens, right? uh execution is a key thing that is missing out. Now, the last one is the integration. Where again same example, the dates which I'm looking for or let's say there are some clash. Uh there are some important meeting there where I'm actually and on those

dates I'm actually searching for my uh Amsterdam hotels. So, all these uh uh third-party integration should be there, right? We have MCP server to do this, but most of the time we uh we cannot see uh that's what happened. So, this is are the things These are the four gaps that we really need to address here, right? And these are the very key uh problem statement, right?

That we most of the time miss out. Now, I'm introducing Agentic API. How [snorts] these four system gaps, right? These systematic gaps we can resolve with Agentic API? So, Agentic APIs are the missing layer, right? If you see the uh evolution of uh Agentic of the APIs, it all uh started where uh it used to wait to be called and it returns the data. The normal raised

APIs, right? Then we got LLM APIs where it generate a response, but it doesn't act. Agentic API are the new thing which I'm pretty sure not the last one because there will be more evolution happen on on API. so, Agentic API it understand the intent. It decide, it execute, and then it learns from it. So, the contextuality, memory, action, and learning all are there in this Agentic

API uh paradigm. Now, I'll take you through uh uh uh, prototype where we have built I mean, I have built a protocol Clarity which will help you to talk to your bank statement. Everybody has a bank statement, right? Uh, we all have a questions that how much did I spend? What is my food spend is high or not? What are my subscriptions? Uh, this one space where

all the other banks, all the, you know, FIs are working uh, across the globe, right? How we do this? I mean, how we can do this? First, we need to post the statement, then we need to classify, we need to categorize all the transactions, then we need to analyze it, and finally the response. These are the four basic tasks whenever you are doing any, you know, statement

analyzer, right? But, with the help of Agent KPA, we can do it in a quickly. I mean, uh, in my prototype, you will see uh, there are around in 1.2 seconds all these five Agent KPAs been uh, triggered, right? It loaded the statement of around 847 transactions. Uh, how much did I spend? If that is a prompt, it will identify the uh, intent. It will classify the

intent with the uh, data category. It will calculate or compute the numbers. It will also create an insight. So, there are five, six different components that we are doing, and all these we are doing with multiple APIs, and each and every API is connected with an agent, right? The first one is the document parser API. It uh, ingest any PDF or any file, right? Uh, then the

merchant classification API where it actually identify the merchant name from the data, then it will categorize it with some based on its MCC code. Then the third one is the analytics API where it will actually aggregate the uh, spends. It will aggregate the data into different uh, duration. The fourth one is the memory where it actually do the profiling of your uh, transactions, that how much you

have spent in which category, where your your your intention is more towards spending on food or travel category or any other uh, uh, different categories. The last one is the inside where we actually detect if there are any anomalies there or if there is any overshot budget or if there are you are, you know, you're spending less on any particular of a credit card. It's so all

these five tasks with the help of Agent T K P I, we call it in 1.2 1.2 seconds, right? Now, uh, a small demo I'll try to run here. You'll see, how it works. >> So, here you just type it that how much you have spent on food last month and how does it compare to the month before. And if I play it, >> So, in this

demo if you see, uh, this is a prototype. Uh, in this I have uploaded my bank statement where I can find that there are many, you know, subscriptions are there, there are gym memberships are there, there are Adobe Creative, unused for our last 47 days. All these we could figure with the help of this Agent T K P I right? My Analytics API identifies that there are

unused, uh, membership. My merchant classification API understood that, uh, these many merchants I'm I'm I'm dealing with in last, uh, 6 months in in, uh, subscription uh category. Uh my memory API actually remind me that uh for 47 days that Adobe Creative uh uh subscription is unused. My insight API says that the last one, the total monthly subscription is going uh I mean, is in an average

uh spends, right? Or if you see that Adobe Creative Cloud, the last uh on the bottom, right? This is the insight that we generate that if we cancel this, it will be around 20,000 uh rupees of saving. So, this one uh use case, and this is completely uh prototype that was built by me, right? Uh here, I can actually identify what all things we can do with

this help of Agent K API. And the time you see around uh on the uh till fourth API, right? It is around uh nine .9 seconds. Now, it is going to uh if I'm going to uh prompt again, right? It will take me to the uh 1.2 seconds of time limit, right? So, this is the demo I was talking about. Here, the uh with the help of

Agent K APIs, we are reducing the time, we are getting more clarity, we are getting more execution, right? >> And this will turn all of your AI product from a chatbot into a system that can actually act. Uh we need a system that can act, that can react, right? Uh How we can do this? With normal LLM, we can generate response. It's very uh reactive kind of,

>> but not proactive, right? Uh without Agent K API, we'll see our model will generate a response, user can read it, user can act on it, maybe. Uh no memory, no follow-through, product gets abandoned in week two. But with Agent K API, we'll see this agent understand the intent, it calls the right API in sequence, it the action executed automatically, right? Now, agentic APIs and it it

works in any domain, right? I mean, regardless of saying I mean, APIs never say that I mean, in since uh the inception API, none of the APIs say that this will only work for one particular industry or one particular domain, right? So, in >> [clears throat] >> similar what agentic APIs also. Now, the next one is that winning AI products which are, you know, built as a

closed loop, right? Uh you see the user uh write the prompt, LLM it intend the understanding, memory layer where the conversation, all the profiling happen, agent layer where the orchestration and decision happen. Then it goes to purse API these uh if you see these, you know, three APIs, right? Purse API, analytics API, and classify API. This is the analogy of the entire architecture that I have built.

Where purse can store the transaction, analytics can store the spending, classify uh API is actually identify the merchant. And on top of that, we have a guardrail and privacy layer where we are controlling all the responses. Uh we are controlling what we are feeding to LLM, we are controlling what LLM is uh fed back to us, and the last one is the feedback loop where we are

actually capturing customers in I mean, feedback. This is how you can actually create an sizable winning AI product. And the value in AI is created at the moment of action. I mean, it's uh little uh a little refresher here, but to be very honest, without agentic API model, then prompt, then response, nothing else. With agentic API, you can actually write a question or prompt, then it goes

to agent, agent will talk to the to API, and you will get the outcome. So, this is not a future, this is a paradigm shift that we should think about. At least I got a response here in my prototype and I pretty sure if you try [clears throat] it at your end, you will also get some good response over here, right? The last one about the clarity

this product that I was talking about, it migrated from chat to execution from product to system and insight to outcomes, right? And this will be proven in every industry, right? Because I will not win by being smarter. It will win by acting at the right moment. So, the right moment is here because now we are sitting on an ocean of data where we need to sniff through

this data, get some insight, and not only just look at the insight, we should take some action on that insight, right? So, agentic APIs or the APIs which are agent ready are very critical here, right? And we should all think about it. Having said that, I'll end it here. It was always very exciting to present before the August crowd of API Days, be it Singapore, be it

Delhi, be it Munich. I really enjoyed it. And this concept of agentic APIs which I have prototyped before this event and I presented in Singapore, I got a very positive feedback and I'm hopeful if you try it response. Uh you can connect over me LinkedIn uh through my scanning this QR code and uh I'm happy to be connected and happy to chat over it and please do

let me know if you are using agentic APIs at your end and you're getting success or not. Even for failure, you can reach out to me. Thank you.