Webinar: The AI-first workflow redesigning for growth

Great. Still people coming in. So, yeah, let's kick off. So, I'm very happy to be here today to tell you about um how AI is impacting knowledge work and how we can redesign it in order to you know, get the most of the value from artificial intelligence and in order to think through this impact, we will draw from an analogy uh from the factory floor, thinking about what happened in the factory floor when robots came in. Okay, so I am Stefan Ohran and I'm managing director at Board of Innovation. I focus on applied artificial intelligence. I am a physicist that moved from the study of the universe to working at the intersection between AI, innovation, and strategy. So, over the years I've uh built and led the development of AI solutions that help organizations work differently. And yeah, so this is what we're going to talk about today. And Board of Innovation is an AI transformation studio. We focus on growth, so we are very interested in understanding how AI, artificial intelligence, can unlock growth for organizations. And we do this across three big topics. First of all, uh value creation, looking at where AI is changing the value equation of business, thinking about products, services, and the business models around them. How AI is changing that, what are the new opportunities. Then thinking about work redesign, so how is AI changing knowledge work and work in general, and this will be the topic of today. And operating model, so how can organizations organize themselves in order to get the most out of all of this, in order to identify the right opportunities and to scale them. And I would also like to invite you to the this year's first edition of the Autonomous Summit. So, this is going to be an edition focused on the topic we call it Obsolete, so it's going to be focused on the transformation of work through artificial intelligence, so also very relevant and close to the topic that we will be discussing today. And yeah, you are invited. You will have you can have a free ticket. So, please join that summit. It will be a great event. Okay, so let's jump in. So, I would like to start with an analogy. Think about a factory in the 1950s. So, it's designed this factory is designed for human work. It's all the layouts, all the tools are made for humans. Now, enter in that factory a state-of-the-art robot. It's able to do wonderful things. But what happens if you just drop the human in that factory designed for humans? Well, what will happen is that the robot remains idle a lot of the time. It's going to be either waiting for inputs from humans or waiting to provide its outputs to a human. It's going to waste a lot of time moving in a layout that is not designed for it. And also it's going to have to rely on quality inspection that is designed for humans where we check that the quality is correct many times during the process because the process takes a long time and it's expensive to restart. And the same thing is happening in the office today. We have state-of-the-art robots that are entering our digital work environment. Of course, generative AI and these agentic systems such as Cloudbot that are enabling end-to-end automation across multiple systems, really changing the frontier of what AI can do for us. And in the same way as in the factory, if we don't rethink the way we work, then we're going to end up in a situation where AI is doing some stuff that is clearly useful, drafting documents, analyzing them, helping us, but still a lot of the value remains unrealized. Right? And the problem, of course, is not the technology. Uh the problem is not AI, the human, of course. The problem is

not the human, of course. The problem is the layout. So, we need to find the right layout for optimal collaborations between human and machine in the context of knowledge work. Now, a first poll, I'd like to hear to hear you a bit. So, be honest, what happened with your last AI initiative? So, you have a couple of options. You might wonder which initiative. If maybe you don't you have not had any AI initiative in the context of your work. Or maybe it worked great in a demo, but nobody used it. Or people use it, but it didn't really change much. And or it's saving real time, but only focus on individual tasks. Or maybe it's you actually redesigned work around it. Okay. So, I'm trying to understand how to see the result, but I see in the chat a lot of E's, so that's great. That's actually great. Some C's. Okay, D. But okay, I don't see any A. That's a great thing, so everybody had at least one AI initiative at work. Okay. Okay, good. Thank you very much and that's a good segue to the next slide. So, today organizations are at a crossroads. There are two very different paths. One is the AI-enabled path where artificial intelligence is retrofitted on top of the old way of working. So, that allows fast adoption. It allows to find efficiencies and perhaps strike a balance early on. But over time it doesn't really scale and it doesn't lead to a lot of benefit. A lot of missed opportunities. So, that's the robot that enters in a factory that is not redesigned. And the second road that an organization can take is the AI-first road where work is redesigned. Where they rethink growth, operations, and competition and they build the business with AI at the core. And this is, of course, where the the winners, the organizations that get the most out of it will be. So, what is an AI-first transformation? So, AI-first transformation, it's quite clear, it's about redesigning how work is structured, governed, staffed, and performed around the technology's capability. So, you can you have here on the left an illustration of an AI-enabled approach. So, you can see that the human is very much at the center. So, the human is the engine. And it's using the human is using the technology to speed up some steps in the process. It could be a draft of an email, it could be drafting a document. It could be a calculation or whatever. But the technology here is being used by the human. The roles didn't change. And it's accelerating, let's say, specific tasks. And the AI-first approach is turning consistent turning this around. The engine is not the human anymore. The engine is the technology. So, in this case the engine could is artificial intelligence and the human is the driver. So, the human intervenes at key moments, but is not driving the process end-to-end anymore. To make this a bit more concrete, consider an RFP response process. So, that's the process of receiving a request for proposal, which is relevant to many businesses, many industries. So, you receive your request the request and then you need to prepare a response. You need to make an offer. So, typically this is a multi-step and fairly complex process. Uh the it starts with receiving the RFP. So, this can be a passive reception where the RFP comes in an inbox or it can be active with some business development involved. Then once the RFP is received, there is a context gathering uh where you collect information, you extract it from the RFP, you identify information inside the organization. Then you need to qualify. You need to see, okay, can you respond to this proposal? Can you uh do you have the right capabilities? credentials? Do you meet the selection criteria? You need to do some compliance checks and so on. And then

finally, you're ready to take a bid no bid decision. This decision typically happens in a meeting with some senior leaders. And then once if there is a positive decision the proposal is defined, the solution, how do we do this? You need to think about pricing, you need to draft the different sections, quality review and submission. So, a quite cumbersome process. And the AI-enabled way consists in not changing the process, but accelerating with AI. And there's a lot of things that you that AI can help doing, extract information, summarize requirements, do the compliance checks, generate drafts, the pricing and so on. But, you're still following the same linear workflow, you still have the same SME heavy model, you rely on you know, the same team to drive things and to approve and so on. And this means that you know, your head count still scales pretty much like it did before. And you can have faster drafts, but in the end marginal gains. So, let's say that you went, okay, marginal. This could be a difference, you know, from 2 weeks to 1 week to respond to a proposal. Now, every organization is different and the timelines change. But, just as an example. Now, consider the AI-first way. Here, the AI is executing and the human is judging and differentiating. So, the process changes. Now, we can think of this in phases. So, the first phase is the intake phase. Instant intake, let's say. And this is a fully autonomous phase where the RFP is parsed, the requirements are extracted, the AI cross-references the CRM and past wins to get the right context. It generates all the scorecards, all the elements that need to be taken into a decision to bid or no bid. It pulls the relevant context, runs the compliance checks and so on. And this can be performed, let's say, in 1 hour. The second phase, and you don't even need to take a decision yet whether you want to do it or not, you can already start drafting because it's cheap to do it. It doesn't take a lot of energy. effort. And here you can have an AI plus human in the loop process of parallel generation where the proposal is prepared. The drafts of all sections are done. The pricing model is built based on past data and so on. And this can be, again, a fast process of 1 hour, let's say. And then you have the third phase, which is where the human finally steps in. So, here you are reviewing and submitting. So, here you're making sure that everything is fine and that you are comfortable and ready to submit. This is a process that will require a meeting. This might require 1 hour or more, depending on the quality of the proposal that arrived to the human reviewer. And so, again, AI is assisting in performing some checks post facto, consistent consistency audits and so on. So, you can imagine you can see here how the process is completely changed, right? So, now coming back to the factory. What can we learn from what happened in the factories? Because the factories went through this quite a long time ago and today the way products are manufactured is very much robot-driven, right? So, what are how did they do that? So, let's take three we'll take three principles, three lessons that we can learn from manufacturing. The first is value stream mapping. And this consists in looking at the process and mapping the flow in the context of manufacturing, it's the flow between workstations. What is What are the inputs and outputs of each workstation? What are they waiting for? In knowledge work, it's about mapping the decisions and the information that goes into those decisions. What are the decisions waiting for? Then, the second principle is task allocation. For each task, you need to decide, is this going to be a human or a machine? And this is not a single dimension decision, this is a multi-dimensional decision. So, you need to define the dimensions that you want to consider. And of course, these are different from you know, manufacturing to knowledge work, but still it's a multi-dimensional task allocation problem. And finally, I we borrow a principle from Toyota, the Jidoka principle, which is about

the human role. So, how do you escalate to a human when the machine makes an error? How do you decide when there is an anomaly and how do you do What do you do? bring the right human in at the right time? And that's a very important principle to design that appropriately. Of course, knowledge work is not a factory. There are also differences, so let's highlight three differences that matter. First of all, quality in knowledge work can be ambiguous. In a factory if an item, a product, a produced item, a produced unit is of good quality is quite, let's say, clear. You can inspect it and you see if there are defects. Or, you know, it comes back after it's been shipped because a customer complained that it doesn't work. In knowledge work, it's more ambiguous. It's not it's less clear, right? So, that's one challenge. And that means that allocation is harder, but perhaps more important. Second, workflows can be chaotic. Factory work is generally linear in generally linear and predictable. You always produce the same thing and you do it in a similar way or in the same way, exactly. In knowledge work, the outcomes are more varied and also the process of creating those outcomes is more varied and it can go off script much more often. And so, that's clearly a challenge and that means [clears throat] that the human role here carries more weight and it carries weight beyond the specific function because it's also about judgment and transversality. And then the scale is different. This is about the outputs that are produced. A factory line produces always the same outputs many times, but always the same. Knowledge workflows produce many different outputs. That means that the economics of redesign change. So, set setting priorities is more consequential when deciding how to redesign. Now, let's think through how to move from a workflow to an AI-first workflow. So, we looked at the example of the RFP, so how could we get to to this sort of redesign. But first, I'd like to again to ask you a question. So, when you introduce AI into a workflow today, what does your process look like? So, A, you pick a task and you plug in a an AI tool and see what happens. B, you map the workflow first, then decide where AI fits. C, you redesign the workflow around AI from scratch. And D, you don't really have a process, it depends. Yeah. Okay, I see a spread but a lot of B's. Okay, some A's. Yes, B. Not a process yet. I think the not having a process is quite common because in the end this is, you know, this happens organically. It it's teams auto-organize themselves with the technology. And you know, different processes can be followed by different teams within an organization. Okay, thank you. So, how do we go from workflow to AI-first workflow? So, here we have a six-step methodology. The first phase is the mapping phase with step zero, which is about setting the priorities. This is, you know, deciding what to focus on. What which one Which is What is the workflow that you want to redesign? And then step one is mapping the decision flow, as we discussed, one of the key principles. So, tracing every decision, identifying what comes into that decision and what the decision is waiting for. Then, the second phase is the allocation phase. So, here you step two, define the constraints and allocate the work to the right agent, human or machine or, you know, some collaboration of human and machine. And then step three, you define the escalation logic. So, that's the Jidoka principle. And then there's the operation. Step four, designing the human role. So, the the human moves from executor to act and judge, and step five, measure the process and reallocate it. So, it's not a one-time exercise. This

is something that should be repeated as the technology evolves, right? So, it's very important actually. Step five flows back into the beginning. So, the first step, prioritizing, is about choosing which workflow you want to redesign first. So, that's a let's say a prioritization exercise, and you can map this in a 2x2 matrix, so very familiar. You have on the one hand the strategic leverage, on the other hand the readiness, and you want to identify the opportunities for which you have high readiness and high leverage. Designing deciding what to redesign is a very strategic question. So, that's you know, there's a lot to say about this. We won't deep dive on it, but this is where you can basically build a moat for your business. By redesigning the right process, you can reinforce your advantage, and you can make it more difficult for your competitors to imitate you. Now, moving to step one. Here, this is about mapping the decision flow. So, you want to understand what goes into the decisions. So, in order to do that, so this is the my value stream mapping principle. For each decision, you want to understand who makes it, what information feeds it. You want to map the transitions, not just the decisions, and this is how you reveal the real waste in a process. For the RFP, our guiding example today, the key decisions are the bidding, whether to go for it or not, the pricing, and you know, the quality. So, is this the quality that we want, or do we need to reiterate to iterate again, and then the final sign-off. So, and each of these dimensions is waiting for some information. So, mapping that is important. So, that's for the decision flow. Then, moving to the allocation. Now, here, we want to understand which agent will perform which actions. So, we have four questions to assign work to three archetypes. What are the three archetypes? The there's the autonomous AI. So, this is AI executing work by itself. There's AI plus human in the loop. So, that's still AI driving, but there's a human that is interacting with it, providing feedback. And then there's the human plus AI assistance. So, that's the human driving. We don't have the human by itself archetype, because in the end there's no real cases when where well, where the human is just performing all the work in knowledge work. I think that's a an archetype that is uh getting extinguished. So, the first question to ask is determinism. So, is the task fully specifiable with clear inputs and outputs? If the answer is yes, then you can think about the cost of making a mistake. So, what are the consequences of an error? What is the risk aspect here? So, if you make a mistake, what are you risking? Does it need to be first time right, or can we get it Can you know, get back on track in a next stage of the process? So, if overall the cost of error is low, so that means that the consequences of an error are low, or you have a chance to recuperate that later on in the process, then you can go for autonomy. You can just delegate to artificial intelligence. If the cost of error is in between, you would say, "Okay, I need to be extra careful here. " But still, you know, there's it's not the worst case where the cost is very high. You can have AI with you know, collaborating with a human. So, that could be the process of drafting the response to the RFP, where the AI is making some drafts and then prompting the human to provide feedback and extra context. Do you like this, or do I need to add something? Do you approve? And so on. So, that would be a an AI plus human in the loop. Or, if the cost of an error is high, then you go for a human plus AI assist. Now, coming back to the determinism, if the task is not fully specifiable, then you ask the next question, which is about judgment. So, is the task at hand uh

uh you know, judgment heavy? Does it rely on tacit knowledge, on knowledge that is not clearly accessible to the machine? If the answer is no, then you can again have a process where AI is driving and the human is in the loop, providing that tacit knowledge. Finally, if the judgment the task is judgment heavy, you can have the next question that you can ask is about the recurrence. So, is this a recurrent process? Or yes or no? If it's yes, again you can go for AI plus human in the loop. And if it's no, you can go for human plus AI assist. So, that's becoming a bit complex. The goal is not to deep dive and spend too much time on this. The goal is to show you that you can have a structured approach to assign to allocate tasks. This is an example. There can be you know, different ways of doing that, but it's important to try to come up with a clear framework for this assignment. Now, if we come back to the RFP, we see that we have the three phases that are clearly assigned to the different archetypes. So, the instant intake is fully autonomous, the parallel generation is AI plus human in the loop, and then you have the review and submit, which is human driving with AI assistance. Now, again a little question. So, when AI makes a mistake, how do you usually find it out? Do you catch it during review? Does it go all the way to the end user, to a client or a stakeholder that flags it, and then you correct or not? Sometimes you don't even find out. Or do you have an automated check built in the process? Okay, yes. So, some people let the clients report mistakes. That's that can be you know, — an efficient strategy perhaps. I see a lot of A's. B again. Yeah, D. And okay, also C. I was wondering. Yeah, okay. There's some spread. Um I think uh the what you want to have is as much as possible D or A, of course, right? So, that's also about the escalation logic, right? And this is actually the centerpiece. This is the jidoka. So, the principle that you know, Toyota defined is to have you know, also for the autonomous tasks, a process of escalating. So, when the machine is doing something, it has to have some built-in checks that allow it to detect an anomaly, and then stop the production line in the context of manufacturing, or you know, stop the process and prompt the human, and then the human intervenes, fixes the immediate problem, and the human is you know, a first-line human, and then with an escalation logic if required, and you solve the root cause. So, the principle is that the machine handles the norm, and the human handles the exception. The trigger is designed, and it's not accidental. An example would be in the context of the RFP would be compliance verification. So, this is a case that's where you we are happy to let AI do the compliance verification fully autonomously, but we still want to flag certain situations that you know, are unclear. And some examples could be there is an unknown requirement. So, it's a requirement that we don't know about, and so let's involve the compliance lead. Or there is a borderline match. So, let's see that's an example that is not very realistic, but let's say we have an insurance of we are insured for up to 4. 8 liability million, and the requirement is 5 million. Then you want to talk to the bid manager. You want to this is a business problem. So, what do we do here, right? Or there are some contradictory clauses in the RFP. So, they actually the RFP is not well written. That happens actually quite often. And then you need again to to go to the human to decide what to do. Now, step four. We need to design also the human role and this needs to feel like an upgrade, not a demotion. This is very important.

And here one lesson that we can take from the factory and again from Toyota that is, you know, a very inspiring organization in terms of designing workflows in a factory context. So, what they did is they decided that automation needs to be less than 8% for certain key processes. That's because they need to keep human engagement. They need to still have humans that know the process inside out because the human is the key to the quality. And if you just automate everything, then you lose your most valuable thing, which is the human expertise. We can draw a pretty much direct parallel to knowledge work. So, imagine that we are automating everything. So, how do we keep the knowledge in our human brains? How do we uh you know, for those that already have certain years of experience, how do you maintain that knowledge? And for those that are entering the world of work. So, how do you learn? Okay. So, that's actually quite important to consider. So, there needs to be a design of the role such that there is enough engagement in the routine tasks in order to maintain the expertise. There is a concentration effort on judgment intensive work where humans add genuine value. So, don't delegate judgment even if one day you could. Uh perhaps, you know, delegating judgment becomes is dangerous because you lose uh you know, the ability to strategize. And then, yeah, you also need to have feedback loops from the human to the system and having the human involved in the process probably makes the system better, right? So, perhaps, you know, it's better to escalate once one time more than one time less. And finally, we get to the final step about metrics and reallocation. So, this is about having a living process. So, you want to reallocate continuously. So, the metrics that matter don't necessarily change. So, if you are talking about RFPs, it's still about win rates. It's about gross margin, the pricing. So, the metrics remain the same, but what changes is the observability. If AI if much it's if the process is much more executed by a machine, then you can observe it. You can measure much more of it. And you can reallocate it. You can adapt it, right? And another important aspect is that with the technology that advances, the equilibrium shifts. So, uh we have today AI models that are um you know, able to do things that 6 months ago they weren't able to do and likewise, you know, in the future the models will improve. So, that means that the boundary of what of AI autonomy the boundary of allocation shifts. So, we need to redesign the process keeping in mind that we still want the human to be involved, right? And if you follow these steps, that's how you can get from the initial process that is a multi-step process that is cumbersome to a process like this one for RFPs uh with three phases and with a throughput that is potentially of 3 hours. So, what you could do, so if you are, you know, inspired by the discussion that we had today, you could uh tomorrow morning you could pick one workflow and think, you know, how does it change how does this workflow change with artificial intelligence? Does it shrink drastically? Can it be handled end to end by artificial intelligence? And is my is the human role elevated or actually not even required in that? These could be tasks, you know, workflows related to data extraction, first pass screenings, reporting and so on. Or does the workflow change form? This is you know, the work still exists, but the how changes. The exception handling becomes AI-aided triage and so on. This is the example that we discussed today, the RFP response case, customer support, compliance reviews. Or perhaps there are some new opportunities. So, you can

workflows that weren't even considered yesterday, you know, become possible because the capabilities of AI, the fact that AI enters as a new agent, cognitive agent change the economics of what is possible, what it's worth doing, right? And this these are tasks where perhaps you are doing some real-time optimization, you are capturing data at you know, at a different frequency and so on and and thinking about new possibilities. Okay, so this brings us to the key takeaways. So, there's three main takeaways. Uh first of all, AI is not about optimizing workflows. It's about redesigning them. It's we have the opportunity to rethink work, to rethink how decisions are made and it's not just about executing tasks faster. Then secondly, the problem is about allocation. It's not automation. So, the performance, the real gains come from assigning its decision to the right agent, human or AI, that maximizes quality against cost. Uh where cost also considers risk, of course. And then finally, the advantage comes from system design and not tools. So, the organizations that win are not the organizations that will adopting will be adopting faster the latest tools, but are the organizations that redesign how AI and humans interact. Okay, so that brings us to the end of this webinar. We still have time for questions. So, happy to take on a few. Let me see in the Q& A. So, I see from Philip that expertise is also built by humans going through the actual flow. Uh that's a spot on. That's actually, you know, one step in the process to think, okay, how it's not just about delegating as much as can be delegated to the AI. It's about keeping a role to the for the human that ensures that uh knowledge is retained within an organization and actually improved. So, which tools do you recommend for building the AI agents? Okay, that's uh that's a question. You know, there are a lot of tools available, a lot of possibilities out there. I would say, you know, uh that you can do a lot with tools such as Claude, Claude Co-work. If it's about, you know, testing new ways of automating tasks and testing new and then it depends on, you know, is this something that should be scaled to many users or just for one person and so on. So, you know, happy to have a conversation about that. So, okay, an interesting question here. How would we redesign the workflow around AI from scratch without knowing how the current one works? So, I mean, I'm reading the question because I'm intrigued by it. I'm wondering in which context you would not know how the current workflow works. So, uh not sure I can answer this. So, in your experience, is there a model that can help us to evaluate if our process is AI-ready? Okay, so I mean, I think, you know, one key aspect of an AI thing, you know, every process will be AI-ready to a certain extent. The question is, you know, is it still relying on a lot of human quality reviews, a lot of that, you know, can be postponed in the process because what changes with AI is that the economics of producing knowledge, of drafting documents changes. So, perhaps you can postpone the review until you've drafted a bit more of it. And instead of having three, four review cycles, you can have less. So, that's that could be you know I would like to thank you very much and

don't hesitate to connect with me on LinkedIn. I'll be more than glad to um you know, to have a chat with with you. So, thank you for Thank you very much for attending the webinar today.