Transcripts
When you wire everything up, the way this AI submission processing looks like is that it automatically extracts it from whatever inbox you have.
Speaker A:It compares it for completeness, it reviews it for accuracy, it suggests routing, then you route it, then person the router.
Speaker A:2 Uses AI to first flag all the potential issues which we're going to see in episode number two.
Speaker A:Then they review, send it back, and then that that loop repeats with resubmittals.
Speaker A:I will show that in episode number three how to track and log everything for resubmittals.
Speaker A:Hello, and welcome to another episode of Design Under Influence.
Speaker A:Boris and I are going to break down the submittal process into three major components and we're going to show you how AI is going to help you navigate this process.
Speaker A:Remove the mundane, stay consistent with your architectural mandate with reducing all the minutiae that comes with this process.
Speaker A:And so the way we're going to do this is in the three episode series, the way we've split up a submittal process is 1 is the submittal intake, then we're going to go to submittal review and then we're going to dig in.
Speaker A:Episode three, we're going to dig in and submittal tracking and resubmittals.
Speaker A:Boris, are you with me to help me navigate this?
Speaker B:I'm with you, buddy.
Speaker B:It's actually, I think you are the expert on this now, so I'm looking forward to learning from you and maybe asking questions along the way.
Speaker A:Let's get one thing out of the way.
Speaker A:We're not architects, but we are an IT company that has been working exclusively with ASC Space Architect specifically for the last, what, 12 years?
Speaker A:15 Years, Boris.
Speaker A:And with that comes the responsibility of understanding the architecture business.
Speaker A:Now, about two years ago we seen AI comment on scene and we started deploying AI here internally at ARC IT to help us serve our clients better, but also deploying secure AI adoption and workflow optimization into our customers shops.
Speaker A:The next iteration of this is really taking something that is difficult to manage, but, you know, in itself submittal process is not that complicated.
Speaker A:But what it does for your firm is it compresses your fee because any project can go between 30 to 1,000 submittals.
Speaker A:And that is quite intense process with the sheer volume of it.
Speaker A:So if it's A one submittal, it's fairly easy to navigate, right?
Speaker A:You get it, you assign it, you review it, you send it back to GC and it gets built.
Speaker A:So that's fairly simple.
Speaker A:Now, the volume of IT times number of projects you have creates this administrative Nightmare.
Speaker A:And this is what we're going to help you solve today.
Speaker A:So the intake work is essentially clerical and you went to school for architecture and you spent last however many years practicing it not to become a human router.
Speaker A:We'll retain the human element where it matters most, where your expertise actually counts, so you can focus on doing what you love, designing buildings and not routing submissions and figuring out deadlines and potentially incurring liabilities for your company for things that were missed.
Speaker A:All right, let me share my screen and show you how the process looks today and how we're going to take it to the next level.
Speaker A:Okay, so here's the process, how it looks today.
Speaker A:Trade partner prepares the drawings, sends it to gc, GC reviews them, stamps them, sends them to Architect.
Speaker A:Architect.
Speaker A:Intakes, downloads, checks, completeness logs, those kinds of things.
Speaker A:Then design team reviews the submittal, architect responds, and then you track and resubmittle, get that cycle going.
Speaker A:So this is how the process looks today.
Speaker A:What we've prepared for you is some sample documentations.
Speaker A:For example, we got a sample submittal.
Speaker A:This is a simplified version of the submittal process.
Speaker A:However, the prompt by prompt guide that we're going to demonstrate and the skill building at the end of this exercise is going to be reusable in your own firm.
Speaker A:So the first document is a submittal transmittal.
Speaker A:This is a demo, but it can resemble something that can actually happen.
Speaker A:This particular one is in relation to door hardware.
Speaker A:We also have our door hardware spec.
Speaker A:And again, this is a sample demo, but we're going to use it in order to build AI process around managing and processing the submittal.
Speaker A:And the last thing we have is the Riverband, which is the community we're building for, we're taking care of the submittal for and we have the full log that we're gonna use to manage this particular submittal.
Speaker A:So far, so good, Boris.
Speaker B:Yeah, absolutely.
Speaker B:I like it, I like it.
Speaker B:Let's dig into it.
Speaker B:I want to see what we're gonna do.
Speaker A:All right, so prompt by prompt, I'm gonna take you through how to automate the busy work in the submittal process, how to where to insert, human decision making, and then we are going to basically package all of it as a skill.
Speaker A:And by the way, this will be available as a downloadable on our website.
Speaker A:You'll be able to download all the prompts as well as the skill itself, which you can deploy in your company and iterate and change continually as your needs differ.
Speaker A:First of all, let me explain where we are right now.
Speaker A:We are in Microsoft Cowork right now.
Speaker A:Now, Microsoft has Copilot, which is a chat, and Cowork, which is the work agent.
Speaker A:And this is where we are.
Speaker A:We're in the cowork right now.
Speaker A:This is a premium subscription product.
Speaker A:Boris, you want to talk, just touch a little bit on Microsoft Copilot.
Speaker B:Yeah, absolutely.
Speaker B:So to take high level, what we're going to show you today can be applicable to either the Microsoft Copilot cowork or other tools, other agentic AI tools that you may have deployed in your workspace.
Speaker B:It could be cloud, cloud, desktop, cloud, cowork, or it could be chatgpt work.
Speaker B:But we're focusing our demo on Copilot Cowork, which is part of the Microsoft 365 copilot business license.
Speaker B:It's a paid license.
Speaker B:It's in addition to your standard Microsoft licensing.
Speaker A:So the important distinction here, this is an enterprise account.
Speaker A:We're not shoving proprietary, you know, GC and our own company's information into public accounts or personal accounts.
Speaker A:This is where our secure AI adoption takes hold.
Speaker A:This is the first step.
Speaker A:So we're talking beyond secure AI adoption.
Speaker A:Let's say you've already developed policies, you've already classified your data, your people already know what they can and cannot share.
Speaker A:You already deployed enterprise grade security account, whether it's Claude chatgpt, Codex, or Cowork with Copilot.
Speaker A:So we're assuming you're already here, and it's important for us to mention this because I would not go into your personal account and start playing with this just because the security concerns and the kind of stuff you will have coming to you through the submittal process.
Speaker A:All right, so let's go with the first prompt here.
Speaker A:Let's start the actual process.
Speaker A:So you see on my screen my version of Copilot cowork, and you see the first prompt.
Speaker A:So what I did here, the documents I just walked through with you a few minutes ago are attached here.
Speaker A:Now, natural question would be, well, do I have to attach all those documents every single time I want to process my submissions?
Speaker A:And the answer is no.
Speaker A:There are steps above this particular prompt that will help you automate the ingestion, if you will, of your submittals.
Speaker A:And this could be done in multiple ways, whether it's procore, other project management system, maybe GC or the owner owns their own system, or it comes to you via email.
Speaker A:All those things can be wired up to have your submittals come to you and the system AI will or Copilot in this Case can identify them automatically and compare them to the specs.
Speaker A:Your specs, on the other hand, would live in a particular space where you're going to point the skills of your copilot AI into where your specs are stored.
Speaker A:So it will know to be able to compare them.
Speaker A:In fact, it can even go and find them in your email.
Speaker A:That's how cool it is.
Speaker A:The log will live somewhere where your copilot, once it has the skill acquired, will will know where it is and be able to access it and update it.
Speaker A:That's the end goal.
Speaker A:But here, for the purpose of this demo, we're building the first set of skills.
Speaker A:We're building the first skill in a whole submittal skill process, right?
Speaker A:And skill one is going to be the intake, verification and intake.
Speaker A:And so what we're asking it to do is using only the attached files, identify each file and extract the intake data, submittal number with source locations.
Speaker A:Do not review technical compliance or change anything.
Speaker A:It's sort of important.
Speaker A:So how we came up with this prompt and how we came up with this prompt, we actually broke down the process on a whiteboard and then we asked another AI to help us create prompt by prompt guide in order to get to our outcome.
Speaker A:And that's where we landed on this initial prompt request and subsequent prompts.
Speaker A:And now we will give you those prompts as a downloadable.
Speaker A:You can have them or you can have them here on the screen.
Speaker A:You can see them and you can adjust them as you see fit for your own organization.
Speaker A:But we've already done the homework compiling the prompt to get to the outcome we wanted to get.
Speaker A:So that's our first prompt.
Speaker A:So without us putting this live into the system and waiting for 5, 10 minutes depending on how complex it is, we already got the answer.
Speaker A:So copilot identified three files, it took this intake data And Boris, what did it do with this intake data exactly?
Speaker A:Can you help us understand where it.
Speaker B:Reviewed the PDF, it deciphered what data was on it, and then it stores its data into what we're going to call, quote, unquote, short term memory.
Speaker B:This is a markdown file that lives inside of this project.
Speaker B:And since we're using the Microsoft Copilot cowork, it lives in the Folder on your OneDrive.
Speaker B:The intake data is now temporarily stored in that markdown file and waiting for the next steps.
Speaker A:Exactly.
Speaker A:So it actually kind of gave us a little bit of heads up here, even though we didn't ask for it to review it, but it sort of found that contractor review Block is blank.
Speaker A:That's good to know.
Speaker A:But it stopped data extraction and completeness confirmation and did not do any technical conformance yet.
Speaker A:It just did what we asked it to do.
Speaker A:It let us know that it, hey, it read all the files that stored in short memory.
Speaker A:We're ready to work on them.
Speaker A:So my next prompt was incomplete.
Speaker A:In fact, I wanted to show you this as an example.
Speaker A:So this particular prompt just said, check intake completeness, distinguish a print and contractor review block from completed approval, blah, blah, blah, very short prompt.
Speaker A:Assuming incorrectly that AI knows what I'm looking for and you know, crap in, crap out.
Speaker A:So my short prompt resulted in, you know, fairly decent review.
Speaker A:But you know, that says package completed is readable, everything is good to go.
Speaker A:All these green marks.
Speaker A:It gave me, you know, kind of couple unresolved flags and said, oh, maybe we should put it as pending intake or maybe, you know, we should put status in review.
Speaker A:So what I did next, I said, okay, that's not good enough.
Speaker A:And so we rewrote the prompt and now the prompt.
Speaker A:I'm not going to redo the whole thing, but the idea is this is a lot more specific.
Speaker A:You can read it on the screen if you want to, but it's a lot more specific as to what we're looking for.
Speaker A:Legibility, GC stamp, conflicting information, dates and things of that sort.
Speaker A:How we our org does it, right?
Speaker A:How we want to review our submission.
Speaker A:Once we've done that, this more complete prompt gave us the information we wanted.
Speaker A:Look at this.
Speaker A:First of all, completeness check.
Speaker A:That's a step one, right?
Speaker A:And right away, contract to review and approval actually complete.
Speaker A:No, it's flagged it.
Speaker A:Okay, because the block is printed but empty.
Speaker A:There's no check mark that GC actually reviewed the submittal.
Speaker A:And so beyond that, it went ahead and did his job, checked everything.
Speaker A:Then it also flagged conflicting dates because apparently we were requested to get it done by 7:18, but contractually we should have 14 days.
Speaker A:So there's something for us.
Speaker A:You know, we have to make a decision here.
Speaker A:Human needs to get involved.
Speaker A:This is the sort of stuff that we don't let AI automate, right?
Speaker A:We need to make sure that this is where it's flagged and that's what the prompt specifically outlined.
Speaker A:Don't make any decisions, Just let us know what's flagged.
Speaker A:And finally, this says this should be the status where it's pending intake because we need to ship it back and put it on back on GC in order to review it.
Speaker A:So we're doing their homework for them, right?
Speaker A:Review it and resubmit it.
Speaker A:So that's what we're going to do.
Speaker A:Okay.
Speaker A:Once we reviewed everything and we agree, we're going to say that next prompt is going to say I approve, go ahead and edit the log and draft an email or draft the correspondence to GC to returning this submission for them to re review and add their stamp.
Speaker A:Okay, so it went through the process.
Speaker A:This in matter of.
Speaker A:In actuality took about what, Boris, like six minutes, five minutes for it to do.
Speaker A:So it went after the task.
Speaker A:It actually wrote, I'll show you this.
Speaker A:It added a row in our submittal log.
Speaker A:It provided notes as well as the information in relevant columns in the correct format, which is great.
Speaker A:Again, if it didn't, this is where we would have gone and corrected it before we packaged this up as a skill.
Speaker A:It tells exactly what changed.
Speaker A:And then it actually drafted an email to Summit Builders.
Speaker A:You can copy and paste this right into the email, but in reality, when you wire everything up, the way this AI submission processing looks like is that it automatically extracts it from whatever inbox you have.
Speaker A:It compares it for completeness, it reviews it for accuracy, it suggests routing, then you route it.
Speaker A:Then the router person, the Router 2 uses AI to first flag all the potential issues which we're going to see in episode number two.
Speaker A:Then they review, send it back, and then that loop repeats with resubmittals.
Speaker A:We'll show that in episode number three how to track and log everything for resubmittals.
Speaker A:But the idea is instead of it creating and sending emails for you, it will create a draft, but in my preference, it will create a draft in my outbox, right?
Speaker A:So I can just press send, review it, add few things and press send.
Speaker A:So what we did next is we asked Copilot to package this workflow as a reusable skill and it went ahead and did that.
Speaker B:So we've just invested probably about 45 minutes to an hour of real time going back and forth trying to identify things that we need this to do.
Speaker B:We don't ever want to go through that process again, especially if we've gotten the output that we like and is going to be useful to us.
Speaker B:So this is where the skill comes in.
Speaker B:Any workflow that can be repeatable should be packaged as a skill so that the next time you can just invoke that skill, you don't have to re prompt it and it will deliver the result.
Speaker B:Maybe it will take some time to actually generate a result, but you can Just start the flow, call the skill, start the flow and go do something else.
Speaker B:Come back to it 15, 20 minutes later and it should be done for you.
Speaker A:All right, well let's go ahead and invoke the skill.
Speaker A:Right.
Speaker A:So we talk a lot about it.
Speaker A:Let's go start a new task.
Speaker A:We will need to still reference the context.
Speaker A:Okay, so here I have the files we did that in original chat, right?
Speaker A:I have them attached right here for reference.
Speaker A:And now what we're going to do, ladies and gentlemen, we're gonna do magic work.
Speaker A:Now let's invoke the skill and see where it's gonna take us now because this is probably gonna take three to four minutes.
Speaker A:So we're just gonna magically impress it and get you right to the outcome.
Speaker B:Before we compress it.
Speaker B:One thing about skills and skill creation, it's better to have multiple skills that accomplish a single task but do it really well versus having one large overwhelming skill where the model can actually get confused and may have conflicting instructions if it's a multi step process.
Speaker B:So having multiple skills that do one thing well and then combining them into what, what are called super skills, long workflows is better than creating one big skill.
Speaker B:Quick best practices there.
Speaker A:And so that's what we're doing.
Speaker A:Right.
Speaker A:Because we broke the process into three stages.
Speaker A:We're going to create a skill for each stage and then we're going to stitch them right.
Speaker B:Correct.
Speaker A:Now in episode two, we're going to cover how to faster cross check without replacement and professional judgment.
Speaker A:Right.
Speaker A:The point of this particular demo is to show you that first intake process that it's good to break a more complex process into components that make sense.
Speaker A:Just know where one starts, the other one ends.
Speaker A:Right.
Speaker A:And then build skills in your AI tooling for each component of the process and then you can invoke them.
Speaker A:And then by the way, when we decide to wire up, let's say we want to wire up this copilot to our email.
Speaker A:The beautiful thing about Copilot is that it has Microsoft grade security number one, but it already has access to all of my Microsoft universe documents, enterprise stuff.
Speaker A:Right.
Speaker A:My work stuff.
Speaker A:And so it's very, very easy for me in this case.
Speaker A:I mean it's easy in Codex as well.
Speaker A:I mean let's Claude as well.
Speaker A:We can access, give it access to inbox but then there's security concerns, right.
Speaker A:Giving you tooling access to extra stuff expands the, what we call Surface of Attack.
Speaker A:In this particular case, it already has all my emails.
Speaker A:So it's Easy for me to give it access to inbox and say, hey, monitor for submittals, anything that looks like xyz.
Speaker A:Again, that prompt can be fairly extensive.
Speaker A:So it knows how to find, knows what folder my specs are, it knows where the logs are, and it's essentially going to prep all the work for you.
Speaker A:So when you come into work, the beautiful state of being is you have four submittals waiting for you.
Speaker A:Mr. Alex, here are some of the first things we found that you should potentially work on.
Speaker A:This is ready for outing.
Speaker A:This is, this is needs to be reset, this one needs to be reviewed.
Speaker A:And then you dig in and it basically presents you with what it found.
Speaker A:And then you use your own brain to work on and finish them up.
Speaker A:And then instead of task switching and spending, you know, 45 minutes to two hours between a whole day tracking those things down, you're gonna have it all tight one place done in 15 minutes.
Speaker A:So here we are.
Speaker A:We invoke the skill renamed this particular task, and I recommend you name tasks that matter to you.
Speaker A:Demo 1.1 submittal prompt by prompt.
Speaker A:And then once we packaged it as a skill, now skill invocation, it took about nine minutes.
Speaker A:Human review gate submittal, Boom, boom, boom.
Speaker A:Ready for approval.
Speaker A:Okay, it's referencing identified the files.
Speaker A:It shows you the data that was extracted.
Speaker A:Now it has the completeness check.
Speaker A:Ooh.
Speaker A:It tells you right away, contractor review actually completed, no flagged.
Speaker A:The approval block is blank.
Speaker A:So it already flags the response date as well, like we saw in prompt by prompt.
Speaker A:And it flags that the thing is not complete and cannot be processed any further routed.
Speaker A:It proposes the log row specifically to how it's going to write into the submittal log.
Speaker A:Here's your routing evidence.
Speaker A:And it's also touching unresolved value, which is due date.
Speaker A:That's what GC is asking for.
Speaker A:But it's not contractual, so it's going to ask you to make that decision.
Speaker A:So recommended next action log this as a pending intake, Correct?
Speaker A:Right.
Speaker A:And then request completed, signed contract, review and approval block.
Speaker A:And this is where our human gate is.
Speaker A:We can just tell it.
Speaker A:Yep, Reply approve to write the row and draft as drafted and tell me which of the four decisions to change first.
Speaker A:So if I approve here, it's going to move to the next step and then it's going to draft me an email to the GC so I can send it and ask for resubmittal.
Speaker A:So what we've learned today is break down your process into different segments, then workflow the process on a whiteboard.
Speaker A:Then let AI help you identify specific prompts to write into an agent in order to ask the agent to do what you want it to do.
Speaker A:Then you package, once you like your output, you package that into skill and then you can just invoke the skill and then you can share that skill across the organization.
Speaker A:As an outcome of this episode, we will share the prompts we use which will help you will get you like 70% there and then we'll share the skill as a downloadable as well on our website getarchitect.com next week we're going to look at episode number two.
Speaker A:Is going to look at the actual review process and how to build the skill to enable AI to help you with the review process and where to put human gates so you are still in full control.
Speaker A:Boris, any party words of wisdom?
Speaker B:No, Alex, I think this was very comprehensive and thank you for doing the demo.
Speaker B:I've learned a lot and hopefully everybody else did as well.
Speaker A:Thank you very much.
Speaker A:We'll see you guys next time.
Speaker A:Thank you for sticking with us later.
