Webit'26:The AI-ready company: from data readiness to AI execution: building the AI-ready enterprise

In this talk, Sandipan Pal from Databricks discusses the critical components of building an AI-ready enterprise. He emphasizes that successful AI adoption is less about selecting the best technology and more about addressing organizational and data challenges. Pal shares insights from his experience in financial services, illustrating how AI systems often fail due to a lack of proper training and context. He argues that organizations should treat AI onboarding like human onboarding, providing the necessary data, governance, and architecture to support AI agents. By focusing on foundational aspects such as data quality and governance, organizations can ensure smoother AI integration into their workflows.

For the past 2 years, almost every company has been talking about AI. The leaders are now doing something different. They are building. Because the question is no longer, do we have an AI strategy? The question is, can we execute it? Our next speaker takes us into one of the most important realities of the AI economy. That successful AI adoption is not [music] primarily a technology challenge. It

is an organizational challenge, a [music] data challenge, and ultimately an execution challenge. Warm welcome to Sandipan [music] Pal from Databricks with his keynote, The AI Ready Company, from data readiness to AI execution, building the AI ready enterprise. >> Wow, this feels like an ungarded boxing ring. Which is appropriate though for this conversation that we are having. The debates and discussion with AI for the past few years

have been a boxing match. Open AI versus Anthropic, GPT versus Claude. Uh regulations versus innovation. Humans versus AI. Ladies and gentlemen, I'm Sandipan. I work at Databricks leading some of the technical initiatives in financial services organizations in UK and Ireland. And in my day-to-day work, I fight my fight on talking to customers on how AI works and how can an organization get ready for AI. And this

is what I'm going to talk about today. 6 months ago, I was working with a bank. The bank hired a new employee in the contact center. And after a few weeks, they had to fire that employee. It's sad. But, what they found out in their investigation is that their customer service a customer satisfaction rating started dropping few weeks after the employee joined. They found that the employee

was giving customers incorrect answers, misguided answers, often forgetting conversations. For one instance, they found when a customer called for waving off their overdraft fees, the the the employee said it can't be according to policies. And which was the correct answer. But, in the same conversation, the employee again stated, "It can be waived off." Which was incorrect. It not only misguided the customer, but it also resulted in

loss of money for the organization. That employee was an AI agent. In their investigation report, they wrote easy answers that the model was not working. The AI did not understand the customer's questions. The accuracy metrics were not up to the mark. Which are easy to answer. But, I was thinking, how does a human employee get into this conversation and not forget about policies just mid-conversation? How does

a human employee has got all the resources, tools, and skills to address customer questions and why can't an AI agent do that? Is it because we onboard human employees with all the necessary knowledge that they need in our organization? If it is so, why don't we do that for AI agents as well? You see, what I found out is for 30 years we have been installing software

in organizations. Software work on code. They are pretty deterministic. You install it. You work with it. If it goes wrong, you know that it's not working. AI is a bit different. AI can reason. It can decide. It can act. It is much more like a human when it works in an organization. And if humans need to leverage AI in a workforce, then we have to treat the

onboarding of AI like we onboard humans. We have to give it enough information about the workplace. We have to give it enough data. We have to train it with the skills that the organization need. And we have to apply appropriate guardrails so that it does not misguide the customer with wrong messaging. So, the core thesis of my talk is an AI-ready company isn't the company with the

best model or the best AI It is a company which builds itself. It makes itself ready to onboard AI agents as the future part of the workforce. Uh last week in the data and AI summit, our founder Ali um mentioned he argued that artificial general intelligence is already here. The models are already pretty good, right? They're solving humanities' last examination questions. And And if you see the

reports, they can answer most of the questions that most humans cannot. But when you put that same model in your organization, it it it's it feels like a dumb It it it cannot do enough it because it doesn't have enough understanding of how your business works. the real question we need to ask here is when AI fails, are we actually giving AI the right environment and the

right system for it to operate in before we start firing AI from the workforce? And what do I mean by environment? By environment, we mean think about an employee onboarding uh an organization. You give it a workplace to work in. What that means is that the employee understands which team he or she is a part of, which teams they interact with, what is their business, how do

they make money, how do they make sales, et cetera, right? Then you give it the um the ma- the manager, right? Every employee has a manager because you want make sure that employees can improve, they can grow in an organization. They are held accountable for they they understand, they get continuous feedback on on their performance. And then there is the building. The employee when they operate in

a building, you need to make sure that the right utilities are in place, there is a toilet, there are fire alarms in place, they are safe, et cetera, et cetera, right? So, you make sure that the system, the environment is ready to before you hire an employee. The same goes for AI. When you onboard AI, the main focus should be on building the foundation and not on

which model you want to use. And that is the common problem I see in organizations. Because we all start from whether this model is better than the other. All right. Which model performs the latest benchmark test very well? And that completely shifts our focus towards model rather than focusing on what do we actually need from the foundations. For example, one of the foundations that you need to

talk about is data. What I find in many organizations is data is messy. It is locked. Not everyone has got right the right access to the data. And often it is hidden in dark places in nooks and corners of your organization in different departments, locked in a remote database, or even as hard copies as files in some cabinets. AI needs all of that context. It needs the

full picture to work well. If it doesn't If you don't provide that, then it's not going to work properly. The second one is governance. I almost think about governance is giving a manager to AI. Because you want to make AI accountable. If an agent makes the decision, you need to know who's accountable for that decision. Because you cannot make an AI agent accountable for anything. Right? There

needs to be a human tied to You need to know which what skills the AI has got, which tools it can use, what kind of permission it has got on on on the data. How what guardrails protect it from not doing stuff that you don't want it to do. That governance mechanism should be there and it's I almost equate it with having a manager in a workplace.

The third one is the building the the the the architectural foundation. We we all see that with AI, we build a lot of demos. The thing with demos is that they are not scalable architecture. So, a lot of the organizations I work with, they look at the demo, they get impressed. We as vendors do that as all as well. Like we are guilty of producing a lot

of demos, but demos show the art of possible, right? We go and show you what Databricks can do in a particular use case. That's the art of the possible, but that's not how it will actually function in production. You have to think about the foundations, right? So, that's where you need architecture that can actually work really well to support AI agents. So, let's dive a bit deeper

into what does ready ready data mean? Like, how can you get data ready? The way I see it is simple. You have got raw data in certain sources. It could be your databases, files, hard copies, external systems, partner systems, etc. or web. You take that raw data and you think about how you can make that data trusted, right? How you can make sure that data is clean.

One of the things we always talk about when we talk about AI is AI needs data, and the main problem with data is data quality. We have got unclean, dirty data, but I don't think that's a big problem. You will always have messy, dirty data. If you start spending time on cleaning up data, then it will take a long time to do that for all of the

data sets that you have got in your organization. I think the the the the thing that you should focus on on how can that data be trusted? Can we put that on a catalog where it can be tracked? Can we actually provide provenance on where that data is coming from? What changes it happened What changes happened from where the data traveled from to where it landed? And

what kind of code made those changes? If you have got access to all of those information, then you can actually trust the data. You need to make sure that you have got a security permissioning model on top of the data sets that you are onboarding in because you need to understand who can access the data. Are the right users having access to the right data set without

compromising on security? The third part is understanding, which is really crucial for AI, especially generative AI and agentic AI. Here is how it works. So, think about you ask the AI, "Find me all the high value transactions in my organization." Right? If you ask it to a human who is in the sales team, they have the context what does high value transaction mean? What does transaction even

mean? An AI agent, when it reads transaction, and it has no idea what transaction mean to an accounting, what transaction means to a sales team, what transaction means to a product team. Because every team has got a different meaning of that word. If AI doesn't get that information, it will not understand what to do. Then you've got high value. What does high mean? Every team will have

a different definition of high value. And the trick is high value will change every week, every month, every year. You have a new different definition of high value. Your business grows, your high value goes up. Right? Now, a human will have those information. Because someone might have your manager would have told you over Slack that this is the high value transaction today. How does an AI agent

know about it? It's not in your database. You did not define your high value transaction in the database. This is where you need AI to have those context. And what that means for us as an organization is that we need to change the way we work. We need to transcript our meetings where we can transcript them and provide that to AI. We need to check how AI

can look at Slack messages. Not personal messages, but, you know, there are there is a security and permissioning model that you can work with. Where it's not only databases and files that AI will need access to. It will need access to the concepts of your business. It will need to understand what transactions mean to different teams in the business. You have to map those concepts to your

data. It's termed as ontology. It's the buzzword nowadays if you have must have you heard everywhere. But then you need to give this the give AI this navigation path to go through. Right? To to to answer these questions. The second is governance. How do you get governance ready? If you go to a typical organization, you will see they have an asset register. They will register every chair,

every furniture, your laptop, desktop, everything. But they never register their data assets. They never register their AI agents. They never register their models. And if you don't register them, you don't know who owns them. Which team is what skills they have got, what permissions they have got. When did they run last time? And when did they fail ownership. You need a way of measuring AI outcomes. Is

it improving? Is it Is it not improving? Is it erroring out? Is it making the same call twice, four times? You have to find those things out. And then the architecture. I mean architecture is a complex topic to be honest, but I I I just simplified it in three pillars. Durable. You You want it to scale up and down in terms of usage, right? How how how

AI usage goes up, your architect You shouldn't have to sit and order a new set of servers through procurement for AI usage to go up, right? So it it needs to be durable. It needs to be resilient. If it fails, and it will fail, we have to design for failure. What happens when it fails? What is our fallback strategy? What happens when something fails at 2:00 a.m.

in the night? Who's accountable? Who gets the call? And who who acts on it? What are the next steps? It need You need to have embed embedded all of these policies in the architecture. And it needs to be connected, as I described earlier, to provide context, the way I've described earlier, you need the architecture to to be connected to all of these systems. Two things that AI

is not similar to humans is judgment and accountability. It cannot judge. You cannot hold it accountable. You have to engineer for those skills. So, remember that when AI gets into the workforce, you need to think about judgment and accountability. The other thing you need to remember is human fails one customer at a time. If an AI fails, one day in the morning, that failure snowballs into a

lot of other customers without you knowing anything because it it fails silently. It will not tell you anything. It will confidently tell you the wrong answer and go on. So, you have to think about that. So, how to think about moving on the ladder? Starting from experimenting to going to employing, you have to think about building this foundation. And you have to you don't have to spend

all year to building them. Pick up one single use case, one single business value that you can deliver. Narrow everything down to that scope and then build all the That happened with the bank that we work. When we relaunch their repair work, actually in the first seven of the eight weeks, we built just the We just chose the model on the seventh week. Choosing the model is

the easiest part of building AI systems. Once you have the foundation, it becomes even easier. So, as I finish this talk, I want to leave you with a question. If you had to make an your most important hire on Monday, which is an AI agent, do you think your organization is ready? Thank you very much.