Shipping AI at ING: From Pilot to Production – with Claire Pfeil & Frank Fischer, ING

00:00:00: Often you have this risk of a competition.

00:00:03: And that's what we really try to avoid, because if your going into it... We will definitely fail!

00:00:09: We aim for collaboration between the global teams and local team.

00:00:14: It is also empower both sides and get best out of them.

00:00:34: Hello & Welcome To The Data Culture Podcast.

00:00:37: I'm Carsten Bange Founder CEO Of Barg & with me is again Florian Wiegelmeier is signaling you that we have another data festival special edition.

00:00:46: We have again two guests of the Data Festival here in the podcast today, it's Claire Pfeil.

00:00:53: she is the COE Lead for IT Analytics at ING in Germany so a local unit.

00:01:00: and then we are Frank Fischer.

00:01:02: he is the CEO lead for analytics engineering at INGE Globe!

00:01:07: And uh...we get some insights into their daily practice aren't we Florian?

00:01:12: Yeah, absolutely.

00:01:13: We will have a look on the question how can you actually mature and create impact from Agentec AI and generative AI in general which means How do we set up proper practices?

00:01:26: What are the upsides and downsides of those models And how to integrate them into processes that AgenteC AI promises?

00:01:33: Also the questions about bringing business experts alongside with

00:01:39: us

00:01:39: to make that all a success for the complete company.

00:01:43: Yeah, so lots of relevant insights and topics.

00:01:46: enjoy the episode!

00:01:49: Hi Claire hi Frank great to have you on The Data Culture podcast.

00:01:53: Hello thanks for inviting me.

00:01:55: Thanks Kastan indeed Great to be here.

00:01:57: So the first question we would like to ask you in this podcast is actually What is your expertise and what are you working on at ING?

00:02:06: And we heard that you're working on chatbots, Chantic AI, Generative AI.

00:02:12: What's like the role in that for example analytics experts or how does it work together with the expertise of data scientists, machine learning expert software engineers

00:02:25: etc.?

00:02:26: I would try first answer You know, when it comes to AI and the modern solution that offers.

00:02:36: It's nice to try them out.

00:02:37: I don't what they can do.

00:02:38: but then in a second step you have to make sure you get results and profits your organization in your day-to-day work.

00:02:45: And thats where we both come because we let say integrate models into the ING infrastructure meaning building data pipelines ascending that necessary data to the models and then, the results of the models.

00:03:06: To the systems in order use them?

00:03:08: And I think since both Claire and i are in tech you also have a different angle on what this means at the end right especially these days.

00:03:17: everyone can do analytics with the LLMs out there but really bringing something into production.

00:03:24: thats where tech is required.

00:03:27: That's also where we come into play and at least I think, a crucial role with our teams that we have.

00:03:33: Yeah absolutely sounds like it!

00:03:35: We see in this discussion.

00:03:37: if you look at the agenda of The Upcoming Data Festival there are lots of organisations exactly on these steps... Many know what they want to do but now its really about operationalising, scaling or bringing to production as mentioned.

00:03:51: so super interesting topic.

00:03:56: But before we do that, maybe I can give us a bit of an idea for size.

00:03:59: So how many people are overall may be working with Data Analytics AI and then specifically AI?

00:04:07: What what size of organization is behind

00:04:11: it?

00:04:11: in my global team We're about one hundred twenty people intact And there's many more on the analytics side.

00:04:18: so we have this matrix organization at IG Of course where we have tribes containing customer journey experts data scientists, data analysts and engineers.

00:04:30: These are separated also in the hierarchy lines between CO and CTO.

00:04:35: The numbers I can tell you a little smaller.

00:04:38: our team is about twenty experts but like Frank said it's hard for me to tell them number because right now really every department is exploring AI.

00:04:56: It's almost everybody.

00:04:58: Not only the customer processing, also internal things software development how to test your software lots of questions that people try to solve and optimize using all kinds of AI methods.

00:05:15: We will go a bit into it during our presentation.

00:05:18: How we organize through this especially in an iterated way?

00:05:25: Yeah, I think that already points to like maybe a little glimpse too what you may be saying in Munich then later.

00:05:34: So IMG is obviously as he already said A global bank and multinational with a lot of local teams And the strong office probably in the Netherlands.

00:05:46: Could give us an idea on how it can be balanced topic of having a strong global team and also still innovating locally.

00:05:57: We tried to innovate an experiment centrally in the beginning, especially when GenEI and Agentec AI appeared right?

00:06:05: we also need to know what is relevant for us.

00:06:08: how much can be do How much ready you get out this And fast?

00:06:14: Can we then also Get that value out it and distribute further into the organization?

00:06:21: Hmm, so there's this topic of balance between central and let say local decentralized tasks as functions.

00:06:28: And ownership in governance is always a big topic.

00:06:33: Did you find it?

00:06:34: A good solution for that would I

00:06:38: think, you know...I've been working now in the data science field for some years.

00:06:44: When we started it wasn't even called Data Science but Data Mining and there were no data scientists who just started from scratch.

00:06:52: And since that time, I would say we are constantly looking for the best way to deal with things.

00:07:00: In there always centralized models competing with more decentralized models.

00:07:05: you can look at the data.

00:07:07: We were dreaming of one huge data warehouse a single point-of-truth where all kinds of data was stored centrally and All You Had To Do Is Just Go There and Grab The Data.

00:07:18: Did It Work?

00:07:20: Not Really.

00:07:21: Does it work when everybody has just his data as he wants to every local entity of ING, for instance?

00:07:29: Stores the data anywhere they want.

00:07:32: That doesn't work either because you want to be able to connect the data and use them in a sensible way.

00:07:39: so I think that there always have been a mixture between both.

00:07:44: You also have advantages on the floor both models.

00:07:49: So I would say it's a constant search for the best way to do it, and Frank and I we are trying to find this sweet spot of.

00:07:59: let us say what you have to centralize?

00:08:02: What can be hand over to local units... And also how can the local unit maybe help the central unit in their experiments.

00:08:16: Building something special for a country helps, although the global organization is building more in general.

00:08:24: It's constant exchange and search for that sweet spot.

00:08:30: I would say so will we find this sweet spot?

00:08:37: We are going to get near it.

00:08:41: Maybe there was some remark because what Claire mentioned often you have this risk of a competition.

00:08:49: And that's what we really try to avoid, because if you are going into this... We will definitely fail!

00:08:55: We aim for collaboration between the global teams and local teams.

00:09:00: It is also empower both sides and get best out it.

00:09:04: Yeah I think its definetely something that adopts in changes Because also maturity change on one side.

00:09:13: use cases change, they're constantly innovative new ideas and technology changes also.

00:09:18: Sometimes there are also no possibilities.

00:09:20: so I think this will be an ongoing state.

00:09:24: maybe that's the takeaway here yeah?

00:09:26: That a fixed date might not be the right answer or one answer to stay for long time.

00:09:34: Let's look into your core task in a way.

00:09:41: What did you find are the biggest challenges?

00:09:44: In trying to bring AI into production, into the organization and as you said Claire at the beginning really generate value of AI?

00:09:56: I think it is not technical thing, it's more an organizational thing because I think we have to get used a new way of working with AI.

00:10:07: When i learned AI there were data scientists and their role was very clear.

00:10:13: They had to build new models, and then they were engineers putting them into production.

00:10:18: nowadays that the large-engaged models already very far in their development.

00:10:25: I don't suppose we will try to develop this kind of model by ourselves?

00:10:31: So the roles inside an organization have to be refined.

00:10:36: so how will IT and data scientists at business departments actually work together to bring one solution into production.

00:10:49: And also maybe the second thing, AI is very great at some things but it's not the one tool to solve every kind of problem and I think that's a mindset topic.

00:11:03: we have to learn to use models for what they are meant to be used.

00:11:11: Any kind of issue that we have?

00:11:14: Yes, indeed.

00:11:15: And like you said in the beginning going from experimentation to really production ready state.

00:11:22: That's I think the biggest challenge because especially in highly regulated environment Like ours comes with a lot of additional challenges.

00:11:30: So not just how to deploy it How to embed it into an application landscape But also, what are the risks?

00:11:37: How do we basically cover or mitigate the risk.

00:11:39: Do we monitor how to address all regulatory requirements and what does this in the end mean for use cases that might have in mind but potentially not feasible because of it?

00:11:51: Yes actually if we're trying now doing our first steps using agentiqi in actual processes And I would say that's a whole different animal than using AI for analytics purposes and maybe marketing purposes, because you're actually giving AI then the responsibility of making decisions.

00:12:15: What we have to do according to the regulator in the EUAIAG law is being able explain why an AI agent decided this or that outcome.

00:12:29: Are we going to allow a client take on the mortgage or are not?

00:12:36: That's an issue.

00:12:38: I would say that it is not really solved yet, to be tackled automatically.

00:12:45: We can use one AI agent looking after another and then have some kind of control And do manually maybe take samples and see what they did in this case But its not yet solved very generally.

00:12:59: And that's still one big thing we have to work on.

00:13:04: Yeah, you mentioned this requirement especially for financial institutions or explainable AI and now we see it.

00:13:10: in the other hand a lot of technology being used are more black-bottom models.

00:13:17: I mean GenAI is not really explainable, neural networks aren't really explain able.

00:13:22: so does that mean how do you treat that?

00:13:26: Do restrict the technology.

00:13:28: So you say, okay for like a credit application I can only use models that are explainable.

00:13:35: or how do work with it?

00:13:38: I wouldn't put in this way because if we tried to use model models that were explainable and nothing else then we would not be able even explore the LLM possibilities.

00:13:49: From what I see, we are doing our first steps.

00:13:54: And as long as we're doing the first step... ...we can still control models manually and have humans watch the model.

00:14:06: For future, what is starting to do now?

00:14:11: We start collaborations with universities.

00:14:14: They are also doing research on that specific topic so they can hopefully help us get a grip.

00:14:24: And which role does all the topics around monitoring and observability play in bad regard?

00:14:31: Because we see like, a lot of companies these days investing a lot money or tired to set up telemetry for a gender gay eye.

00:14:40: understand what happens in which step.

00:14:43: Is that also something that helps you, explain what models and agents are doing?

00:14:51: Yes to a certain extent It definitely helps for monitoring And we use it as kind of a challenger.

00:15:02: We call it LLMSA Judge where we have models challenging.

00:15:07: Let's, for example take the chatbot as an example that we have in some countries already based on LLM.

00:15:13: All of our conversations are also actively monitored and challenged if there was a behavior where you don't want to be signed towards your customers And then we've got few other controls afterwards As well.

00:15:27: which means at this point Have a look or do random checks on what happened in these conversations, maybe label specific pieces of conversation that need to be better next time.

00:15:40: Or go the wrong direction?

00:15:43: So we have multiple layers and lines of defense especially with those customer facing systems.

00:15:51: But also for internal systems, we always try to have all the categories in place that they can imagine.

00:15:58: We have a lot of tests and again the LNM is a judge.

00:16:01: And

00:16:02: which role in these multiple lines-of-defense does then human play?

00:16:07: For example if you talk about a gentekai You're talking about processes where you have subject matter experts usually like a bank clerk or someone else.

00:16:18: Does it also need some new skills?

00:16:20: How do you approach

00:16:21: that?".

00:16:22: That's something we need to learn.

00:16:23: So they are definitely part of the project because, of course ,they have their knowledge.

00:16:27: They're all so hard ones To assess whether output is correct or not and in The end We learned right especially a geneticist.

00:16:37: You really knew still even Jenny Iis relatively new steel And when he to learn how much can be rely on that?

00:16:44: how much does it help in the business?

00:16:46: those innovation related topics always a change.

00:16:49: And change also means, change for people right?

00:16:51: Change like you said they are skills and so on but change comes with concerns.

00:16:58: often that's something important needs time.

00:17:03: we need to address trainings or general education of the topic what it means for their day-to-day routine, and how they can benefit from it.

00:17:14: How they can be more efficient or even more effective in what they do.

00:17:19: if we assume back office people that could help them to make better decisions.

00:17:28: Same with sales,

00:17:29: e.g.,

00:17:30: how are they accurate?

00:17:36: We have some internal... AI projects.

00:17:40: also, I can give one example when there are very let's say complicated regulations for different countries and maybe internal IG regulations.

00:17:53: And in addition to regulations by law, then it can be overwhelming for a clerk to make the decision.

00:17:58: When they onboard a customer what kind of due diligence do they have to do?

00:18:02: In there we develop the tool that can help and let's say guide employees through this document jungle and extracts what is relevant for special use case.

00:18:17: but yes you are right Everyone has to learn about what can I do and my responsibility in the process.

00:18:27: Up-to-now, a clerk was responsible for making customers happy giving them their right answer.

00:18:34: now they still are because not every use case or request is answered by a chatbot.

00:18:39: sometimes people want to talk with humans.

00:18:43: we should allow that.

00:18:44: additionally they're responsible So they have to know what is an appropriate answer.

00:18:52: How can I correct the system?

00:18:55: It's interesting, so you're mentioning that you've developed a data and AI culture internally.

00:19:01: You already mentioned training as part of it.

00:19:04: What else do we do on board with your colleagues or develop more people embracing AI really using it in the end, generating more benefit out of it.

00:19:19: What's your approach there?

00:19:22: Communication and I think that is all about communication.

00:19:27: We also offer Microsoft Copilot.

00:19:29: maybe not everyone has it but you can play around with it right.

00:19:34: And thats how we start to learn what are advantages.

00:19:36: so people get familiar.

00:19:40: The level of fear might decrease a bit because they see the advantage and what it can do.

00:19:46: And how this helps them to be more efficient in their work, so personal productivity can increase faster... get away with boring tasks or at least help from that again.

00:20:04: then also under conditions data protection and all those things that we need to keep in mind.

00:20:11: So, that's also an important part right?

00:20:13: You cannot just communicate all the nice advantages benefits people get but what are other risks connected with this if you use LLMS Right?

00:20:22: Yeah I completely agree.

00:20:23: communication is always cornerstone yeah To develop a culture With you internally.

00:20:27: What Are The Most Effective Ways To Communicate?

00:20:31: How Do You Reach The People?

00:20:32: Or What Did You Learn?

00:20:36: Okay, so maybe I can try to answer this one from my experience not only in the bank but also as a consultant.

00:20:46: What does that always work?

00:20:47: is when an expert tells everyone else what-to do and what-not-to-do?

00:20:53: That could be at the beginning of something... ...but it won't very successful.

00:20:57: if you want to reach many people.

00:21:00: But let's say user groups that are exchanging their experiences in a more informal way, that helps.

00:21:09: For instance like Frank mentioned with Microsoft Copilot there is some people who just created their special agent solving an email problem they have or anything.

00:21:23: I wouldn't be able to state you these not as said but then when i was on this round and noticed These very specific examples created high interest in many people and then they exchanged their very specific ways of working.

00:21:42: So that seems a really good way, not having one expert telling everybody else what to do.

00:21:49: but as now AI is more democratized than before.

00:21:54: also the communication can be more democratised.

00:21:58: Very good experiences, thank you.

00:22:01: Let's have one question in mind since you are really at the forefront of building this deploying this etc.

00:22:09: we hear more and more people in your roles really struggling with a speed off.

00:22:15: technical development that you need obviously needs certain time to develop something roll it out educate And halfway through that process, you have already completely new capabilities in JNAI foundation models or on other systems.

00:22:33: How do you work with it?

00:22:34: First of all is this a problem also for you?

00:22:36: Secondly what did we do about it?

00:22:38: Yes definitely!

00:22:39: That's the topic.

00:22:40: and another challenge so what we not to always immediately jump onto next model version even still... Usually immediately start to assess.

00:22:54: So we have kind of a process where you assess, okay Does this model work?

00:22:59: Will we have performance issues or whatever?

00:23:02: so at some point it will be white listed for further use.

00:23:05: It comes with another challenge because often The newest version is the one being supported Also from compute power perspective.

00:23:14: that's what he experienced which means if You are still using a lower version You might have less resources you can count on and being in a production system then already, that of course can trigger lot issues as well.

00:23:33: If the compute power is not there anymore because yeah... ...you are behind with your model version

00:23:38: Maybe I would give more pragmatic answer or more pragmatic approach.

00:23:42: Because if we had ten use cases to accomplish on our project list And if you are doing one use case over and over again, because your always trying to use the very newest technology that is on the market.

00:23:59: Then you have one use-case with top notch maybe or don't ever finish it... ...and then there's nine more use cases which didn´t work.

00:24:08: just used what was best for now see where it takes you and learn for the future.

00:24:17: So I have a more pragmatic approach.

00:24:19: Yeah, but i like the appeal just get started with it because if we don't get started uh... You can wait forever until your AI is solved

00:24:29: in no way.

00:24:30: Like We said earlier The real issue Is not to use always the newest product But the real challenge is To be able Put it into production and get results out of it.

00:24:41: And the pipelines you built will probably work for the newer version of the model, so that's the main issue I

00:24:51: see.".

00:24:52: an ML Ops process already, maybe something you are building on which she had earlier because I mean we have like data ops in the past.

00:25:06: We've got ML OPs and also the software engineers obviously has their DevOps experiences.

00:25:12: so is there something that you could learn from?

00:25:15: And does this help with all of these changes to your background?

00:25:19: Like i said a little bit regarding the LLMS in general right What we try to combine in the end that be also use Yeah, basic standards of software development.

00:25:32: when it comes to embedding Analytics no matter if classic analytics or Jenny I and so on into a production environment right you cannot just deploy it and then hope for the best.

00:25:50: You really need everything in place, you also need the standards in place controls in place... ...you need to do your risk work and so on and so forth.

00:25:57: So from that perspective we just adapt everything there was there in the past And try fit new technologies as possible into this because otherwise I would be quite concerned right?

00:26:16: We

00:26:18: didn't hear the word agent yet which surprised me a little bit, but obviously that's for many organizations.

00:26:24: That is next step in AI to

00:26:27: be honest.

00:26:27: sorry

00:26:29: then I want her here more about it.

00:26:31: so what?

00:26:32: What's going on?

00:26:34: first of all yeah Is It On Your Roadmap or it is Obviously But How Do You Treat It?

00:26:41: how do you approach it?

00:26:43: Are there certain use cases, you find it like for first steps being especially easy or beneficial to us them.

00:26:51: Let's hear more about your approach to the whole topic of a genetic AI?

00:26:56: I will give general answer.

00:26:59: It is definitely on roadmap but as Frank also explained earlier its always trade-off between new technology and safety issues especially when it comes to agenda AI, where you have too much agent talking with each other and taking decisions.

00:27:18: It becomes very untraceable.

00:27:20: so we need to put guardrails in place And make sure that we know what kind of decisions or freedoms one AI agent has.

00:27:32: I would say it's innovative yet still a safety-concerned response.

00:27:40: My specific question would be, did you already find an idea or framework on architecture?

00:27:47: I think one of the main tasks is now combining probabilistic and deterministic workflows.

00:27:52: That's for many right now that you can have this pointed agent maybe take certain decisions but we had to model basically taking a decision or preparing it.

00:28:06: But like if you think about more complex or bigger workflows, then typically we have probabilistic and deterministic parts of it.

00:28:14: And that's where I see a lot of things happening especially also in banks yeah?

00:28:17: That they start to combine now workflow engines rules based approaches plus JNAI-based things.

00:28:26: is there something that were already working on?

00:28:29: yes so we have projects looking into this.

00:28:34: We also approach it step-by-step.

00:28:35: So like I said, having already a landscape with multiple agents being involved that would be quite complex for the first step.

00:28:45: so starting with one agent That comes with what you expect from an LLM right?

00:28:53: But then have APIs in place where you just trigger the action.

00:28:58: i think thats at end What makes difference.

00:29:03: When it comes to the agent itself, most of the magic is done by cloud providers I would say.

00:29:08: And that's an important part.

00:29:11: we also have a lot of software engineering.

00:29:14: basically you need your backends in place and processes clear not just tech but data as well.

00:29:20: The process itself needs to be in place.

00:29:23: We do kind of readiness assessment too.

00:29:28: see if our process gets ready rebuilt or optimized via agendic flows?

00:29:38: Or not, because otherwise if the foundation is not there you can try a lot with agents or GenEI.

00:29:46: Likely put that your successful.

00:29:47: it's not very high

00:29:48: now.

00:29:49: did you find certain type of class processes being especially suitable for agent usage?

00:29:59: or maybe a certain area of the business, more customer facing on my internal or regulatory.

00:30:05: Is there something that sticks out as pattern?

00:30:08: We're just at the beginning.

00:30:09: I would say so.

00:30:11: we have few areas where many steps are involved and you can translate those steps into something agents could do then right again when they also have data in their back end available or at least where you can develop it.

00:30:32: I could give maybe a personal view, but i would like to try out agents when it comes to the governance work and documentation work our engineers have to do.

00:30:49: Most of needed information that is needed won't be able put into necessary forms Instead of engineers doing it manually, I would like that to happen.

00:31:02: But i don't think there is any boundary.

00:31:04: where can we use the agent?

00:31:08: It's a domain or specific topic... What risk are you willing take and how risky will be used by an agent and how much benefit do they create?

00:31:24: whatever you try to do and automate with an agent.

00:31:28: Of course, when it's customer facing You have more risk can also probably More profit.

00:31:35: so we always have to do this assessment And that's not an IT thing That is more a business things.

00:31:41: So It's Not really my place To answer This specific question.

00:31:48: Maybe you contribute to the answer, because that's another big topic for many.

00:31:53: How do we measure and communicate the value of AI?

00:31:59: We see more and more organisations coming under pressure... ...that they also have to now show their impact on investments into AI.

00:32:06: I'm sure it is something you are confronted with.

00:32:12: What can be done there?

00:32:14: So if we take the chatbot example Of course, we also measure.

00:32:18: how does it land with the customers?

00:32:20: What's their deflection rate.

00:32:23: How many cases does the chatbot solve basically itself?

00:32:27: or do they still end up in a call center at the end?

00:32:29: Or are customers happy afterwards what the chatbots gave him and how he communicated And so on?

00:32:36: That of course is also in the end difference.

00:32:38: pro-projet How can KPIs look like Basically...what you want to measure?

00:32:44: It isn't always the same And we always try to do a bit of, yeah.

00:32:49: A discovery phase in the beginning to see how much can you get out of it and smaller scope?

00:32:54: How much is able to scale at end before start just big project that consumes lot effort money time.

00:33:04: To be honest It has been.

00:33:06: with AI machine learning You as data scientist were building model As best you could and then, oh maybe it's not what I was looking for.

00:33:16: Then start to look at new data to use better information.

00:33:21: like Frank explained about the KPIs.

00:33:23: we have to decide What do even want measure?

00:33:27: Like said earlier It is way forward.

00:33:30: We need to start somewhere and improve on a way.

00:33:33: Always same answer i'm afraid.

00:33:36: At this stage, we've got bad news and good news.

00:33:41: The bad news is that our episode is slowly coming to an end but the good news it's that if you can continue your conversation on June sixteen and seventeen at the data festival in Munich We are really happy to have you there!

00:33:58: You already gave a glimpse of the organizational perspective.

00:34:04: So I guess you're talking a bit more about that on in your talk, but maybe it can give us a quick teaser.

00:34:10: Quick outlook what?

00:34:12: What uh Can participants expect?

00:34:15: You will see Frank and me working together locally and globally And find common ways the common ground.

00:34:23: This isn't the end.

00:34:24: It's not just about data processes and technology is also about like we said About people and the organization you build around it In order to be successful at The End And that's also what we want to share.

00:34:36: That sounds super interesting, I'm looking forward it!

00:34:38: So see you soon in Munich!

00:34:40: Yes, see you in Munich.