3227: Dataiku on Managing LLMs Without the Chaos

[00:00:03] Genuitive AI is no longer a futuristic concept. It's here. We're entering our third year now since it arrived, and it's reshaping how businesses operate. But as companies rush to implement AI-powered solutions, many find themselves facing unexpected risks, governance challenges, and even mounting costs.

[00:00:24] So, how does your organization scale Gen AI responsibly? What are the hidden risks of AI bias, and how can your business mitigate them while maximizing value? Well, today I'm joined by Emma Irwin, Director of Sales Engineering at Dataiku, who's going to be talking about the most common mistakes that companies make when deploying Gen AI,

[00:00:50] how they can avoid them, why governance, access controls, and ethical AI frameworks are critical to success. All washed down with some real-world examples. So, if you're wondering how to balance AI innovation with business responsibility, you're going to love this one. So, enough from me. It's time to get Emma onto the podcast now. So, thank you for joining me on the podcast today, Emma.

[00:01:17] How can you tell everyone listening a little about who you are and what you do? Emma Irwin, Hi, everyone. My name is Emma Irwin. I'm originally from Canada, but moved to the UK just over two years ago to take on my current role at Dataiku as the Director of Solution Engineering. For those of you who don't know, Dataiku is a universal AI platform that makes it easy to build and deploy Gen AI solutions with confidence. Our technology enables rapid application development while ensuring security, cost control, and compliance.

[00:01:47] It provides both code-free and code-first tools. So, Dataiku can empower everyone in your organization to create and use AI and Gen AI applications. What that means for me is that I run a team that supports customers through their whole journey of evaluating and acquiring our software and then realizing value through use case implementation, which at this moment in time is pretty exciting with all of the hype and actual results from Gen AI use cases.

[00:02:14] It really is. And thank you so much for joining me on the podcast. Your co-founder, Florian, he's been on the podcast, I think, two or three times now. How's he getting on? Everything good there? Oh, my goodness. Well, he's doing extremely well, although I'm not his co-founder. I am just a team of three here. He recently spoke at our company kickoff last week in Prague.

[00:02:37] We had the entire org together to get excited about the new technology, where the company is going, and hear some pretty exciting stuff from him. So, I think he's probably riding that high. Awesome. Well, exciting times. And one of the things that I try and do on this podcast every day is talk about real-world business problems and technology problems and try and put it in a language everyone can understand.

[00:03:00] And I think many leaders right now are eager to implement some form of Gen AI and trying to work out the value proposition of doing that and understand the kind of problems that they can solve, too. But common cost and risk pitfalls can often derail these efforts. I think we've all seen examples of that. And even those that have been successful are starting to think, well, where's the ROI on this project?

[00:03:23] So, I'm curious, from what you're seeing, what are the biggest challenges organizations are facing when deploying Gen AI at scale? It is so true. I've actually never seen senior leadership so eager to implement anything in my tenure. But I think that, you know, the issues stem really from two major themes.

[00:03:48] One is governance related, and the other one is about the workforce and the skill sets that are available. So governance really has to be a priority. I think companies are still trying to figure out what is the framework that they want to work within, you know, without a framework or even if they have a weak governance framework,

[00:04:09] then organizations are vulnerable to privacy breaches, to bias incorporation in their models or to compliance failures, none of which they want to happen to them. And on the second point, upskilling the workforce, you know, if companies really want to get ahead with this kind of cutting edge technology, they have to embrace AI literacy, which starts with data literacy. They need to really invest in their employees so that they are preparing them for the demands of the use cases today and tomorrow.

[00:04:39] And I think that those two things together are ultimately what will help them reach their business goals. And the goals are huge. I think we recently ran a survey that said 88% of organizations plan to increase Gen AI investments in 2025. And that's on top of the 66% who had already spent a million dollars at least just in the last 12 months. So lots of lots of money.

[00:05:07] You want to make sure you're getting it right with the right governance and the right skill set. Yeah, I mean, some huge stats there and another huge one. I think recent data also shows that 88% of businesses, they're doing this, but they lack the specific applications or indeed process for managing large language models because it's fairly new to them. So what guardrails should companies be putting in place to ensure the safe and effective Gen AI deployment

[00:05:33] rather than running into mistakes and problems further on down the line by just blindly putting everything in and getting it over the line? Oh, my goodness. I know this part of the industry is moving extremely fast. I think the uptick in Gen AI projects across businesses has actually taken so many people, even us who work in this industry, by surprise. So I would say that there are three things companies should consider when hoping to ensure safe

[00:06:01] and effective Gen AI deployments. The first one is around access. So how will your teams access various LLM related services? You know, at Dataiku, we recommend that there's some form of abstraction layer, which would allow you to decouple a business solution from a specific technology because they're moving so quickly. There are new options every day and ensuring you have flexibility between or to swap between,

[00:06:26] you know, different providers or different brands and then actually lock down the access in an appropriate fashion is, you know, the first thing to consider. The second is about discovery, which may seem a bit odd, but I think it's a question to ask. How will you provide a mechanism to discover and document your LLM related objects?

[00:06:51] If we truly believe that this technology will be embedded in the business across the business, then we need a way to search and sort through where it is and how it's impacting us along the way. And then the last thing is about value quantification, because like we mentioned already, you know, we're spending and businesses are spending tons of money on these initiatives. So asking yourself how will you quantify the effectiveness of these Gen AI projects?

[00:07:18] If you're asked tomorrow to showcase the ROI of a Gen AI initiative, what would you do? Well, and even maybe a step further in that is that think about the two things potentially holding you back, which would be difficulty isolating the Gen AI's impact from other technologies or initiatives, and then also a lack of clear metrics or benchmarks for measuring success.

[00:07:43] I think that's one thing that when you get really excited and into the hype of a new tech, you don't always stop to assess the current status of things. And so making sure you have that benchmark is a great place to start. So really, you know, you said, what are the guardrails? If I have to summarize them, the three guardrails to consider is access, discovery and value quantification. A hundred percent with you on all of those.

[00:08:09] And another well-documented challenge that unfortunately refuses to go away is algorithm bias, which continues to remain a major issue in all things AI. So on that side of things, any steps that you'd recommend that businesses should be taking to reduce things like bias and inequality inside their AI models that actually promote fair and ethical usage?

[00:08:32] Because we've heard a lot of stories around the black box AI and people not knowing what that or how the AI is coming to some of the conclusions there. So any advice around this? Yeah, I've, you know, in the industry, we've thought about fair and ethical usage of AI for years, really.

[00:08:55] And I think the thing that we come back to each and every time is that to promote that fair and ethical usage of AI models, you need diverse AI development teams. So whether that's the individuals creating the model itself, which other companies and organizations are taking advantage of, or it's the organizations taking pre-built solutions and accessing, you know, the gold standard models that are out there,

[00:09:24] you need the diversity in the team who's implementing them. And so one of the things we do here at DataIQ is obviously build a platform that allows for collaboration between business stakeholders, data analysts, data scientists, executives, et cetera, really regardless of professional role or skill set. And it expands the number of people who are touching and creating the analytics, the models, the agents,

[00:09:50] by meeting them where they are and getting all teams across the org, sort of speaking a common language. I think that's the primary way that you ensure that you identify dApps in either the data or in the application. And you ensure that a wide range of demographic groups are included to find the best like system and generalizable results.

[00:10:18] So not something that just works for a specific subset of people or clients or anything, but really can be used in the most fair method. And something I don't think we talk about enough is how most IT leaders, they want to use fewer tools to simplify their AI implementation. It's music to their ears, but we need to talk about siloed systems that often make scaling up incredibly difficult, not to mention technical debt.

[00:10:48] But how can organizations better streamline their AI strategies while maintaining security and governance and all that IT stuff that I love talking about? All that IT stuff. Don't let them hear that. I think that I'd argue, you know, IT leaders, they don't want to just use fewer tools for fewer tools sake, obviously. And what they're looking for is a way to ensure that each step of the journey, in this case, we're talking about AI strategies,

[00:11:17] but I used to work in like big data infrastructure and it was kind of the same thing, right? Making sure their end users have an experience that is effective and efficient and hopefully problem-free so that they're not answering the red risks all the time. And from an AI project perspective, that spans design, deployment, monitoring, and governing. But so there's tons of elements we could really consider.

[00:11:43] But I think that if we really focus on security and governance, there are two key things that they can streamline. One is having a way to monitor everything in one place. So providing a single source of truth for any and all relevant individuals to have complete visibility across the data that is used and the AI projects that are created, all the way from data preparation, traditional ML,

[00:12:12] and now these more advanced Gen AI use cases. Two ways they could do it, you know, investing in a centralized model registry for all projects across all platforms, and then also agreeing on an enterprise-wide set of performance metrics or deployment status monitoring rationale. The second is, and I hate to say it because it is not sexy at all, but being always audit ready, right?

[00:12:41] We've seen more and more regulation and kind of government involvement in these technologies. And so staying ahead of audits or upcoming regulation by automating your documentation, clearing your approval trails. Two kind of standard ways we work with clients is implementing standardized approval workflows. So just having like a structured sign-off and audit trail

[00:13:08] for all your stakeholders so that you can point back to it and say, yep, we did our due diligence. And then the second one is comprehensive documentation. You can, with LLMs these days, just automatically generate your project documentation, your model records. You can have full transparency. And it is one of the usually most boring tasks to be done. But the thing that will save you time and time again, when someone leaves the business, a new regulation comes in, you're looking to optimize or streamline,

[00:13:37] that documentation is really a game changer. So I would say the security and governance really come down to monitoring everything in one place and being audit ready. And you probably know where I'm going to go with this, but of course, those are two key components of our platform as well. We really build, or DataIQ builds governance into every part, ensuring that regulatory compliance, ensuring business alignment, which ultimately lead to maximum impact as well.

[00:14:06] And a huge complex issue we're seeing playing out on the global stage right now is the UK and the US appear to be aligning on AI security rather than AI safety. And we've seen the UK recently shift from AI safety institute to an AI security institute. And the EU, on the other hand, they remain focused on prescriptive regulation, but over-regulation risks them getting left behind. And I think in the workplace, I think many enterprises are struggling

[00:14:36] with that balancing of innovation and risk when it comes to AI adoption. So what best practices should businesses be taking to optimize Gen AI's potential and not losing innovation or the ability to innovate while also maintaining control and compliance? It's a huge balancing act, isn't it? It is. And I think, to be honest, if I had the silver bullet answer to this one, I might not have to do my day job anymore,

[00:15:05] but really go and make some cash as a consultant of some kind. But in general, this is the same, you know, innovation versus risk seesaw that we've seen for years and years when it come to adopting any new technology, when it was the traditional machine learning techniques of the past and now when it's Gen AI.

[00:15:31] So the thing that is sort of most important is I think providing your team a safe and secure sandbox or design environment where they can really test out a wide variety of these new technologies. So whether those are different LLM providers or, you know, other techniques that may exist in the industry,

[00:16:01] it's benchmarking these different options, ensuring that they can be scalable, ensuring that there is a way, like we mentioned before, to monitor performance between the options or to assess fairness or applicability, view for bias, right? Kind of the standards we would want to do with any other AI project. But just noticing and noting that when it comes to Gen AI and the speed that you can actually apply these things, it can be more difficult

[00:16:31] and that the framework that you put in place first and foremost has to be there. I think three axes, if you want to call them that, that you can kind of review these projects on is cost. So having a way to actually trace and monitor the usage, especially of LLMs, which, you know, can really balloon in cost. Anticipate, manage, or spend there because a cost risk can be huge.

[00:17:00] The second is, you know, let's call it safety. So evaluating the requests and responses on your LLM usage, whether that's checking for sensitive information, checking for data abuse or leakage. And then finally, there's the quality element, which is like the output of the projects that you're building. Checking, you know, is the response objective?

[00:17:28] Is it productive really for the use that you're doing? And I think all of these things come down to the fact that at this point, we still need humans in the loop. So humans to build that framework that we want to work within humans to at some point document the process that's happened and also to validate that the responses, especially when they're external facing use cases are appropriate.

[00:17:56] And I was reading earlier today before you joined the call that you've worked with some pretty major brands like here in the UK, for example, the NHS and Aviva and help them integrate AI into their operations. And I think very often we struggle to see the evidence of real world solutions out there. So to bring to life what we're talking about, are there any examples of how businesses can successfully implement Gen AI without avoiding common pitfalls

[00:18:24] that maybe you can talk about and share today just to help them understand how it might operate in their world? Yeah, of course. This is the most exciting part of working here, I've got to be honest, is working across the wide variety of customers and getting to see the creative solutions they come up with and the ways that AI and Gen AI can impact their working life and then also their customers. And you mentioned Aviva, long-standing customer of ours, you'll find on our website

[00:18:53] and there's reference to a number of projects they've successfully used Dataiku to complete. So all across different organizations inside of the business. Most recently, they used some of our Gen AI functionality to perform customer sentiment and market trend analysis. So checking their customer feedback and being able to enhance their online experience, showcasing on their website what is most appropriate or interesting, making sure they have the products

[00:19:22] that are relevant and also the information about existing products. We also find stats like making their data science team five-time more efficient, you know, really cool things to see with a big customer like that. More recently as well, we've worked with a number of customers across Northern Europe. So Orsted, Fallender, MyTeams, Remit, and they've done some really cool work with generative AI where they basically have access

[00:19:51] to either 300 to 500 proprietary news sources every single day that they would like to take advantage of because they're in the green energy space, right? It's technology that's moving super quickly. They want to be on top of the trends. They want to know where to invest. But even with big teams of individuals reading and summarizing those, it's pretty much impossible to stay on top of it. So they use DataIQ and a variety of LLMs to do the summarization

[00:20:21] and the categorization and also ranking of the top articles. So there's an executive dashboard now that shows top 10 articles most impactful for their business that they can read every day, as well as a secondary option, which is for the entire company. I think it's top 50 most interesting or relevant articles. And in the first week of launching that, they had over 500 internal employees sign up to get the alerts, to get the information. And they said that it saved

[00:20:51] hundreds of hours of employee time. So they're quite happy about it. And maybe the last one, which I think is quite relatable to everyone, is the idea that now using DataIQ and our inbuilt like RAG chatbot functionality, you can actually set up a chatbot in under a day. LNG chemical implemented one so that their employees can actually talk to the chatbot

[00:21:20] and find out answers to safety regulations and guidelines rather than going through these huge manuals or internal documentation. So it's, you know, something that they had tried to implement previously, which was taking weeks and weeks and weeks and they were able to do it in just a couple of hours. So a lot of fun, fun and exciting use cases to be had. Awesome. And if we go beyond AI and AI governance, I know you're also a strong advocate

[00:21:48] for women in AI and tech. And it seems like great progress was being made and then more recently it feels like we've obviously gone a few steps backwards again. But from your experience, what are you seeing and what challenges are women still facing in the industry and what steps can organisations take to ensure that they foster more inclusivity and more mentorship, bring people along? Yeah, I am a huge advocate for women in technology in general and AI specifically, obviously.

[00:22:18] I am super lucky. I want to give a brief mention. My boss is a woman who has demonstrated just an insane amount of like grit, determination, but also kindness and empathy in her leadership style, which I really admire. And I think that gets to the point of you need these types of role models and people to look up to to really showcase what is possible.

[00:22:49] Women in the industry still face underrepresentation in the field. And like I mentioned before, diverse teams really make or break the output of an organisation's success. So it's not just about women being in the industry, but also the point of view and the education and the background that they bring and this ability to round out a team and have the effectiveness increased as a whole and as a result. So I think an increase in women in leadership positions,

[00:23:18] obviously mentorship, strong communities can all ensure better representation. And one thing an organisation can do is obviously assess the current situation, find any gaps and then decide how they want to approach it. That assessment is kind of the first step. Here at Dataiku, we have a women mentorship programme, which has been really great. And then just recently, yesterday actually, we took part in, I don't think the UK calls it Take Your Kid to Work Day,

[00:23:47] but similar in a way. And I actually had one of my colleagues' young cousins visit and she got to interview some of our data scientists, myself and a few other people about what do we do every day? What were the skills? What were the courses in uni that we took that were relevant? What did we wish we knew before starting? Things like that. And so I think it's lots of small actions that build up to the benefit

[00:24:17] that we want to see. And so anything that ORCs can do to support it, you know, I think that one was a pretty simple one to support and it was pretty impactful to see the way that, you know, the office experience resonated for the young people. I love that. And if we were to look ahead, are there any other trends that you see shaping the future of Gen AI and the enterprise in workspaces

[00:24:45] and how IT leaders and business leaders should be preparing for the next phase of AI? Because obviously, we're in the third year of Gen AI now. We're starting to see maturity and solving of real problems. But anything that you'd advise around this? Yeah, I mean, I think it's just a continuation and acceleration of the current trends. So one thing that's been really explosive is this expectation

[00:25:14] of AI and Gen AI being built into our lives. And I think that's going to continue, right? The consumers, employees, everyone expects these assistants and technologies to be embedded. And so I'd say that more specifically, it's going to be a push by and for the industry leaders to build and deploy agents into their everyday life. And so, you know, I think I saw a stat that was

[00:25:43] from Gartner by 2028, at least 15% of day-to-day decisions will be made autonomously through agentic AI up from what they're calling 0% in 2024. I think that's this critical piece, this desire for, you know, agents and specifically custom agents that is, you know, autonomous, LLM-powered systems that can work within your business to actually execute actions

[00:26:13] rather than just responding to human interaction. And, you know, that customization being even a step further. So rather than taking a generic AI capability from some vendor, the ability for businesses to create their own AI agents, maximize their competitive advantage and really invest in the tech that gets them there and then the people that work with that. So,

[00:26:44] we'll see. We can check back in a few years if I was right about it, but I'm all in on agents. I love it. Well, we're going to have to get you back on next year and see how things are evolving. And one other question I'd love to ask by guests before I let them go, you've shared your insights, so huge thank you for that, but I want you to leave one final gift, one more piece of wisdom, and that is a book that you would recommend that has inspired you or means something to you that we can add to our Amazon wishlist. But what would you like

[00:27:14] to leave and add to that list and why? Yes, I love this question. I am an avid reader, so it was hard to choose, but I thought in honor of International Women's Day coming up and sort of my greatest personal inspiration, I would advocate for a book published by my own mother. It's called Please Stay, How Women in Tech Survive and Thrive. And as she puts it, you know, women are often inspired to join the industry, but inspiration is not enough. So this

[00:27:43] book is all about workable strategies and techniques that make women actually thrive in the technology space, whether that be software, hardware, AI, etc. And then she wrote it with her co-author Deborah Christmas, quite a name I know, and their mission is really to provide the advice and to create a community of like-minded professionals so that we can all thrive together in this exciting tech space. Beautiful. I'll get

[00:28:13] to find out more information about DataIQ, connecting with you or your team, etc. Where would you like to point everyone? Yes, I would point them directly to the DataIQ site. We've got a ton of great resources there and we also host a number of events all over the world. So if you are looking to meet with us in person, we'd be happy to engage there. Fantastic. Well, thank you so much for sitting down with me today discussing the steps that

[00:28:43] IT leaders and business leaders need to be taking to avoid the common cost and risk pitfalls that come with implementing Gen AI and also why businesses should adopt guardrails that are designed to address some of these challenges, mitigate those common risks when building, deploying and managing Gen AI in the enterprise. We often just hear about the buzzwords and the quick fixes but it's great to dig a little bit deeper on this today and a language everyone can understand. So thank you for joining me.

[00:29:13] My pleasure. Thank you so much for having me. I think AI is evolving fast but as Emma Irwin highlighted today in our conversation without proper guardrails businesses risk running into serious challenges whether that be spiralling costs or biased outputs and let's be honest scaling AI without governance is a recipe for disaster. Businesses need clear frameworks for access,

[00:29:41] discovery and indeed performance tracking. But ultimately bias in AI isn't just a technical issue either, it is a human one and you're going to need diverse development teams to play that critical role in creating more ethical and responsible AI solutions. But what is your organisation's AI strategy? Are you building with the right safeguards in place or are you just diving in without a plan, just innovating and not worrying about safety and

[00:30:11] security? Let me know. All concerns, raise them all. techblogwriteroutlook.com LinkedIn, Instagram, X, just at Neil C. Hughes. Send me a message on any of those platforms. But that's it for today. Thank you for listening as always and I will speak with you again tomorrow. Bye for now.