Hi everyone, thank you for coming. This webinar today is brought to you by nPlan. I'm going to introduce you to Ben Koops, who's a project controls consultant at nPlan, and he's going to be showing us what 780,000 real project schedules can teach us about building the next one. So without any further ado, over to you Ben. Thanks Jareth, really glad to be here. So yeah, I'm going to pull up my slide deck and we'll get started on this. To introduce myself, it's Ben. I lead Schedule Studio nPlan. What that means is I spend most of my time speaking with planning and scheduling teams across contractors, owners, consultants, essentially getting under the skin of how they actually work and which parts of the job quietly eat the hours that should be going into judgement. So quick word on nPlan in case anybody's unfamiliar. We are an AI construction company and we've been doing this since 2017, so long before AI was cool. We've built the world's largest dataset of real schedules and how they actually played out and we use it to help teams see risk to coming and get a credible program built in a fraction of the time. Schedule Studio is that second part. It's the part that builds them. One thing I'm going to say, a mix of scheduling and planning and programming throughout, it's all the same craft. We work with a global audience, so planners, schedulers, this applies to both of you as I interchange between that terminology. So quick plan of what I'm going to cover. I'm going to start with why planning feels the way it does today. Run through some of the options available. Lay out what I think good looks like and then show you Schedule Studio on a real project and then there'll be plenty of time at the end for any questions and please do pepper them in as we go as well. So yeah, before I jump into the product, I just want to see how programming actually gets built today because when you lay out, it's actually quite a thing. So a project lands, tender, feasibility study, client asking whether a date's even possible. And the first part alone is reading the documents. It can be a week or two and you're not building anything, you're just reading. That's hundreds of pages of scope, specs, a drawing pack, super thick and nothing but a door open and the best planners can carry a whole job in their head, which is kind of a superpower in its own, but a superpower kind of has a cache limit as well. Page 300 doesn't always remember page 40 and when contradictions are hiding between those two documents that came three days apart, it's a very easy thing to miss and that's not sloppiness, it's just what hundreds of pages does to anyone. Then you get to the second stage and you're building WBS and this isn't just detail at the bottom, you're structuring the entire job. So top to bottom, the big breakdown, sections under it, the detail under that, deciding the whole thing hangs together before you've built a single activity. You need to get the structure right and everything else stops in. To get it wrong near the top, you're probably not going to feel it for a while until suddenly nothing fits where it should. Then you've got durations, that's a lot. This is the really rigorous bit, every activity, signed off quantities, output rates, your norms, historic productivity and then the new sequencer. You link the whole thing together by hand, what weight on what, dozens, hundreds, thousands of links and I've got to say a perfect logically linked schedule really is a thing of beauty. The spaghetti balls sometimes happen, slightly less so and I'm a spaghetti fan. Someone told me recently that they had to go for a lie down after a full day of linking and honestly I can really empathise with that. Hours of hooking up dependencies, one at a time, hunting down for the loops and the open ends and you can only really see what it adds to, your dates, critical path, whether a loop snuck in when you hit run. Then once you've done this, you've got a review in terrain which sounds nice and gentle, you can relax at this point but it's not. Something's always moving, a method changes, a date shifts and then you're rippling back through the whole web by hand, running it again and re-sequencing and stitching and again. This is weeks of work and it's been done this way for decades. The quiet cost of all that time, it takes so long to build one credible program that you never get the chance to build a second. There's no time to model different sequences and see which is faster, cheaper, less exposed. You've got to commit to the first approach because not because it's the best one but because it's the one that you had time to make. If we step back for a second, you're doing all of this from a blank page but we're sitting on more scheduled data than at any point in the industry's history. If you think about any other data rich industry, finance, logistics, they all live and die by their data. Our industry is just as data rich but we're planning on kind of memory and best guess and maybe some outdated production rates. Why do we build every project completely new as if the last few hundred thousand haven't happened? That's essentially the kind of the shift that I want to talk through. What changes when you don't start from nothing, when you can start from the world's hindsight and you've got a credible program early enough to influence the decision whilst all the options are still open. I'm not going to pretend that we're the only way to do this, to kind of get to a program faster. If that's what you're after, there's a handful of options. You can build it the way we always have. Do it the same way. Build it by hand in P6 raster and there's a lot of things right with this. It's rigorous, it's yours, your name goes on it and you understand every single line because you built every single line. The trade-off is kind of the one that I just spoke about. It's weeks of work, you kind of get one go at it and it ties up your most experienced people on mechanical grind. Their judgement is the scarce thing and it's being spent linking. You've also got option two. You can just throw more people at it. You can hire more planners, you can pull people from different teams, you can bring in consultants and that works in a pinch. I'm sure we all know that the good planners are scarce and the budget to keep hiring usually isn't there. That option doesn't scale the moment you want to chase more bids. There's a huge lag to get there. That's more of a fix rather than a cure. The whole point of today is trying to get you out of the fix and better still not to be in it to begin with. So option three, you could build your own AI. Lots of people are having a go at this. I've spoken to some teams that have built some genuinely clever things. Every single one of them say to me upfront the same thing. It's capped by their own data. Some of them have a few hundred projects and it's a lot of effort to build these models and it's a ton more to keep them running and it can be chaotic for some. So option four, you could point it at ChatGPT, Claude, a swarm of agents and it will hand you something that looks like a schedule. Maybe even it will be able to be read by P6 but it's kind of guessing what sounds right from language. It's not grounded in anything that actually happened and getting it to behave like a real linked program that you could take into P6 or Astor or Powerproject is a very different thing. And I want to lay out all four because they're all a real choice. And if you're tuned in today, you're probably weighing up one of those choices right now. So that brings us to the actual question of if we were going to design the ideal thing, what would it have to do? What would it look like? For me, it's four things. Firstly, it's got to be grounded in how real projects actually finished. Not how they were planned, how they actually turned out and with data, it's all about scale. And I want to stay on scale for a second because it kind of is the whole game. One project outcome is a story. It ran late, maybe for reasons that will never touch you again. A few hundred thousand of them is a pattern. Now you can see what drives a day and tell it apart from one job's bad luck. So scale is the thing that matters because it means it's seen work like yours before, your methods, your sequences, your ground conditions, even on a job that you've never built. That's the difference between an anecdote and something you can actually lean on. So the second thing is it has to be fast enough to change how you work. A credible draft in hours, not weeks. And I say draft because I'm being realistic. We're not trying to build perfect programs. We're trying to build the first draft here. And that speed is kind of the whole point. It lands early enough that you can still change your mind and you can build more than one. You can run one, two, three scenarios. You can see which sequence wins instead of committing to the only one that you've had time for. Three has to fit into how you already work. So your tools, P6, Asta, MSP, your specific coding, your workflow, it just hit around the way you work and it shouldn't rip anything out. And four, it's got to leave the judgement with you. You have the expert call. This is the sequence sensible. It is the critical part of credible. That's today's yours. In Plan, we obviously as an AI company, we are big AI adopters. We all use it very much in our day-to-day work. But when your name is put to anything, when you submit any type of work, the onus is on you to evaluate that and interrogate that and review it properly before you put it forward. So the judgement always has to be with the people working on it because they understand the full picture, not just the kind of limited inputs. So that's my kind of scorecard. There's four there. And whilst I say it's mine, it's actually kind of what I keep hearing for planning and scheduling teams, which is kind of why I'd love to hear if there's anything you guys would add. If there's something missing or something you would change, please put it in the chat. Hopefully we can get around to it later as we go. But with these four, kind of as our starting point, let me show you how we've tried to meet them. So, NPLAN, where does this credible program you can get in an hour actually come from? So for us, it starts with our nerds. And that's not me being cheeky. That's the official designation of one of our teams. That's NPLAN's Experimental Research Department. They go deep on the hard stuff. So large language models, graph neural networks, forecasting science, stochastic machine learning, uncertainty modeling, like seriously clever people. So we took that lot and we pointed them at our 780,000 destruction schedules and crucially, how they actually played out the outcomes, not how they were planned, how they finished. So this is by far the largest dataset of its kind. This is billions of activities that we're learning from. And that's the bit nobody can shortcut. You can get some engineers, but you can't conjure the learnings of three quarters of a million finished projects. And here's the thing that people don't expect. Schedules are actually almost perfectly built for machine learning. If you open up anyone, B-Sigs asked MSP and you look under the hood, it's the same kind of rigid structure, scope, sequence, durations, links, calendars, and machine learning lives on consistent structures like that. So the thing planners have done for decades turns out to be exactly what we needed to learn from that scale. Each activity carries around 160 contextual features, what came before it, what comes after the link type, where it sits in WBS and more. And maybe you'll say, isn't this just ChatGPT? And I get that a worrying amount. And I say worrying is the right word because not in the way that you might think, is worrying how much stuff out there is just kind of like a chatbot in a trench coat. And that's not us. Here's the main difference. It's two things working together. So we've got our brain, which is our proprietary statistical model trained on 780,000 schedules and how they actually turned out. And then the other half is the translator. And that's the language model. It reads your documents and lets you edit in plain English. They have to work together to get anything kind of substantial. And as I said, we're not building perfect schedules. We're not trying to. It's about getting you 80% of the way there fast. Okay. So when you get there, you can apply your expertise. And that last stretch, the judgement is exactly where schedulers want to spend their time. Anyway, this just gives them more of it. So I've said enough there on me theorizing. I'm going to jump out of this and I'm going to jump into our product and show you what we do. One second there. Cool. Right. So let's paint a familiar picture. It is, let me make this a bit bigger. So I'm not squinting at you all. It is, I think I'll be on the first day. Something lands and needs a program. Maybe a bit's come in. Maybe you need to prove a date is feasible. Maybe a client just moved a key date. And it's not just your problem. There is a whole cube behind it. The bid lead, you can't price without it. The project director who needs to know the date is real. Procurement, which interplace orders. And you're the one that it all files through. So normally that's two or three weeks before you've got anything. And you're already running the maths. It's late one tonight. Friday's packed. So your weekend plans are looking shaky. So I'm going to show you what that next hour could look like instead. And a quick word on what I'm about to show you. In terms of the environment, the demo I'm showing you, we work across all sorts. Rail, highways, water, buildings, everything. So I had a bit of a choice what to demo. But there's so much happening in data centres right now. And I've been having a ton of conversations with data centre teams. So I thought why not take some public documents and build a program for one of the biggest, most complex data centres coming up anywhere. That's metas hyperion. And to be clear, this is all from public data. This is just to kind of demonstrate what the tool does. It's not anybody's program. So this is the first stage. And this is where you upload whatever you had a junior planner. It's as little or as much as you've got. Design drawings, scope document, planning docs. Or you can even go just a prompt describing the job. For this case, it is a bunch of public material that I've put in here for Hyperion. Now you'll notice also, just to say, I've got all the that is because a few of the steps in this do real work and they're not instant. So this is just for the demo's reference. In terms of that, the first is here between uploading documents. And that's the fact extraction gate. So before it builds anything, it reads every document you gave it in 100, sometimes thousands of pages, just like a planner would. So it pulls out all the facts that define the schedule. Geography, the key milestones, the contractual dates, the methods. It sorts them all. It strips the duplicates and it flags where two documents contradict each other. So remember I was talking earlier about the planner who's on page 300. You can't always remember page 40. This doesn't have that problem. That's the factual basis it builds from. And that's why it's not instant like a chatbot. And that's why I have tabs open. Because it's doing real work in between and it's not completely instant. So once I've uploaded my documents here and I progress through to the building the WBS stage, which is the first of a few kind of stage gates where we are center checking what has been uploaded. So the WBS will be generated from your documents. The schedule overview, the milestones. There's a few stage gates as I click through and show you them where this has been generated. And before it generates, before we build the whole program out, it's just sense checking, like, okay, you're happy with this, you know, do you need to move things around? Do you need to edit things? If we go through to the milestone section, the overview section, has it understood your information correctly? Cool. Yeah. Does it understand the program? Yeah. I think started all the right information before we get started. Cool. Then you want to go through any milestones. Yeah, these all the right milestones. Cool. So it's a gated process. And once you've, once you've gone through that, and you've uploaded all the information necessary, you've confirmed it all. You just click a button and there you have it. A full sequence, logically linked program in a full Gantt view. So this is, as I said, it's not client ready. It's a credible first draft. But in that draft, on a job this size, in the time it's taken us pretty much to join this webinar and needs to waffle on. So if you remember the scorecard I was I was talking about earlier, the first thing I said was grounded in real projects actually finished. So if we click into a random activity here, we can see the information. Okay, so we've got durations, we've got, you know, the labor, the resources, calendars, the ways task driven, an explanation of how we've got to where we are. You know, this isn't a guess that, you know, this kind of sounds right based on language. This is a read from what has actually happened. And I guess we're talking about speed. This first pass, the first pass I built was well, it will go to 500 activities. I press expand once in the schedule tools here, I press expand, expand schedule. And I got this version, which is 1600 activities. So as of this month, so October, 2026, the cap for paid users is 8000. That's 8000 activities. And if I wanted to get, you know, in that kind of region, I just expand a few more times, it's going to break these things down, break the activities down into more into more detail. And when I talk about speed, that 8000 number, that's less than a month ago, 6000. And honestly, I bet that's outdated before long, because it keeps advancing. And that's one of one of my many favorite things about working with nerds. Every few months, things jump again. Every few weeks, things are changing. There's some really exciting stuff coming up next week that I will wait to share with you if you get on our mailing list, which actually got delayed for this process. So we wait to do it afterwards. But essentially, I'm talking about the pace, right down to the small stuff. Okay, so a few weeks ago, we got told that activity names were too long and wordy to verbose and planners wanted a tighter kind of verb noun format. About a week later, done, you know. So I'm saying this because if you've got any feedback, we'd love to hear it. Please share it. And you can influence what we're building, because we want to be talking to the planning and scheduling community. So second thing on the core saw card was to build fast enough that you can build more than one. This is kind of where that lands. You can come here and you can iterate. So you can break out different versions. You can go back to any one of these stages and it will build out a second version within the project. And you can build another version. If you want to edit and adjust that new version, you can jump into the editor here. If you could talk to your schedule, what would you ask it? What would you ask P6 to do if you could? That essentially is the editor. You know, all the things that you lose hours to by hand, whether that's re-sequencing, paralysing activities, adding a curing lag after every pull actually came up on a call today. Dropping in a win to shut down. That's the old way. That's unpicking and retying dozens of links. Hunting the loops, rerunning the CPM, best part of the day. Here you type the request, here's some examples that I just pulled up for reference. And you're reviewing the change in minutes. Which means that you can actually build that second, that third scenario, and then you can see which one wins. And remember my third point saying, you know, you have to be able to check it. So however you're reviewing your schedules today, if you live in DCMA 14 points checks, it's all here under schedule integrity. So the activities, the link assessments, the constraints, the calendars, the various bottlenecks, it's all here. And you can click into whatever you want to drill down to. So, you know, have a look at the long durations. Might take a second to load. Review them all here. So this is your kind of mechanical quality review. The qualitative call, that stays with you. That's your judgement. And now you've got time for it. When you want to interrogate the critical path, you can just jump into driving paths. Have a look here. We can increase the view on the float list. Let's look at all the different activities involved. We can jump in to driving those paths for you. And we can review the various different critical paths that could be involved. And the last one that we said was important. It has to fit with your tools. So XER files, straight in and out of P6. You can upload from these tools. You can from P6, Asta, Powerproject, XER Powerproject with the XML. You can download a PDF for the teams on the ground and share with them. It works natively, kind of, within your existing workflows. And look, I've intentionally not gone through absolutely everything. You can actualise the program. You can show critical activities. But I want to show you the shape of it rather than just run through a million features. If there's something specific you're wondering, it does. You can ask in the Q&A. Better still jump in and check out yourself. If at that point you're still not sure, get in touch with me. Happy to answer any questions. So that's me rattling through. I'm going to pull back to my slide deck now. Okay. So remember that question I started with? It was what changes when you don't start from a blank page? When you've got a credible first draft program early enough to shape the decision? Whilst every option is still open, what changes? That is what the hour buys you. You can pressure test a sequence before anybody's committed numbers to it. You can build a second, third version you never had time for and pick which one actually we're in, instead of committing to the only one that could get built. You can chase more bids with the same team. You can spend your hours on judgement calls that win work, not mechanical rebuilds. You can bring people into the conversation sooner. SMEs, procurement, subbies, with the working draft instead of a blank page. You can just start the conversation sooner. You can have something to work off in all your various different working groups. But the real point I'm trying to make here is, yeah, it's time, but it's actually a lot bigger than time. It's not really about saving time. It's a different way of working. When the program shows up early enough to change your decisions, because it's grounded in what actually happened on hundreds of thousands of real projects, you can stand behind it. Everything else kind of changes. Yes, it hands your team the capacity back, but capacity is not the headline. The headline is decisions that you can make that you couldn't make before. The best thing you can do is to try it. There's a free trial. You can jump straight in, no credit card, test it out on a real project of yours is always best, because I can show you a schedule that I've made, but you're going to be like, yes, that looks like a schedule. It's better if you know the project. You can see, does it stack up? Does it make sense? Just see what you can build in an hour. See what you could do. If you want to go further than that, grab me. Happy to talk more. I think that's everything. I would love to open up to questions now. I'm aware I've rattled through a little bit faster. It's funny how you speak a lot faster when nobody's replying. Bring your questions. What kind of things would you like to know? Thanks, Ben. Really appreciate that. We've had a few questions coming in the chat, so I'm going to have a go reading them out. asked where the 780,000 schedules came from. Sure. As I said, we've been doing this since 2017, and the bulk of our business is actually on the risk side of things. We build it up with our main side of the business, and that comes from our customers. 50% of those are owners, so we've got some real depth and specificity there. The other 50% is from contractors, so we've got some real breadth. That is a truly global and trans vertical dataset. Okay, thank you. When you were demonstrating the properties of an activity, you explained that you could get into the detail on what the duration is and how many hours have been worked on that. Could you talk us through a little bit more detail about the resources that you could assign to each activity or what gets generated to that activity between what the user provides in context or edits and the bit in the middle that I guess generated by AI? How does that work? Sure. With regard to resources, as of today, that only comes from the context that you have with it. The documents that you upload, if that has context around resources, we'll extract that, understand that, and then we'll apply that. If you had a resource sheet, you could upload that. We can extract that, add that in. You can also additionally edit and add resources yourself on an individual level. You can adjust activities to be task or resource driven, as well. That's where it stands today. How does the tool generate resources? Has it got the data from the 780,000 schedules to say there's these many resources needed in prelims or whatever in M&E? How does it work? It doesn't generate the quantities for resources. It will assign the resources based on the you assigned to each task. That is left for you. Another question we got was, can we add schedule assumptions and that schedule narrative? I might go further. Perhaps the basis of schedule that could be produced, perhaps? What kind of narrative do you mean? As an artefact, as an output from Schedule Studio, can you generate a basis of schedule that captures assumptions and tells you what the basis is of the schedule? What you can do, actually, and there's something I didn't share here, there is a fantastic section where you can pair schedules and you can get a full narrative of the schedule in there. For a standalone schedule, it's not current. Let You can get narratives of your programs, but only currently in the comparing. If you build multiple iterations, you can compare the two and then you can have narratives of each one, how they compare and contrast. But for an individual one, no, you're just going to get an XER XML file that you would plug in and play and move forward or PDF. The durations that get generated by the Engine in Schedules Studio, are they based on deterministic planned durations from your database or a mix of planned and actuals? Yeah, good question. That's usually one of the most often. So they're grounded in how our dataset, the real schedules actually performed and it's not just kind of copying a number across. It's infused with the context of the scoping documents that you're sharing. So it draws from the dataset and then it goes to the documents you shared, extracts all the relevant information and then says, okay, when we see context like this, when we see stuff like this, what do durations tend to look like? That's the way it comes up with. I also didn't mention a key thing on this, which a lot of people really enjoy. You can share previous programs that you've built, so a reference schedule. If you want to infuse your particular way of working, your durations, logic, codes, all of this stuff, you can share a reference schedule you can ground not just the structure that we're working on, a maybe previous similar program that you've built or schedule that you've built, but also you can have that influence your durations as well. Yeah, from the 780,000, what it established the duration based on the estimated productivity from that schedule or we have to do new productivity? No, yes, productivity rates is not a factor for us. The productivity, so the 780,000 and how they actually performed productivity rates is you would have to build, you know, work with your own ones of those. It will give you a duration and it'll say how it's reached the duration, but it won't be built from the bottom up in the same kind of way that maybe planning teams will build with productivity rates. It builds them based on other things it's seen and how things tend to happen within that type of context based on statistical analysis, but it's not grounded in productivity rates like a planner would. Okay, sorry, one more. In first draft, what is the score which we got on the duration? So again, is it required again, we are to call subject matter experts and to see what is the realistic duration for this one. So like the workshops required or how realistic the first draft which you showed in this webinar. So are you asking how accurate the durations are? Yeah, yeah, sure. Sure. I think the best thing is to check it out. The reality is about 95% of them people are very, very happy with. There'll be a couple that are a bit off. That's the whole point, you know, it's that first pass, you review it. Sometimes there'll be maybe like a long lead item that's been a bit compressed. It's still human in the loop. They're not going to be perfect, but you need to review them. The majority, bang on, you know, we've built programs which have been almost identical to the ones that tier one contractors have built. There have been still the activity that you need to adjust as well. At the start of your presentation, I like the opening where you've explained the use case as a schedule of doing a schedule and the time it saves. That's what it's all about, right? I was wondering if there are particular use cases that you've explored on the types of projects or even by sector that you've noticed that are completely different to one another, say a nuclear energy project versus a water one or a contractor side one to a client side one and so on. Whether you've helped clients navigate that with particular data from that 780,000 days, that makes sense. Yeah, I mean that's a big question. There's a lot to say on that. I mean, as the conversations I've had, they have been incredibly broad. Everybody has different ways of working. They're working on completely different projects. The actual definition of a project is, it's entailed within the definition that it's unique, you know, it's bespoke. But one of the really interesting things and exciting things, especially for machine learning, when you break down a project into activity level, and as I was discussing, you've got all these contextual hundred, you know, 60 contextual features. When you break it down that way, you actually notice that a lot of these projects that seem vastly different because of their vastly different context actually draw tons of parallels from each other. You know, they all have to lay concrete. They all have excavation, maybe some demolition. There's all these transferable activities. So whilst one job might look at first glance completely different to another one, there's all these learnings because you had to do similar, similar actual activities when you break it down on that level. So whether it's a port, it's a data centre, whether it's rail, there's all these transferable aspects that makes it really valuable and perfect for machine learning because it's very hard for humans to kind of join these dots together, but a machine can do it very well. There's one use case that in mind. I've recently seen an owner needing to put together a summary schedule of a contractor's PFA level five. Can you put lower level schedules into the system and get it to generate? Not necessarily 100% using the seven to eight thousand dataset, but contextually aligned to say milestones from the PFA or elements of it, but use the logic. I don't know. The machine learning model that you've got from the historical database. In other words, you use a bit of the data that you've got. You use a bit more of the context, the schedule that you're putting in there to generate a schedule using the engine in the schedule studio. There's more context. Sure. I mean, if I'm understanding you correctly, that's pretty much what we do with reference schedules. So if you start with something, you can start with that and you can. Well, there's two ways. You can either upload a program schedule and then just start editing it. So start working with it, or you can upload a reference schedule, which means that will influence the outcome, but rather than being static and tied to it. So you upload a reference schedule and then additional scope, additional information, and then you can build something from that. Another question I had was on the structure, which I think was kind of asked earlier, but just explore it a little bit further. Say a client wanted to establish the WBS dictionary and they didn't want a one generator from those loads of other project dates that you've got, but they'll want to use your engine to generate something quickly. So in the context, they could feed in, say, the WBS dictionary would be able to narrow the stochastic nature of a generator schedule into needs to be built against these groupings and guidelines and whatnot. Yeah. I mean, again, if you've got a similar project, you can upload a reference schedule and you can use that WBS. You can upload a straight WBS as well. If you have a structure that you want to work from, you can use that as the starting point. So at these various stages, and then they're kind of gating, you can choose how it's built so that when you get to the end, you don't have to make as many revisions. So choosing your own WBS or a schedule to influence your destination are two ways to start that. And like with all kind of AI tools that you might build from, it's all about getting a good foundation starting right and building as you go. We're coming up to the end of time now, Well, I mean, if anybody does have any extra questions as well, please feel free to get in touch. Thank you everybody for coming as well, and I hope to see you all soon at the next event. Thank you. Thank you. Cheers. Bye.