278 | How To Build AI Agents: A Step-by-Step Playbook for Business Leaders with Jim Spignardo

23:52In that case, sometimes I'll just scrape the one there and then just upload it. The other thing, too, is after you've created your agent, you have the ability to share it to your organization. You can either share it to the whole organization, you can share it to a subset of individuals, either as the group or individual user accounts.

24:11Jim Spignardo:So here's an interesting IT question for you related to that, which I truly don't know the answer, but I'm curious. Let's say that this agent is using a specific folder in my drive. And let's say that I share this with the entire organization, but some people do not have access to that folder on their personal logins, will the agent prevent them from seeing that information? Did Microsoft close that loophole? Or if I gave them access to the agent and the agent has access to a folder, you're going to get access to that folder even if you're not allowed to see that folder? Yeah. So if that content exists in N365, maybe that's your OneDrive or Teams or SharePoint, the agent can warn you that the people you're about to share with don't have access.

24:59And you can, if you have the privileges, right, you have to have the ability to assign permissions. You can do that as part of the agent creation. The thing you have to keep in mind is that if you do do that during the agent creation and you then remove the agent, it doesn't remove the permissions. You got it. If that content exists somewhere outside of M365, it can now just give you permissions to my C drive for instance but but it doesn't work the

25:26Jim Spignardo:other way around meaning if I let's say I do not assign permissions to the people who do not have permissions would they not be able to use the agent like what's going to happen then like would they be they wouldn't right got it okay so it does work this way it protects the data structure and access from a company perspective. Yeah, absolutely. Yeah. And, and, and so that's something you need to think about. If you go ahead and share it. And again, that they've just added that recently where the kind of similar to when you share a link and outlook, it will say, Hey, some people in this email may not have access.

26:03It's the same way. I would say some people might not have access. Do you want to forgive them access? That wasn't there before. So people were setting up agents and then people were going in and go, this doesn't work because Because they now access the pods where they were located. So it does still respect your permissions. Awesome. You can't get permissions if you don't have that privilege.

26:21Jim Spignardo:Got it. So I'm out here now in my agent catalog. And you can see all these agents that we have and bill. And what I want to do is pull up the RFP response bot that we actually have for our own organization. And so I do see that one right here, but I know I think I actually have a second one. So let me go ahead and type that in. Yeah, this is the newest one I have. So I'm going to go ahead and find that in my list. What I want to do is go in and edit that because I want to show you what's occurring behind the scenes. And so it immediately jumps to the configure tab. And you'll notice that similar instructions, although it's a little bit different because it has some things in here about tone and style.

27:04And I actually used the second method that I talked about to create this one, which was to collaborate with Copilot. and tell it what I wanted to do to build me the instructions, which I then brought into the gym. And so that's why I'm getting a little bit different outplay. I have things in here about compliance and ethics, parallel processing capability where you can handle multiple RFPs at the same time. We also have some limits and transparency, response optimization, and then continuous learning and feedback. You'll notice if I go down a little bit further here, this is where I have given access to these resources.

27:47So within our organization, I've given in access to our services SharePoint site, which goes in and talks about all the services and solutions that we can sell and position with a client. We also have information in our sales and marketing website, which probably is going to fill in a lot of those sales pitchy parts of the RFP. We also have a document that specifically talks about RFP response sections. This we built over time responding to RFPs, and it's essentially chunks of data within this Word document that typically get filled into an RFP. We also have a PDF about what are winning formats and strategies.

28:28We want to give it the ability to say, hey, when you create that first draft, make sure you're relying on this best practice knowledge about what a winning RFP looks like. We also gave it our location for all of our previous RFPs. And then we also gave it a couple PDFs for best practices for winning a proposal and then mapping your ideal RFP response. I have up here, allowed it to search all websites, which I think actually In hindsight, I'm going to turn off. And you notice that I do have the ability to use specified, only the specified resources. And over here, you can see my starter prompts, essentially.

29:10Jim Spignardo:So what I want to do at this point, and I've said before I can show you, this is where you would share it. And you can see anyone in your organization, specific users, it does give you a little bit of a blurb about permissioning. They'll respond with knowledge that each user has access to. Okay. So it is permission bound to the user unless you change it. So keep that in mind. So let's drop out of this and go directly into the agent. And one of the other things to be aware of, at least in the Microsoft ecosystem, is when I'm in a chat, they could be in a chat in any application or just general chat.

29:48If I want to call up an agent, I don't have to necessarily open up the agent directly. If I type in the ad symbol within the chat window, that is going to give me a list. And of course, I'm in an agent right now. So let me show you in a new chat. If I go in at the ad symbol, it's going to bring a list up of all the agents that are available to me to call up. So I can pivot any time into using an agent in a regular chat. Like I said, that works in any of the Microsoft applications as well. PowerPoint, Outlook, Word, Excel, you name it.

30:20Jim Spignardo:Two cool things about this. First of all, most people don't know that, and that's sad because it's a very helpful thing. So again, you can use this from anywhere, which is very helpful. But the other really interesting thing is it means you can build multiple agents for a single process. So let's think about the RFP for a second. You can build one agent that analyzes the RFP, another agent that analyzes the data in your company that aligns with the requirements of the RFP, and a third one that actually writes the proposal outline, and a fourth that writes the details for the proposal. Like you can do it this way.

30:53Jim Spignardo:The benefit of that, well, two benefits. One is you have more control over the process because the in-between steps, you can stop, go back, finesse, fine tune, et cetera, before you continue. But the second is you can get much more granular instructions for each and every one of the steps, which will mimic more what you're actually doing in real life. But then the reason that connects to what Jim said before is you can have a regular chat and then add the first agent and say, let's start with the research. Here are the RFP documents for this RFP, and then we'll do the research. And then he comes up with an output.

31:26Jim Spignardo:And then you add the second agent that now knows what happened in the conversation because it's the same context window. And this is how you can keep on going. So that add symbol is not just for being lazy instead of trying to find the agent somewhere. it also is very useful because you can use multiple of them in a single conversation. Yeah, absolutely. Good point. Absolutely. And here's the rule of thumb I usually put around agents. If you are in the process of building an agent and there's more than one primary task you want it to complete, it's a good chance you probably want to create another agent.

32:02We could definitely build this as multiple agents, although typically we found that with our users, this is the best process for them. But we have some examples that we're building one for our HR department right now. There's a sub-unit in that department, which is our talent acquisition department. And they said to us, we want to be able to answer questions like they have. And we said, no, we're not going to do that. We're going to basically just link to that agent when it gets to that topic and go, is there a question about talent acquisition? We will hand off to that agent so that their agent can answer.

32:35And so we're not duplicating effort in our organization either. All right. So I have my first, I have chosen the initial prompt here, which is the RFP analysis. I'm going to go ahead and click the plus button and I'm going to upload a synthetic or simulated RFP that I asked ChatGPT to create for me. It's a school district. Let me see if I can bring up. Oh yeah, I do have that. So this is what it looks like. It's a PDF file. It's pretty sparse. I know a lot of RFPs are a lot more detailed than this, but it created this fictitious school, Redwood Valley Unified School District, gave it an address, gave it an RFP number, proposal date, due date, demographics of the school, you know, the purpose, the current network environment, their data center, their objectives, the scope of work.

33:26So, I mean, it has all the elements you'd be looking for. Usually they're much more detailed than this, but for the purpose of this demonstration, this will do. So we now have the PDF for that RFP, and we'll go ahead and click the submit button. And now you'll notice what it's doing here. With Copilot now, because it has a model router, you can actually figure out whether it needs to think deeper or it can answer fairly quickly. So you'll notice, you'll see, it responded fairly quickly because it probably had a decent bit of information. So below is a structured RRP, read it, ready analysis, and draft response.

34:06If you want to reformat it into a full proposal, exec, and summary, or compliance matrix, you just got to let them know, right? So here's the key highlights, the purpose. My screen is frozen there for a moment. What's going on? There we go. We just got stuck for a second. Some of the gap risks and proposals are for the proposal considerations. This part here, it said for the RFP compliant, the content is structured so that it can be pasted directly into a proposal, which is cool. Keep in mind, it can actually put it right into a Word document as well. There's a little bit of an executive summary draft, then goes into ProArc's technical approach aligned to the RFP, talks about what we would do as it relates to that specific part of it, talks about the wireless network assessment, the security architecture review.

35:01It talks about any cloud and SaaS connectivity. Then it also talks about the deliverables. Here's our vendor qualifications. Again, this stuff actually came from stuff related to our teams internally. So, you know, in the PCI-E level network architect, we have multiple CISSP engineers. We have Azure and cloud network specialists on and on and on. And then it actually...

35:26Jim Spignardo:And again, for people, just a second, And for people to understand, the way it knows that is because you give it access to your company data through the relevant SharePoint folders that has this kind of information. So it's not making it up. It's actually pulling from the data that you gave it. And this is where the real magic happens, right? You need, going back to what Jim said in the beginning, you need to know which data it needs access to. It's exactly this. Like you need to plan ahead and say, I'm going to create a folder or a subfolder somewhere that has exactly the information this agent needs, no more nor less.

36:00Jim Spignardo:And I'm going to point it to that. And like you've seen from GM, this could be five different folders. It doesn't have to be one that has specific aspects of information that will help it execute what it needs to execute. And because we told it to only use this data, that's the only source of truth it's going to use, and hence it knows how to do these kind of things, down to the level of what certifications and how many people with those certifications exist in the company to be able to comply with the requirements of the RFP. Yeah, exactly, exactly. And it was even able to go down here when we started working on pricing models to know that we actually do penetration testing and access point lifecycle design.

36:39right so these are things that may not have been mentioned in the rfp but we can state that to the customer hey there's an optional add-on you can also do these things for you as well now goes through a compliance checklist right which basically um is checking the draft to make sure it actually addressed all of the elements of the rfp right so and he wanted to go through and say let me check my work let me ensure that all the things that are laid out there because i'll be honest, we have answered RFPs where there was a section that we thought we addressed, but to the client, they thought, well, you didn't address it to the extent that we thought you would.

37:18So if they're doing it this way, we get much more consistent responses and we get the AI to be able to check our work for us. And then down below, we get kind of these outstanding questions for finalization. So there might be some things that we want to additionally add to this. So the one is, it's asking us if we want to have a complete proposal in a PDF of Word format. Do you want a set of narrative sections to paste into an existing template? And or do you want a compliance matrix and executive summary only? You'll notice then I also get some suggested prompts based on the questions it's asking me.

37:59Do we want the narratives only or do you want to generate the compliance matrix and executive summary? What I'm going to ask it for is, and I didn't see this. Can we build out a task list and associated based timeline approach?

38:24Let's see what it gets me there. this is typically a lot of times what a customer wants to see in a in an rfp so this came out there and said okay assuming this kickoff date and we can adjust those if necessary uh we could complete by july 15th uh the recurrent recommended phased approach is aligned to a 60-day window phase one would be the project initiation it knows you know our project management philosophy because again Again, we've fed it documents and knows how we run our project management office and what that would look like. Down here, we have deliverables out of that phase one, which is the project plan, a documentation request log, and a stakeholder engagement schedule under the network infrastructure piece of this.

39:12Notice this information is coming right out of the RFP. I don't think we've shown you that, but there were some specs in here about the high school had a 10 gig network, the middle school had a 5 gig network, and so on. Again, the deliverables, again, the wireless heat mapping exercise we will be doing, even talking about which environments we're going to look at that typically have density issues. Phase four, we get into the cybersecurity, heart protection, compliance.

39:40Jim Spignardo:Yeah, I think the phases themselves are less relevant, but the clear thing is that it knows how to approach it from a systematic way, right? It knows how to read the RFP, understand the requirements, understand what the company knows how to do, and then put the two together into a very well-structured format. the other thing that I will say that I do a lot when it comes to these kind of things like when you're trying to write a really long document and usually a response on RFP is not going to be five pages it's not going to be 10 it's going to be 50 to 500 sometimes yes and the right way to do this is kind of like to follow the process that Jim defined so you start with an outline and then you go okay let's work on section 1.1.3 and then you start working on that with the AI and then once you get that, okay, let's go to 1.1.4.

40:32Jim Spignardo:And this way, because if you ask AI to write the full document, let's say it's supposed to be 150 page of answer, you're going to get six pages, sometimes 12. The only way to get it to dive to the level of details you need is actually to do the step-by-step process of let's create the outline, let's understand the components we need. Now let's start working on each and every one of the components separately. And as you do more of this, you will get a feel of what the level of granularity you need to impose versus what it will know how to do on its own. Yeah, absolutely. Yeah. And that's the problem.

41:09That extends to a lot of different types of collaboration that I do with AI. And it could set you in my section. You kind of work out the details, work out the kinks. You have it, analyze it over and over again and then say, okay, I'm happy with that section. Maybe copy that out, put it in a Word document and move on to the next section. And, you know, it sounds pretty labor intensive, but I will tell you it's been a lot less labor intensive than it was in the past for us. You know, if I think about what on average it would take and how many people it would take to be involved in an RFP, you know, it's probably no less than three people, usually four to five people over the course of 10 days putting all this information together.

41:57We recently responded to an RFP in less than a day and a half because we were able to pull all this information together so rapidly and get a good first draft. And then, like we always tell people, have a human review it and go through with it and say, OK, we don't like that section. This section reads a little inauthentic for who we are. And you can even say, hey, can you make this seem more on brand? And then we'll do that and we'll change it up for you. But, you know, the amount of time that it takes us to draft now is probably down by 75, 80 percent of what it used to take. And then, again, our win rate has dramatically grown up.

42:41We probably, again, we won about as many as we lost. I would say we're now closer to winning three quarters of our partner RPs. And sometimes it's not because of AI. It's because we're not qualified enough to do the work. or there was somebody already had some sort of unknown advantage over us that we didn't know.

43:00Jim Spignardo:Two things. One, a bit of an interesting add-on to what you just said, and the second is a question. So adding a little bit on, a lot of the work in the RFP is really evaluating whether the company actually has the expertise, the skills, the knowledge, the background, the past experience to actually win this. That by itself used to take days of work. Sure. And then you may get to the conclusion like, okay, hey, we're actually, we can't win this. Let's not bid. You already spent the two, three, four days, two weeks, depending how big the project is to get to that conclusion. Where now this part of it, if you built your data set correctly first in five minutes, you will know whether you should invest more time in this or not without investing any serious resources.

43:45Jim Spignardo:And the other thing is usually these resources are the most capable resources in the company. So it's not only you need to invest these two days, these are the two days who are the most senior project managers with the most experience with like, these are the people that can evaluate whether you can do this or not. These are the people you don't want to get out of their day job because they're the one that generally need the most amount of revenue to the company. The question that I have in the ChatGPT universe, when I do this, I use Canvas to write the answers. The benefit of Canvas is that it's a fully editable document that I can add text myself.

44:19Jim Spignardo:I can highlight text and ask AI to work with this and so on, which makes the iterative process of creating an outline and then working on 1.1, 1.1, 2.2, and so on, very easy to do. I don't believe that's possible in the co-pilot universe, but I'm asking about saying. Not to the same extent, but what you do have the ability to do is to, I'm looking here, you can actually send this to pages, right? And I'm not seeing that option right here, but if you send it to a page, the page does open up a canvas, which you can then edit things in and more easily copy from the chat window into that. The feature you're talking about with ChatTBT, yeah, it's create per coding too.

45:07You can see the changes being made while you're asking for the iterations.

45:10Jim Spignardo:So your best practices suggestion is working two windows or two monitors, one that has the chat, one that has a pages version of the first draft, and then you go back and forth. And the nice thing about the page, when you send something output to a page, is that after you're done editing it, there is a button in the top that says, send this to Word. So you can take all of that ongoing kind of workshop benching stuff and say, now I'm done, put it in Word, and we'll send it off for approval or last draft or last review by humans. Great stuff. Quick recap and summary of everything we did, and then let you say some final words.

45:52Jim Spignardo:First is you need to know which data you have, where you have it, what's applicable for what purpose. And again, now generalize it, forget about RFP for a second. That's true for any aspect of the business, right? So having a solid understanding of what information you have in your company, what of it is good versus not so good versus bad, labeling it correctly and putting it in places that will make sense to humans and agents was always true, but it becomes way more important right now. Because if you don't have that, everything that Jim just showed you is not possible because you won't have the 12, five folders and six documents that it gave the RFP agent to be able to do the things that it knows how to do.

46:39Jim Spignardo:So number one, you need to have your right data intact. Number two, as we said in the beginning, you need to figure out the use cases. And as we said, the right way to figure out use cases is things that can help you grow the business while hopefully saving you time. One or the other would still probably be a good option, but in this particular example, it does both. It reduces the effort and it generates more revenue. So it's a perfect fit to be something you want to build agents for. The next thing is really how you build the agents. And as Jim and I mentioned, the best way is to just have a conversation about it.

47:13Jim Spignardo:Just talk to the AI in a regular chat to say, this is what I'm trying to do. Here's the information I have. This is the goal that I'm trying to reach and just work iteratively and then in the end said, okay, now write me the instructions. And then the last thing is really, it's never really done, right? This gets you 90 % there, but they're like, oh, we should have added this. Go and add this. Like every time you learn something you didn't do before, you can go and add that additional thing to the agent. In many cases, I do this in the chat itself. So I'm like, oh, I would like the agent to also do this.

47:46Jim Spignardo:Can you add me a set of instructions to the current instructions that will add this new functionality capability, check whatever you want to do. And it will write you another paragraph or two in bullet points and will tell you exactly where to put it in. And you copy and paste it in the instructions to just keep making it better and better. And then like Jim said, you can share this with relevant people in the company so you're not the only one using it. The only really annoying thing that Microsoft did, but I understand they're trying to make money, is the other people need M365 licenses as well in order to use the agents that were created, which to me is really annoying.

48:20Jim Spignardo:Like it would have made sense. There is a little bit of a footnote to that, by the way. Okay, I'm curious. So if you do not want to license your users all up and still want to take advantage of agents, Microsoft does have a pay-as-you-go model where you now can buy token packs. And so if your agent is associated and sent out for the pay-as-you-go model, it's great for frontline workers field workers yeah they can use the agent every time they use it you get charged the penny per each message you can buy also now buy packs that are a discount so they can assign those as well and so now you're not in that model where i don't know if they're going to use 30 worth of copilot every month right yeah and now you have 300 employees i go So$30 is not a big deal.

49:10Jim Spignardo:$30 times 30 ,000, that starts being interesting. So that's awesome to know. I did not know that. Jim, this was absolutely fantastic. I'm sure people will learn a lot, both from the mindset as well and from this specific example. I think it's a very powerful example that, as I mentioned, even if you're not doing RFP work, you're probably writing proposals. So don't call it an RFP, call it customer requirements. And you go through the same exact exercise and you can do this. I can tell you that I haven't written a proposal on my own for at least two years. And I keep on upgrading and updating the process.

49:47Jim Spignardo:So the process that I have right now is way better than I had even a quarter ago. But it now writes better proposals than I ever did by a big spread, by a big spread. And it does it in about five minutes when it would have taken me two to three days. and so there's really zero reason not to do these kind of things that's all i'm saying it's there was the old joke you know in the in the in the project world of faster better cheaper pick two you don't have to pick anymore you can do faster better and cheaper true i didn't think of that yeah that's true absolutely i've heard that expression but yeah ai kind of jake puts turns on its head so yeah if people want to follow you learn from you work with your company what are the best ways to do that?

50:33Sure. I am very active in LinkedIn. I write about three articles a week, topics in AI and technology. So if you just search for Jim Spignardo, I noticed that I typed my

50:44Jim Spignardo:name in with my last name right on there. It's S-P-I-G-N-A-R-D-O. I'm the only Jim Spignardo on LinkedIn, probably the only one in the world. So you can find me there. My company is ProArc. so it looks like it's pronounced proarch but it's proarch p-r-o-a-r-c-h and we have all of our services listed there we just did a webinar yesterday similar exercise where we demonstrated building that hr agent so there's lots of free content on there that you can avail yourself of i'm also coming out i'm looking around my desk here coming out with a book soon an ebook that will be free to download. It's called the AI Turning Point, which really discusses how organizations have to fix what's wrong in their businesses to be able to take the best advantage of AI, both from a business process perspective, but also from a business mindset as well.

51:44Jim Spignardo:Awesome. Thank you so much. This was really well thought after and really well explained. So I really appreciate you taking the time and sharing with us. And thanks everybody who joined us live. I appreciate you spending the time with us and learning with us while we're doing this. Have a great rest of your day, everyone.