80% of the Workload Done Before a Human Touches It

Sean Weisbrot: What's it like to spend all day every day thinking about how to disrupt your own business model so that others don't do it for you?

Ray Meiring: It's a huge amount of mental energy, but incredibly exciting to be thinking about that all day, every day, to be reading about it. Sometimes the ideas are coming to you at, like, 03:00 in the morning, and you're scrambling for some device to write them on. But extremely exciting, very energizing to do that.

Sean Weisbrot: What are you thinking about in three in the morning? Give me something specific that came to you recently.

Ray Meiring: I'm thinking about how do we change the proposal process, and can we do everything from cloud? Do we even need a UI on this thing? Is it completely headless? Or, where's the value of the UI? Should we be challenging old thinking and old beliefs around this? And you know those kind of three AM thought processes that happen? Sometimes they're absolutely daft, crazy, and and they should just be put straight into the trash. But other times, they're pretty good, and you come up with a good idea.

Sean Weisbrot: How do you know which ones you should throw in the trash and which ones you should keep thinking about?

Ray Meiring: When I wake up the next morning, there's some clarity that normally comes to some of those ideas, but you've got to test them. So it'll be straight into a meeting with maybe my product team or my engineering team having a discussion about, you know, what we might do, or even better, with a customer. Hey, I had this thought. What do you think about it? Has this got any merit to it? How's your organization thinking about these things?

Sean Weisbrot: Have you ever considered talking to Claude about something before you talk to that customer or your team?

Ray Meiring: I do it often, actually. I I I might bounce the idea of of Claude, get some critical thinking on it. It's been really good at doing that, at giving me a perspective on things. Sometimes it kinda tells you what you wanna hear, which is, which is the risk with that. So you still need further validation, but absolutely.

Sean Weisbrot: Do you ever go to another AI with an idea that that Claude gave you, like Gemini, for example, and say, hey. Claude gave me this idea. What do you think about it?

Ray Meiring: I haven't done that yet. I haven't done that yet. So yeah. But maybe that's something to to go for.

Sean Weisbrot: How has AI helped you or hurt you in the process of coming up with the next generation of RFPs?

Ray Meiring: The challenge is obviously trying to track where this is all going. Right? And what is possible and what isn't. So on the herd side, the herd comes in that you're constantly thinking, and you're trying to keep up, and you're trying to you're trying to move in that direction. But where it's been incredibly helpful is taking the current UI that we have in our system, all the current flows and process, and saying, what would this look like in a fully agentic world? And then having that conversation of challenging yourself or it about what that looks like. So it's hurtful when you've got to be constantly deliberating those points, but very, very helpful in being able to challenge and run various processes through it.

Sean Weisbrot: One of the things that worries me is that I can have an idea and implement it faster than I can think about whether that was a good idea or not. I imagine with a business like yours, where mine is new, I don't really have customers yet. So it's okay for me to iterate, but it's also bad because I don't have other people giving me feedback. You have the benefit of that.

Ray Meiring: Right.

Sean Weisbrot: In an industry that hasn't really changed in, like, what, fifty years, people are still using, like, binders to, like, share information. Right? They they probably mail each other binders.

Ray Meiring: Right. Some some dude. Yeah.

Sean Weisbrot: How do you get to a point where you can iterate at the speed that your customers can tolerate where you're not driving yourself crazy going, goddamn it. I wish this would move a million times faster.

Ray Meiring: Probably one of the most challenging things to do right now, and I'll I'll tell you the reason. Many of our customers, they're still using AI to help them write a better email. Right? That's the full extent of their understanding around what artificial intelligence can do. So going to them and presenting this agentic idea where it does 80 to 90% of the of the workload on your behalf, it blows their minds. And I say this with love and respect to them because their job is not to think about agentic, it's to think about building RFP responses. Having said that, when you're in a new customer interaction, no one wants to see the old stuff. They all wanna be involved with what's what's coming down the track. They wanna see the art of the possible, and they want that reflected in your software. So what we started doing is our first release is into sales. It's not into existing customers. It's actually into sales. And go put this out there and get some feedback and some perspective on how is this resonating with the prospect. Right? Can they see themselves in that? And then we bring that back and iterate on that through. And then we've got these kind of leading edge customers that are very much working, you know, step by step with us, where we can put the good stuff into their hands and make sure that it works. And I'll I'll give you a third piece. But

Sean Weisbrot: Sure. No. Go for it. Go for it.

Ray Meiring: yeah. Yeah. The third piece is where in the past, we've built kind of horizontal tools and then applied those into various industries in their unique ways. What we've seen right now is that taking a horizontal approach to the problems we solve, like RFPs and proposals, that's really challenging. We need to go vertical. We need to be really specific in the problems that we solve in those use cases. And so in the last eighteen months, we've hired industry professionals into the business that we can bounce and can work directly with our product teams to to make sure that these agentic ideas that we come up with, they actually are grounded in the reality of how people work.

Sean Weisbrot: I wanna take a step back first before we go further into agentic. Why should businesses care about this process?

Ray Meiring: Because when they apply this technology, call it a Genetec AI, whatever we want, It is gonna be a huge time saver that will unlock the potential for them to focus on the more strategic human based elements of their jobs, which in our world of RFPs and proposals is really understanding the customer's needs, being empathetic towards what's important to that customer, and developing stories that are going to motivate the humans that are making the buying decisions.

Sean Weisbrot: Won't the agents just be creating those stories for each other and then can helping them to understand what the humans want so that you don't really need any humans involved in the process at all?

Ray Meiring: Well, I think there's part of it that's going to be fully automated. You know? Asking an RFP question like, do you have data redundancy, or do you support pro bono work, or some basic question like that. Sure. That could just be agent to agent. But if you imagine that you have a complex legal problem that you're looking to solve, right, and you're a big corporation, are you gonna be happy with an agent pitching you the fact that they can solve that problem for you? Probably not. You wanna sit across the table. Right? You wanna sit across the table from another person that tells you, I got this. And the agent can make the documents and can route the processes in the background. But you're going to want to sit across the table

Sean Weisbrot: They're not gonna know it's an agent.

Ray Meiring: as that big corporation with a big problem from another human who you trust, who will tell you, I'm going to use all of my technology and people to solve your problem.

Sean Weisbrot: But why why do the humans need to sit across the table from each other when the agents can do all of that thinking and all of that work and all of that planning without any human intervention if done right?

Ray Meiring: Oh, it can do all the work. It can pull pull the document together, but it will not have the same connection with the human decision maker as another human. Right? Unless it's a highly transactional, commoditized type of work. But if there's a some kind of lawsuit that you're dealing with and you need assistance with that, you're gonna wanna talk to another human who has empathy with your problem, who's done this before, who you can fully trust, Not just make a selection from a a smorgasbord of agentic outputs. It's that it's that trust element.

Sean Weisbrot: How is all of this important for smaller companies that have a dream of serving enterprise but have no idea what the hell they're doing yet?

Ray Meiring: Smaller companies need to think through the strengths and weaknesses of those two entities, right, and the two entities being the agents and the humans. Because enterprise is still going to have a big human element to it, particularly in this in this more sales type world of of what's happening there. So what is the human strength that we really want to drive to the forefront to free them up from the stuff that actually is the agent strength so that the agent can take care of those pieces? But if the thought process is just agent, it would be like the thought process being just human. Neither of those two are a complete package without each other.

Sean Weisbrot: And I feel like the smaller companies are gonna be the ones that get into agentic work much faster than the enterprise. So how do they work backwards with people that aren't there yet? Because if you say, oh, my agent will take care of it, and they're like, what's an agent?

Ray Meiring: I think I think the enterprise customers know what an agent is. They're they're familiar with the concepts and the and the principles around around agents. And I actually think it's a better time than ever for smaller companies to be selling into the enterprise with agents, because if those agents can tackle one discrete problem and do it incredibly well, that frees up the humans to do other things. There's a real opportunity to do that. In fact, we've seen that in our industry where we're we're kind of midsized. We've got these agentic start ups coming in, and they're not focusing on everything. They're just focused on this one specific problem that the agents can solve incredibly well, and the enterprises are engaging with them and buying into into that, that approach.

Sean Weisbrot: Why should enterprises not focus on automation instead of agentic work?

Ray Meiring: Isn't it the same thing to a large degree?

Sean Weisbrot: No. Because automations don't have to be smart. They can be dumb.

Ray Meiring: Well, I I see that agents automate processes. They just automate it with a lot more intelligence. Right? So like the old way of automating workflow type tools, it was very specified, declarative in the way that those steps were laid out. And if anything deviated from that, it had to have exception processing. Now with agents, you don't need to be as definitive. The agent can operate within its boundaries to come up with the best outcome there. But the process is still being automated to a large extent.

Sean Weisbrot: I guess the reason why I ask that is because I've implemented dozens of automations into my podcast operations.

Ray Meiring: Right.

Sean Weisbrot: And I I did it as automations, not as agents even though everyone on the Internet's trying to sell me agents. I did it because I don't trust agents, at least not yet.

Ray Meiring: That's an interesting perspective. It's we see it as a blend. Right? We see that agents will form part of that automation process. Right? And it may be an agent handing off to another agent to do work, but that's the process that's going to run. I'll give you a real example. If you're doing a capability statement for a law firm, there's typically five steps in that capability statement production process that would happen. Right? Now each of those steps can be performed by one or many agents. And overlaying that whole process is another agent who's driving those next steps and moving the gates along in that. In the past, we would have built workflows and human based, yes, no, I'm done, I'm not done. Now we don't have to do that anymore. It's just an agent to agent kind of approach there. Do so does the kind of the primary agent that overarches everything have some very definitive instructions on what how it needs to execute on the steps? Absolutely, yes. It does do that. Does it have quality gates built into that? Yes. Quality gates to check that. But it's orchestrating and handing off to other agents to do the small amount of work or the the units of work in that process.

Sean Weisbrot: I guess my fear is that they will hallucinate information if they don't have access and don't have the freedom to figure it out. So for example, like, if I have an automated workflow and it breaks, it just breaks. It doesn't like, there you know, you have console logs. You have network logs. You have things that you can look into to figure out, is this thing actually working or what happened, why not? You have scripts you can build so that you can force it to happen again, you know, in the case that it it missed its scheduled cron job, whatever. I don't I don't wanna get too crazy for people that don't understand.

Ray Meiring: Right.

Sean Weisbrot: Essentially, you have fallback systems, and you know that the automation is not gonna delete your data or hallucinate a response. So if you have an agent that's tasked with doing a job, if its goal for existing is to complete that job, but it doesn't have what it needs to complete that job, it's not gonna error out. It's not gonna time out. It's gonna probably hallucinate what it needs to get done in order to finish, and then you have to trust the orchestrator to realize that that is not acceptable and stop them and have them do it again, provide additional information. Like, this is where my fear is. And when you just have an automation, if if the automation works correctly, if the files are designed correctly, if, you know, your systems if your back end supports it, whatever however you do, I use serverless. You know, it just works. It works over and over and over and over and over and over and over and over because you you know that that thing is not gonna change. I think the only benefit of an agent is that you need some level of fluidity if your process changes, and the agent can handle that. But my fear is, yeah, hallucination or access to data that you overwrite or, you know, like, OpenClaw scares the hell out of me as an example.

Ray Meiring: Yeah. That's pretty cool. Could go down a rabbit hole on that one. But I I guess it depends on the type of process that you're running. Right? Like, in our world, again, building a capability statement for a lawyer. Step one in that process is do some client research. So primary agents asking a research agent to go and look back at our CRM, look back at any RFPs we've sent to this person, look back at our last engagement, look for any news articles online, and produce a client research document. In that process, primary orchestrator agent then looks at that document and says, Hey, did we get did you get this right? Is there a problem with the document? Now, there's human in the loop in certain of these cases as well. A research document, you don't really need human in the loop too much because it's just an internal document. But when you turn that research and run it all the way through that process to the point where it becomes that capability pitch document that's beautiful with bios and experience in it, you need a human still to go and check that. Right? And you might even want that human in the loop a little bit earlier. But the orchestration is happening to move that process along and kind of loop back if something's wrong, fix it. Because we're dealing with a lot of research, a lot of unstructured data that needs to be presented here. And so those agents are reading and writing and refining a lot of that written and read content.

Sean Weisbrot: So how do you build your agents? Like, are you using n a n, using MEG, Zapier? Like, because, there's a lot of these different platforms out there. And I feel like they all do the same, but I could be wrong.

Ray Meiring: Our team grew up on the Microsoft c sharp background, and so they're in that world, that Microsoft world, and they're building these agents on top of those frameworks. So the agent SDKs that Microsoft produces and allows. A lot of it is being built from the ground up using those baselines that Microsoft provides. A key part of what we do is we integrate into the Microsoft Office platforms and the Microsoft M365 platforms, because that's where, let's take a law firm, they want to operate. So we have a very Microsoft centric engineering approach to building these agents and deploying them as well. What's been interesting with that is you know, we see Microsoft's Copilot out there. We also see a huge use of Claude, legal specific tools like Legora and Harvey, some of these big names that are out there as well. So while the actual core capabilities are being built with the Microsoft frameworks and the SDKs in the background there, obviously, MCP is the thing that's, you know, putting that out there and making this accessible to everything else.

Sean Weisbrot: It's interesting. I interviewed a guy years ago who was building apps for Microsoft Teams,

Ray Meiring: Yeah.

Sean Weisbrot: and he was telling me about how freaking difficult it was to get anything published on Teams because they had this manual process of of review for every application that could take, like, weeks or months. Do you have the same issue? Okay.

Ray Meiring: Still there. Still there. Still the same process. They need to get an agent sitting behind that. You know, the good news with that is it's a pain to get it there. Once it's there, it's pretty you don't have to keep updating it through their through their offering. You update the source code. It goes through a review once. If you change the manifests that they have, then you gotta go through the review again. But most of the time, it's a one and done.

Sean Weisbrot: That sounds like, Google developers for, like, like, Google developers and I iTunes developers' programs for, like, their apps. Yeah. But I've heard that those are a lot easier.

Ray Meiring: Similar to that. Yeah. Could be. I I don't have another reference, but I know it can take anywhere from a week to a month to get through the through the store. I mean, last year, we did a we did one through the Adobe marketplace because we work with not just law firms, but engineering companies. And they use Adobe InDesign extensively, and so we built an app, put it through the App Store there. That one was a little bit more complicated, but same principle.

Sean Weisbrot: So you're building agents and then deploying them, and then the users are using them how they want.

Ray Meiring: To some extent. To some extent. So we'll have these primary agents that deal with specific use cases, capability statement, RFP answering agents. Those are kind of primary agents. And the users will use that, and the agent will guide them through the bounds of what needs to be done and the best practice steps that need to take place there. And they can access these agents through Office, through Adobe, all over. But it's the primary agent that's driving what the other sub agents are doing.

Sean Weisbrot: How long until agents are talking to agents?

Ray Meiring: Eighteen months. Eighteen months.

Sean Weisbrot: Why do you feel that way?

Ray Meiring: Because we've started on this journey already. You know, we're working with, the RFP issuer type organization who sends the RFP out to have their agent talk to our agent to answer as many of the transactional questions as possible in advance. Maybe just a singular human in the loop review on some of those those questions. Shortcut that process entirely. It's it's already it's already been done. It's already been worked on.

Sean Weisbrot: So what was a typical cycle for the start to finish of one of these processes where it was human to human, where like, to now?

Ray Meiring: So back in the day, and this is still today, but back in the day, a system like SAP Ariba would produce an Excel spreadsheet. An Excel spreadsheet would be sent to another to the the selling human to complete that and then sent back. And depending on the complexity of that RFP that's received by the seller, it could take anything from two weeks to six months for the seller to complete that RFP. Again, depending on on complexity. Typically, three to four weeks, of cycle time there. In an energetic world, that process happens straight from the buying platform to the selling platform. The selling platform does a bid no bid. Do we really want to do this? Is this in our wheelhouse to do it? Have we won this before? Qualifies in or out. It qualifies in, goes ahead and answers majority of the baseline questions. Now it can, moves past the first gate. Now we're into the high value discussions around the more kind of strategic elements of that of that RFP.

Sean Weisbrot: And how long does that stage take? Because it seems like the the base layer should be fairly simple if both agents already have the vast majority of the knowledge they need to be able to generate and respond.

Ray Meiring: Well, the the first step of, like, the basic bid, no bid, and then answer the questions, that's gonna be days, hours. The real long pole in the tent there will be the human just validating that and signing off on it before it goes back in. Because that step's still gonna be important for a little while until everyone trusts that the agents are gonna get it right. But that could be within a day that you've passed that first gate. The second gate, where it's more of the strategic stuff, that may still take some time, but at least you know you're working on the good stuff as the seller and not just, you know, you've got this RFP now you need to respond to.

Sean Weisbrot: Right. Because you could potentially send out a dozen of them and only get one done and through through and and the sale is made, and that that could take a year.

Ray Meiring: Right. Yep. That's exactly right. Yeah. And, you know, depending on the, the industry, like, the legal industry, you you're still gonna have an RFP. It's gonna have the questions, the basic questions, and it's gonna have the more complex questions. Then you're gonna get shortlisted, and then they're gonna want somebody to pitch, some human to pitch. So there's still that full process that needs to follow. We're just shortening the time of the first gate on that so that it's agent to agent. So we get to the really good stuff that we want to look up.

Sean Weisbrot: Should it be illegal for companies to be that large? That they to be so large that they need, they need months or years to figure out if they wanna buy something?

Ray Meiring: Well, my example of a month to year is something like, hey, we're RFPing for you to build a bridge, across this huge body of water, or we need to rebuild the, you know, Hangar five of big airports. So the reason it takes a year is because there's just a lot that needs to go into the planning of that that that RFP response. But you can imagine how expensive it is to bid on a bridge construction. Right? It's expensive to bid because you almost gotta plan this thing quite far into it. You gotta be showing pricing. You gotta be showing, like, a draft plan on this. This is what we would do. If you can cut out the ones that you're not gonna win early on in the process, you're actually saving money by focusing on bridge building that we're gonna win versus just going for every bridge that we get our people.

Sean Weisbrot: And it do you also I'm just trying to think, like, if there's a way for the agent to also work on the the actual quoting. Like, the actual okay. Let me because, like, I know that you can show something to an AI and say, hey. Make this better, and it'll it'll come up with a a better jet design, you know, that humans can like, so I imagine if any part of that process includes an AI that specifically focuses on figuring out how much it's actually gonna cost.

Ray Meiring: Right. There are totally gonna be agents that can do the costing, the planning, the resource kind of layout on that. That's not gonna be us, but we'll talk to a specialized agent that's a pricing agent. We'll go ask that agent, hey, run the pricing calculation on this thing. Here's five previous ones of similar size that we've done, come up with the best pricing based on that pattern that's gonna work for us right now. Same for design, same for resourcing. Yeah. It's gonna be a it's gonna be an org chart of agents passing work off to.

Sean Weisbrot: I think that part's really interesting because I, you know, for the software that I I've been building, I asked it to look at every API call we make and every software that we touch that it knows of and turn it into a markdown file with all of the costs and all of the plan all of the upgrade opportunities when we should consider upgrading into the different plans, what are the the the, you know, API limits and calls and all of this stuff for all of those things so that we know how much it'll cost. Wait. It's just me. So I know how much it'll cost to run the business as it gets to a certain size.

Ray Meiring: Right.

Sean Weisbrot: It's so cool that it

Ray Meiring: And how good was this?

Sean Weisbrot: I mean, the cost to run the business right now is $50 a month.

Ray Meiring: K. Great.

Sean Weisbrot: But, yeah, it it knows, like, it it won't be that expensive with the if the pricing model I have remains, it won't because, for example, like, an image generation is, like, a fraction of a penny, and I'm charging a dollar to generate an image.

Ray Meiring: Alright. Alright.

Sean Weisbrot: And the the you know, there's a transcription API. There's the post generation API. There's the image API. There's, like, a number of those kinds of things. But then there's also, you know, the the database provider. And so there as you're aware, there's a lot of things that, you know, are involved in this, but, luckily, there's a lot of businesses out there that offer free plans for people that are just starting out. And so I can essentially start

Ray Meiring: Right.

Sean Weisbrot: I I started and, you know, launched a production ready software by myself in less than a month, and it cost me $50 a month to manage it. But it cost me probably a thousand dollars to actually build the software and the website and all of this stuff because I was stupidly using thinking mode, which triples the output cost. It it they charge you based on output tokens, not input tokens, which I didn't know. So, yeah.

Ray Meiring: Okay. But I think you hit on a key point there, Sean, because we're still coming to terms of what actually running these agents is gonna cost us. We've got a lot of theories. We've got a lot of models. We've got a lot of views there. But as usage increases on this, we're going to learn a lot about how much it costs to operate this. And then, of course, there's the whole pricing discussion and scaling discussion around that, margin discussion around that. But I think that's a key area that's still coming to light as we deploy and use AgenTic AI more and more.

Sean Weisbrot: I think that's why people look at using AI tools as like, they're they wanna charge for usage rather than a subscription because they they always have profit baked into each token.

Ray Meiring: Exactly. Yeah. Yeah. That that's that's great when you're building a more technical product. But for us to go to, like, a a head of sales and try and sell on a token based costing model, it's such an unfamiliar concept for them to get their minds around. So there's a there's a transformation that's gonna need to to take place here, especially around the cost side, cost and set selling price.

Sean Weisbrot: What's the most important thing you've learned so far in your career?

Ray Meiring: Speak to the users that understand and know the real world use cases that you're aiming at. I think it's great to be able to come up with a lot of ideas, but speak to those users and understand. And I'd say something else. I would say some things are worth innovating around and making changes and and and trying to find the next thing. And other things, you just need to actually stay the course on certain ideas for a long enough time period to actually see it through. And let me give you an example of that. And this is especially the case with with AI today. You come up with a great idea. Right? You put it in front of your users. They say, this is a this is an excellent idea. And you start rolling this out and trying to build scale around it. And immediately, you get nervous because it's not working. It's not going in the right direction that you thought it would. And your inclination is innovate and change your way out of that. Sometimes, that's not the right approach. Your right approach is just to stay the course on the original idea or slightly adjusted idea and go full board of that. There is actually a lot of damage that can be done doing this versus just slight tweaks to a really great idea.