AI, Actually – Episode 13: Building Software 10x Faster with AI: A Real-World Walkthrough

Mike (01:02)

Okay.

Alon (01:15)

I think I saw the last bill – we’re clocking a thousand bucks a week on on Claude, just on Claude, not including the Nano Banana and all the other ones I’m actively using. So ⁓ as as Pete was saying, we embarked on this mission. We’re in several industries and in one market.

Pete Reilly (01:24)

Mm-hmm.

Alon (01:36)

where we got really interested in the way customers interact with the service provider. From a service provider viewpoint, there’s a CRM, right, a system to manage those interactions. And from the customer perspective, there’s a customer portal. Here we’ll sort of deep dive into like, what is the service provider’s view of their customers and how they interact.

And if you think about CRMs, their core job is to track, right, is to track what happened with the customer.

Track all the activities. so one is, you know, help me close deals if I’m in that mode of a sales representative. Help me understand the customer behavior over time from a marketing perspective and from a management perspective. How’s my business performing? Am I going to hit my targets? You know, are our customers churning and all that kind of stuff. So so the underlying data is what’s critical, right? Knowing who your customers are, is what they’re doing, how they’re interacting with you.

is the source that you crave to make all those kinds of judgment calls.

So historically, would say that we’re talking to a customer and they’re like, OK, we’re ready to go implement something. It’s going to be a multi-month, call it $100,000 plus kind of effort to minimum to get started. And we were just curious if is there a better way, given all this hype around AI, is there a better way to think about it? And we said, OK, give us a couple of weeks to try to work through this in our own minds. Like if you’re in a you’ve got specific needs, how

of those needs get addressed today with classic CRMs versus what you could build bespoke. And so we started down a path of first building requirements. What does a CRM system for a specific industry do? In this case, we’re dealing with basically providers to the low voltage audio video entertainment kind of solutions, They serve both commercial and residential. What does their CRM system look like?

And at the end of the day, we said, okay, here’s a dozen things you have to do, everything from integrating to other systems, integrating to telephony, integrating into email and scheduling software.

Obviously all the database record keeping some industry specific modules for doing proposals and other kinds of things. You know, so you’ve got a fairly large ecosystem you’re going to touch. And so we looked at it, it was kind of daunting and said, Hey, this is, know, traditionally you would say, fine, you know, fire up Salesforce or Zoho depending on which end of the market you’re in and start customizing. Right. And that’s where you get to the a hundred thousand plus kind of like just configuration effort.

And so we went down the road saying, let’s use our core know-how and underlying platform that we built over time that helps us write new code and see how far we can get. And what we found is as fast as we were able to build good designs, which you would need either for implementing in any system, whether you’re configuring a system right from scratch.

As fast as we could write those designs, we can get the code and code that’s production caliber code pretty quickly to get written. So the overall experience was such that we said, well, we can, you know, listen very carefully to a customer, use domain knowledge that we, you know, obviously have gained over years, and then to specify what we want and specify in a way that leans on a technical architecture that’s

you know, that’s tried and true and scale up a system at the speed of which it would historically take you just to configure, you know, ⁓ a traditional CRP CRM or ERP kinds of systems. That was that was sort of the big learning, I would say.

Pete Reilly (05:13)

Right.

Right. So maybe if you wouldn’t mind, I think pulling it up and sort of just seeing a little bit of an action will be.

Alon (05:23)

And we’ve

talked enough, let’s just show some stuff.

Pete Reilly (05:26)

Right, exactly. Because,

and while you’re doing that, I was just going to sort of chime in. It feels to me too, like job descriptions probably morph. I mean, I could see a world where a business person who’s running some line of business actually becomes almost like the product manager because the capabilities to sort of define things and design things and provide…

quality specs and so on into a development team are much, greater. So I think it’s going to be really interesting organizationally. think those lines do start to blur.

Alon (05:57)

Yeah, for sure.

Pete Reilly (05:59)

So

maybe just a quick tour and then maybe talk about some of the just the concepts of how you sort of thought about this maybe differently than a traditional CRM

Alon (06:08)

So now when we’re building CRM functionality, we’re not saying how do we just build a system of record? We’re saying, well, how does a sales rep just, you know, work their to-do list and which part of the to-do list can be automated so they can just flag those as, hey, these are all the tasks I need done. Can you help me get them done? Right. The idea is that you want to start all the way at the beginning of saying, what is it you’re trying to deliver and be ambitious enough about

what you can deliver. And then there’s all the make, how fast you could deliver it and with the right quality and all that, right? There’s the implementation, but it starts with you having the right vision for the products that I don’t think is the same vision that we had two years ago.

So quick, of get familiar with the system. Think of it as like most CRMs have this funnel, right? Where you’re starting with some kind of an opportunity, the early stages, the marketing leads, right? So you’ve got a bunch of leads that come in. You’re then going to work the leads through the sales funnel, right? You got to qualify them. After you qualify leads that turn into opportunities, you work the opportunities through the…

stages of the pipeline, hoping to get to a positive result right in closing.

Along the way, you are tracking the accounts that are related to those opportunities and those accounts are then related to contacts within those accounts. That’s the key entities that you’re working with in any CRM. And then around the periphery of that is a bunch of processes. How do you do what you do as a company, which leads to a lot of configuration in those CRMs. Either you’re buying something and then you want to, depending on your standard operating procedures, you want to customize

it to do the work the way you expect to do the work. And then there’s a series of, know, so what’s very rich in many

⁓ systems like ERPs and CRM is this configuration area. how flexible is it to meet my goals? And no different than others, we wanted to make sure that we could address market worth of needs as opposed to one specific customer. And so even though this is a very young project, in a sense, we’ve built a lot of that stuff that only accumulate after years of being in the field.

I’ll pause there, Pete. to see where you want me to go

Pete Reilly (08:24)

You took a perspective that was much more, how do I enable the salesperson as opposed to how do I just sort of capture the record? And maybe we could talk about some places where you see that.

Alon (08:32)

Yeah.

Yeah,

so an interesting spot is if you think about even just the start and the funnel of the lead generation. So let me sort of show that here. the typical kind of like, OK, enter a form that contains all the lead information. Now, this may be populated automatically if you’re wired to, let’s say you have a web form, a contact form on your web or other.

channels by which you take leads in, right? This may get populated automatically. One of the things that we added upfront is the idea to say, okay, look, if you have any sort of records, right, whether they are audio files, right, a call recording,

you know, an image, a screenshot of something, we should be able to take that in and automatically create a lead for you from that. So the process that you saw here is we essentially piped in an audio file. So this might be the call recording. We transcribed it with, you know, using a model for transcription. And then we took the transcription and fed it all these purple fields that you see are automatically extracted on your behalf. And so you’re in process, you

you’re sort of viewing and approving this, right? ⁓ In order to create a lead, but all of that is sort of automated, right? And so there’s not this time of like, okay, cut and paste and carefully edit stuff. And the sophistication of that, we sort of see it as like, this is equivalent to, well, whoever was responsible with typing stuff in, Like entering those lead forms, that’s the job of the agent. The lead agent here is to capture all that.

What’s nice is that because we’re doing it this way, not only do you get the record, but you get the history of how did this come about. We know for a fact that this was brought in.

AI extracted all these bits of information from this node, and that’s how we got populated. And that’s part of the history, right? So now, whenever I work this deal, eventually I’m going to decide that it’s qualifying, I’ll convert it into a lead. Again, I have the same kind of capability alongside any form where it’s heavy duty, I’ve got to type a bunch of stuff in to actually just having the AI, speaking to the AI or ⁓ giving the AI context and say, hey, do this on my behalf.

Pete Reilly (10:48)

Yeah.

Alon (10:48)

that’s kind of part of the journey that we think is great. by the way, and we keep the audio file in this case, is tagged along and saved with this lead record. So AI always has the ability to review what’s happened to date. If I was to kind of go further here, Pete, for instance, as you work into the opportunity records, similar but different here, there’s always an AI summary of what’s going on with the deal.

Pete Reilly (11:13)

Mm-hmm.

Alon (11:14)

it can reflect based on activities. the typical thing that you see is,

I can certainly like, this is a nice format where I can see a lot of information about the deal along the side, but as activities pile on, right? And the deal has moved through multiple stages or a long period of time to get a nice, concise, you know, summary of what the deal is about, right? That is a thing that software couldn’t do before, right? It could give you the numbers. It could give you, you know, 10 activities, these three, but it couldn’t write you a, you know, a summary of what’s going on. And now we just take it for granted, right? That’s just sort of part of what happens. So

as you imagine using the system and I’ve got an abstract summary for every opportunity, I can now have a conversation say, well, tell me about the three deals that ⁓ are stuck in the pipeline for last 30 days. And it’s got a narrative that it can borrow and talk to.

Pete Reilly (12:05)

And one of the things that I’ve noticed you do here, which I think is sort of plays into that is, you know, in a traditional CRM, have, you know, account records and lead records and opportunity records and whatever you happen to type in. But to make the agent really helpful to the salesperson, they need more more context, right? And it’s not just the records I typed in. It’s the sales, it’s the call that I made last week. It’s the email I got yesterday. It’s you know, the diagram of their house or something.

that’s all really important to capture. And it seems to me that you’ve done a lot here to try to expand the context that we can sort of grab by building that right into this capability. Maybe talk about.

Alon (12:43)

Yeah. Yeah.

I think as you think about designing new systems, there’s a few, a few bits of, guess, that we learned. One, as you said, is context is king, right? The AI is only as good as the context that’s provided. And context comes in the form of, you know, what actually happened in the real world.

But another kind of context is how does my business actually run? So if you look at this concept of what we’re calling sales playbooks, in this case, it’s playbooks generically. In this scenario, it’s adapted for sales.

You know, there’s a methodology in every company of how, you what is the sales process? And to the extent you could describe it, describe it by saying, okay, well, in this situation, we’re describing it through the stages that a sale goes through. But within each stage, what are the tasks? What are the things I have to do as a salesperson to move from one stage to the next? And how can AI help me? Right? So if I look at it, sort of the, let’s look at an email step I’ve got here, I’ve got a step that says, Hey, acknowledge the email for the lead. Here are some best practices. Here are some insights.

instructions to the AI of how they should think about this and email templates to use when it’s sending out this acknowledgement email, right? And so just by setting this up and I set this up across for the various stages, now suddenly the AI has the capability to follow along, right? It’s not just naively saying, know, sales happen this generic way. No, sales now happen in the context of my standard operating procedures.

And what’s nice is even though this looks like a fairly, you know, kind of lengthy setup, we actually used AI to generate the setup, right? So we have a, you know, generate your own playbook, right? But here I could upload my document that I already have that describes my playbook and we’ll generate these, these playbook steps on your behalf that you could, you know, tweak it if you need to. And so now that we have this, you can go back to, if you look at opportunities and we dive into an opportunity that we’ve got here, you’ll see these suggestions of what to do. And so for example,

Here is, okay, let’s notify qualification outcome, right? So here’s an email that we should be drafting.

this is a test system, so I don’t know exactly what we got configured. But at the end of the day, it’s able to auto draft an appropriate message using the context of what’s happened so far, using the email template and the business practices that you said are important to you. And then importantly, this email is not a standalone event, right? An email is just part of a long flow that you’re having in the sales process. And so it’s important to understand that, hey, when we send this email, for instance,

want to be reminded, let’s say if I’ve got no acknowledgement in the email, no response in some period of time, so I want that escalated. But importantly, after the email is sent, the system tracks the email and decides what to do next. if there’s a response, how do we parse that response into a next set of actions according to the playbook? If there’s no response, what do we do? And so there’s a key system here, and credit to Mike for building this quickly and professionally.

for us, this looks like another daunting thing, right? If you think of playbooks as sort of the human version of, what’s supposed to, how does a process work? Workflows are sort of the AI version or the machine version of the set of rules of what to do. And so this gets very detailed in like, hey, send the first email. After you send that email, we want to decide.

If a response went 24 hours, what we do, if there’s no response, what do we do, right? And so this whole workflow, if you will, is all about escalating a series of email and outreaches to decide whether this is a qualified lead or something that needs to go into kind of a nurture bin because they’re non-responsive. And so, and again, this looks sophisticated, but you can build this using AI as opposed to having to wire these boxes one at a time.

Pete Reilly (16:22)

Right?

Right?

Andrew Sweet (16:24)

And

this also starts to feel like, I don’t know about the rest of you, but I never felt like I sold the deal because of Salesforce, right? And maybe I’m being unfair, right? But this is actually enabling and it is taking away mundane tasks so that I can focus on as a human, maybe what I do best, develop relationships and I’m not spending my time in the CRM. And it’s actually giving me recommendations. then the accidental by-product is much more

Mike (16:31)

Yeah.

Pete Reilly (16:32)

Right.

Mike (16:49)

And

Andrew Sweet (16:54)

accurate view of the sales funnel for executives so can get predictable business results. Nobody thoroughly enjoys it, I know, going in and adding a lead, converting it, etc.

Pete Reilly (17:04)

Exactly.

Alon (17:06)

Yeah, think

Mike (17:06)

the system.

Alon (17:08)

the evolution of all this ultimately feels like you’ve got a sales assistant effectively that you’re delegating all the stuff that can be delegated, right? All the stuff that you want delegated, right? It’s not the build the relationship part, but it’s everything else. Like before the meeting, send out a reminder, right?

Hey, if we got a prep for a meeting, go do the prep work and bring me, you know, draft the agenda on my behalf, right? It’s all, all those things. I think the interface essentially shrinks to more or less in its ultimate form factor to some kind of.

an action inbox, as we’re calling it, where all those activities are just easy for you to scan down, like what are we going to do now versus later, and press the button to get started with whatever AI assistant’s available for that action.

Mike (17:55)

And it’s, it, it’s not black box, right? Like, you know, that it might be suggesting this next action, but right underneath that, you can see the workflow that is the best practice of your business.

You don’t need to go there all the time, but you know it’s there, right? And it’s helping organize you and your supervisor and your customer and your marketing team all to get the job done

Pete Reilly (18:13)

once you start collecting all this context, you’ve enabled, you know, you’ve really made this agent sort of really enable the sales persons to sort of do get their job done. This also in many ways sort of becomes the system of record, not just for the, for the sales process and sales team, but for everybody else downstream in the organization in a lifetime of the customer.

Andrew Sweet (18:30)

Yes.

Alon (18:37)

Yeah, that’s right. think because context is king, you want to retain it, you want to ⁓ grow it as much as possible. So if you think of the full lifetime journey of a customer in any business, there’s sort of the initial engagement with the customer, qualifying them, prospecting them, winning them over.

And then that’s just the beginning. Right. After that, it’s all about, well, how have they engaged with us? What issues have they had? What have we done on their behalf? Right. What do we want to inform them of as the business grows and you’ve got new products or new offerings that you can market to them? It’s critical to retain all that information because the more that you retain, the more personalized their experience can be as you continue to work with them.

I would say this is really very new, right? This is three months old, like Claude Code’s been around much longer Codex OpenAI has been around longer, but the ability for them to stay on track, long running tasks, building things in, you know, what takes sometimes hours versus minutes, following a very detailed specification, that is only a few months old. This was not possible six months ago with, you know, at least not productively right now we’re sort of

expecting every time you run through what we call a module design, that you’re have a pretty good solution in kind of the first shot and then you’re going to spend hours after that continuing to tweak and tune the direction that you want it to go.

Pete Reilly (20:01)

If you look at what you’ve built Alon how long has that taken you and how long would you expect that would have taken you a year ago?

Alon (20:08)

Yeah, yeah, and I’m going to shy away from talking just about me because there’s more than me. it’s a small

Mike (20:13)

This is the system.

Alon (20:16)

team, let’s say a of three. Yeah, a team of three work on this stuff. So I would say we spent between Thanksgiving and Christmas, basically, building version one. So in that time frame, with days off here and there,

Pete Reilly (20:16)

All right, like, keeping moving.

Alon (20:31)

I would say it’s on the order of magnitude of 10x ⁓ to get to the same place with a team that knows what they’re doing, not learning things on the fly.

Mike (20:35)

Yep.

Pete Reilly (20:42)

maybe you could talk a little bit about sort of the stack that we specifically used and then how you see that, how do you think about that going forward as models change and things sort of continue to evolve?

Alon (20:54)

Yeah, it’s important to continue to put good abstractions in places so you’re not sort of locked into today’s technology. That’s probably the most important part of designing an architecture.

So we’ve leaned on sort of what I would call traditional patterns. Now what’s interesting is as you’re building these out, and so for instance, know, choosing the right model to do the job, right? So there’s a section where you can figure which model do you want to do this thing? And so, and your choices, you know, for a model that’s kind of, let’s say, draft the email versus a model that’s extracting fields from a transcription, right? It can be very different models. So that’s sort of one layer. I think the other thing is you have to

Pete Reilly (21:29)

Yeah.

Alon (21:34)

Acknowledge is and will get a little more to engineering side is there’s some core decision you have to make about, what’s the database strategy? What’s the front end strategy? What’s the interaction pattern between them? Like where are you caching stuff? How are you optimizing for speed and scale as the stuff grows? So, you know, if you’re doing proof of concept, you don’t care about any of that. You know, you just throw something in that generally works. But if what you’re trying to build is a reference architecture that will work for many kinds of applications, then you do have

to lock in those patterns. And we made some choices around using ⁓ an RM and using Postgres in this scenario as the database. some of those decisions are changeable. Postgres is just a relational database. You can plug others in. The fact that we use one model versus another is, you can plug and play those things. But other decisions are really your system level decisions. Like, I want communication to happen in this way.

we pretty much emphasize the idea of strong data typing front end to back end. So we used Prisma actually as an RM that lets us infer all the data types. the one place declares, what is the actual data type of anything? And that ⁓ ontology right continues to whether you’re in the front end or back end code. And it’s important because AI agents love that kind of consistency. If you can point it to the source of truth and say, is true everywhere in our project.

It makes it much easier for them to write code that doesn’t break, code that’s right the first time. So there’s a few of those deeper dive decisions that we iterate on and made so that to make this project friendlier for AI coding agents. And I think that’s a change of mind for a lot of folks who may have been religiously stuck in, we do things this way because that’s the way we’ve done them too. Now you have to think about, how good does that support the other capabilities of AI agents coding?

agents specifically.

Pete Reilly (23:29)

Right. What’s the coding agent stack for you guys right now?

Alon (23:35)

Yeah, it’s a heavy mix of cloud code, cursor, we’re a lot on Opus 4.5 where possible and some codecs. ⁓

Mike (23:43)

Mm-hmm. you

Alon (23:47)

5.2, I guess, is now. that ebbs and flows, I would say, as you design tasks, you, know, like, some variances, conveniences in certain places, and there’s disappointments in other places that we usually feel like, it’s so close to doing this well, but it’s not quite there.

But yeah, those are primary tools for me, at least.

Mike (24:01)

Yeah, and there’s

definitely mixing models like like, you know, there’s times when a certain model gets stuck and you, you know, it’s stuck and you go get an idea somewhere else, right? And or certain modalities specifically generating images right now, you know, Nano banana hands down. So but but as alone pointed out, that’s just a matter of saying vigilant, right? And it’s in and it comes from nothing more than knowing those models are available and trying them out. It’s not not like we’re ranking by some benchmark or using

Pete Reilly (24:17)

Yeah.

Mike (24:30)

You know, it’s literally try it out, see how it works. Okay. I get it and go.

Alon (24:32)

Yeah.

Pete Reilly (24:34)

It’s very easy to switch, right? Go ahead.

Alon (24:34)

I would love if the models didn’t get smarter so you could just get your favorite and get really good at it. It’s so annoying when you kind of know like, really need to try this other app because where I’m using doesn’t quite get this part and maybe this other model can do it. I’ll still use like 5.2 or Pro basically version of 5.

Mike (24:40)

Mm-hmm. you

Pete Reilly (24:46)

Yeah, it might be better.

Alon (24:57)

to build out thoughtful design for things that I then will use as input into actual coding. So I’m doing more brainstorming and spec building, I guess, outside of coding tools and then bringing that into the coding tool.

Pete Reilly (25:09)

Yeah.

as folks at enterprises are going down this path from what you guys can tell, is there any one platform, anyone hyper scale or anyone right? That they will generally have everything to do what we talked about, or do you need to be a little bit more flexible than that?

Mike (25:27)

definitely need to be more flexible. Now they’re all going to try, right? And that’s important. there’ll be a great solution, I’m sure that’s full stack from one vendor. But the best solution is at least for the foreseeable going to be a mix of capabilities or even a third party stack that doesn’t use. That’s right.

Pete Reilly (25:40)

But yeah, yeah. So have a best of breed perspective is just

what you’re recognizing.

Alon (25:47)

Yeah,

I mean, so like the hyperscalers of the folks, you know, the big tech companies, they’re going to have a lot of technology for you to leverage, but that technology is not really last mile kind of application technology in those cases, right? It’s the infrastructure building blocks around infrastructure capabilities. What I’m talking about, like, let’s say you are trying to, you know, deploy an ERP solution or an HR management solution.

you’re not getting enough from them, you’re getting it from an ecosystem of vendors who specialize, let’s say, in solving those kinds of problems. I’m suggesting that those kinds of vendors will increasingly need to modules that you can flexibly redefine for yourself without the traditional cost of.

you know, we need to be able to have lengthy design discussions and then map those onto what’s possible and what’s not possible within the software. I’m much of that should, it should be much more additive, much quicker to deliver and much more flexible to deliver.

Mike (26:57)

Yes.

Andrew Sweet (26:50)

it feels like a little bit for a business executive, this is starting to flip the whole paradigm of how we traditionally have evaluated build versus buy

much easier now to build and you get a specific system or solution for your business almost quicker and cheaper than if you were to buy it and then have to configure it

Alon (27:10)

Absolutely. Yeah, I think it totally makes that boundary much more. It shifted in the direction of bespoke is much more appealing, I think, in many situations. And then, you know, it may be the kind of thing where you can pick and choose your spots, right, where you want to be bespoke versus not.

There’s still a thing under the hood is a system of record that you want to trust. Right. And so you may be ingrained in systems that you’re not looking to replace because you’ve got a lot in there, but how you build systems on top of it. So even if you think about Salesforce, right, Salesforce is not the end all of all customer relationships that you’re doing. Right. You may have additional software for the sales team for productivity gains. You may have, you know, portal oriented for your clients. You may have, you know, marketing modules, right?

but you may still want to retain, you know, sort of the core data somewhere that you’re not reinventing. But in other situations, you could say, no, we can build it from scratch because it’s very bespoke, right? It’s a part of the business that needs its own thing. And suddenly that becomes not a million dollar project, you know, upfront, becomes build it, you know, for hundreds of thousands or less and then have a very rich,

kind of solution. ⁓

Mike (28:23)

You

know, I think there’s a chance that a lot of the people listening to this are going to think, yeah, a little bit of hyperbole here. You know, what these guys do is make software they’re all about AI. And I think it’s important to take a step back from that.

for a second and realize that there is a breakthrough that’s occurred. And it’s not a breakthrough in the headlines. In the headlines, the models have gotten 5 % better, 8 % better, 6.2 % better at this particular benchmark. The tip that’s occurred is that as an engineer using the tool, it is now producing something that is good enough for me to start with, as opposed to being something that I would just look at and say, no, no, I better start with something on my own.

If it’s going to do something good enough for me to start with and it’s going to have the power to iterate that over time, by definition, it goes 10 times faster than me to begin with, right? So it’s going 10 times faster. And it’s the thing that I’m willing to get started with and it helps me change it on ongoing basis. This is, it’s a step change. It’s a step change in the space.

Alon (29:16)

Yeah,

yeah, I call it a couple of things. One is I think the models are piece of the puzzle, but the harness around the model is just as important.

I think we’ve had a couple of years of maturity now of building harnesses around the model to keep them all sort of honest or keeping on track to a task. So it’s not just like a one massive prompt that I’ve engineered and suddenly everything comes back. It’s the fact that that product gets broken down into a plan. The plan has steps. Each step is carefully monitored and evaluated. So I think that layering above the model continues to mature and is a big part of the unlock.

⁓ So that’s a, you know, something that’s definitely.

evolving, right? It should only get better with additional engineering kind of focus on it. The other part is the entire flow of building software, right? It’s not just let me go engineer a final thing. It’s the, do I gather requirements? How do I simulate what users might want? Like part of accelerating this build out is to say, okay, I’ve got an application that’s working. Run it as if you were a user, make notes about where there’s a lot of friction or where it’s unintended.

intuitive, make those notes into effectively, know, PRDs or new specs that I need to implement and then go implement those, right? Now, that’s not all one go. You kind of have to guide it along the way where you’re asking it to make observations, write those observations down. But

you’re effectively trying to figure out how to delegate every bit of the work that you’re doing to where you’re just making the highest level judgment calls. Like, oh, I need more of this. I need to understand when a user is trying to do this thing, what’s the issue? Or, hey, this is a problem. when I’m doing this, I’m getting inconsistent results. Right? An end-to-end test. You’ll run it 10 times. Figure out why there’s inconsistency, right? Like, it’s a tricky timing problem or a state manager or something that’s not obvious when you’re just staring at code.

All those parts, the models have gotten good enough to where, and the harnesses around them, to where they are now highly productive. So part of the mission here is to figure out like, how do I leverage more and more of AI in the workflow as opposed to saying, okay, the bottleneck is now, I need to go get user feedback. How do I contact the hundred people to get their response? Well, there’s a few different ways. One way is you could simulate what a user might do. In others, could build an application that sends them a request to view a video.

to make comments, right, or simulate what the underlying experience would be. We’re testing all those things and they all incrementally add to the acceleration of how fast you could deliver working software.

Pete Reilly (31:49)

Thank you.

Andrew Sweet (31:56)

In that point around, what makes this effective is the scaffolding around the coding agent is actually applicable to agents in general. So I think there’s the ability to absolutely learn from what we’re learning with coding agents and apply it more generally in the need for that scaffolding. And so it’s a great place to start.

Pete Reilly (32:16)

If I’m running an organization as a big development team, maybe a software company or some other kind of company, and maybe I have a different sort of an offshore strategy and maybe I’m doing some things in Europe or I’m doing some things in different parts of the world, should I sort of look at that potentially differently because of this?

Alon (32:36)

Yeah, I think so.

Mike (32:38)

Yeah.

I think it’s a it’s a reboot of your ⁓ software development lifecycle and I don’t mean the geeky part of how to use GitHub. I mean, it’s a reboot of literally how you think of the timing of your software releases, how fast you think you can achieve what what markets you want to go after. I mean, it’s just it’s such an enabler that you you may you may be looking at parts of your space that you would say no, I don’t I can’t tackle that too much infra is needed too much investment but

Pete Reilly (32:53)

Mm-hmm.

Mike (33:07)

Now, if this can be automated, how fast can you move through that? Right? So I think it’s a big reboot. And then when you start talking mechanistically, all right, if you say, yeah, Mike, I know I’m going to reboot my business, but what do I do now? Right? What do I do right now? I think that’s where some of the kinds of things you’re pointing out. Look, mean, any developer anywhere in the world can use this technology. Maybe not in China, I guess, in some situations. But it’s broad use, and intelligent people are going to be able to use it everywhere.

⁓ So it’s not excluding any one strategy that you already have in place. What it really is doing is saying, where are you focusing your efforts on? Designing your product features ⁓ or on good coding architecture, right? That’s the reboot. That’s the sort of bitter lesson here is to spend more time figuring out what you want out of the solution and letting the AI take over more of the work rather than getting stuck.

in the minutia of what you think you needed to do because of the strategy that you had before you could take Gen. AI into account.

Pete Reilly (34:07)

Yeah.

Andrew Sweet (34:08)

In addition to

the velocity, you know, it does shorten the distance. I’m not just talking geographic distance to offshore kind of development, but the distance between development and the business. You’re closer to the business and you can iterate much tighter in your own time zone and sometimes even in your own head, right, that you don’t have to even go to other humans. And so I do think it does and it will cause a fundamental rethinking of how

Mike (34:26)

Mm-hmm. Thank you.

Andrew Sweet (34:37)

offshore and remote developments done because again that distance can be shortened and oftentimes it’s that distance that creates the disconnects between what the business needs and what’s delivered.

Alon (34:50)

I mean, I think it’s very fundamental. It’s kind of like the rules of change, right? You were playing tennis and somebody then inserted pickleball rules and changed the court on you and all that. It’s not the same thing. I mean, the end result might be, you’re winning by delivering something, but first you should set your aspirations to solve the problem in a way that gains from the machine intelligence

Mike (34:58)

No.

Andrew Sweet (35:00)

Right.

I’m wondering as I’m listening to this, is there a potential now that business executives are gonna have to look for? We used to talk about shadow IT and what that typically meant is marketing would go off and hire a vendor and they’d go build a project without IT being involved. It seems like more and more, maybe not today, but the opportunity for

now marketing to go off and actually write their own applications, not even get a new vendor, but actually develop applications. And so that level of governance, I think is going to become more more important, but open to thoughts.

Mike (35:42)

Okay.

Alon (35:52)

Yeah. I mean, I think there’s going to be some fascinating opportunities ahead where, for example, software vendors have to consider the

downstream effects of whatever they build in a sense that whoever wants to customize or create bespoke iterations or, you know, mashups of their.

of their solution, how do they enable that, right? So if you think in kind of the, you know, the roadblocks of paradigm of, you know, lots of folks can create content on top of the underlying platform.

⁓ You know, every business has got bespoke needs. And if the, if the underlyiing software you’re delivering has a set of well-known APIs, but more than that, right? If it has coding reference examples and best practices that a coding agent can latch onto, I think you’ll see a ton of, and I don’t know if this is shadow IT or just the IT, right? The idea of customizing the CRM or customizing your ERP becomes much more like, okay, ⁓ there’s, I choose

my favorite vibe coding tool. I feed it, you know, whatever the, my software vendor gave me as the bundle. Like, Hey, here’s the thing you give your vibe coding tool and then just ask, ask for, you know, the new custom report that you want or the new, you know, entry form that you want. And the vibe coding tools can adhere to the guardrails that are set up by, know, by your software and the testing pattern that’s set up by it. So I think we’re going to see a world where a lot more customization will happen, but it’ll be so much easier. Right.

Like it won’t require.

the long, like as fast as you sort of can nail down a good spec for something and design it out. Basically that’s the hardcore after that, it’s sort of automated and you wait an hour or two and you probably have a decent prototype where you’re trying to do. And then you go through some productionizing that whatever you did in terms of figuring out, well, how does this get deployed and how does it get managed and so on. Right. And there’s agents all along the way to help you do that. And I think it’s on the software providers to ensure that that ecosystem works as opposed to

you know, sort of trying to deliver it all like in a single process. Here’s our solution. It’s done.

Andrew Sweet (37:57)

Yeah.

Mike (37:59)

Yeah, one of those

famous phrases in 2025 was AI slop, right? The idea that you generate it, you ask a five word prompt and you get a five page document and you run around with it like something that’s amazing. And that’s happening in software too, right? That you can ask, make me an application that and you will get an application that, right? So really quickly something comes out. That doesn’t mean that it’s something that’s production worthy.

Alon (38:23)

think we lost them.

Mike (38:25)

that can be used and that’s going to be part of how the future flows, right? It’s much more likely that you’re going to get a good starting point that then needs those best practices from IT and the infrastructure compliance and all that to be able to fully carry it out.

Alon (38:40)

Yeah.

Yeah. Well, my, my take maintaining it, cause it’s a really interesting question, right? If you think about it, it’s sort of, you know, AI coding agents all the way down. The, the, When something is broke, like historically you would say, okay, who’s the right person on the team to fix that? Who’s got the insights on how to fix that? Now we’re sort of looking at like, okay, who’s got time and everybody does roughly the same thing. They go in and say to the AI, Hey, this is not working. Here’s my prices or logs or whatever. Tell me what’s going on.

Mike (39:09)

Right.

Alon (39:10)

You

and so you end up being much more of a generalist as a person working on it.

I think that the implication of that to customers is pretty interesting, which is one, your own IT department, to the extent you are provided, like when you’re customizing thing, if you’re provided with the right kind of context, so again, this goes back to what does a vendor publish to the customer? Well, if the vendor publishes a variety of things that includes source code level or least abstractions or pseudo code or something, IT can go ask this question, hey,

our customization XYZ broke, fix it. Here’s a reference to the latest API changes or whatever. So I think the reliance becomes less on human understanding of the underlying code. It’s almost like you’re looking, everything you look is through an AI lens. And you say, look, the underlying code is the source of truth. Nobody reads hundreds of thousands of line of code except if you’re an AI agent. And what historically we used to try to do is

Mike (39:59)

Right.

Alon (40:11)

write a lot of documentation to sort of, you know, take a hundred thousand line code and explain them in 2000 lines of English. And, but by just doing that, you’re sort of miss a lot of the details now you’re like, no, no, the agent is going to go read the, you know, the underlying explanation of every line of code and give you a very credible answer as to why you’re seeing a change in behavior or whatever the outcome is. And that to me is, is just, you know, just as big as, as writing this stuff from scratch. Cause suddenly you don’t have this around like

Pete Reilly (40:35)

That’s really interesting.

Alon (40:37)

I just written 10,000 lines of code. How will anybody support them or understand them? like, yeah, that’s not really.

Pete Reilly (40:42)

Yeah, in some

ways, I think what you’re saying Alon is in many ways, it’s way easier if I have sort of there’s a premium on having access to the source code. Right. But if I’m buying from a vendor, I have to get in line and I have to submit a feature request and I would provide the source code. I can make the change like that.

Mike (40:52)

Mm-hmm.

Right?

Yeah. And so it’s almost like an iceberg, right? At the top is what the human provides. This is the expected behavior. When I do X, I want to see Y. And then, you know, all the way down to the bottom is the source code that does that, right? And so in between what you do is you get the AI as you’re building it, you get it to keep that list growing of the expected behaviors, right? And you get it to materialize that into tests that it can apply automatically. So now it becomes straightforward for anyone.

Alon (41:01)

Yeah.

Pete Reilly (41:08)

Yeah.

Yeah.

Mike (41:28)

who has AI coding skills, not a CRM development skills, not database skills, not networking skills, but AI coding skills to be able to make those fixes. Right.

Pete Reilly (41:38)

Yeah, which are much more portable, right? ⁓

That’s really interesting. Obviously, continues to sort of in the build versus buy favor. And it pushes in the region.

Alon (41:47)

Yeah. Well, yeah, I mean, I

think that as you’re saying out loud, like, Maybe that’s not the best way to set this up because build versus buy in the historical world meant very specific things like you’re buying a whole thing or you’re building a whole thing. I think we’re talking about a hybrid thing, right? You should be buying the pieces that you don’t want to rewrite the pieces that are, you know, sort of evergreen, right? The things that need to be in place like

Mike (42:02)

you

Alon (42:12)

things that do enterprise authentication, things that do database management, all the kind of you want to build anywhere we have a custom need, where your need is a little different than that. Yeah, yeah. So what you’re trying to buy are the pieces that enable your customization, right? a thing, let me buy a solution that’s closest to what I’ve got, and then let me try to hack that pieces up.

Mike (42:16)

A database, right?

Pete Reilly (42:17)

Yeah, yeah, yeah.

or it touches the business process of the job or, you know, yeah.

Yeah

Well, because the software

providers generally would tell you, you need to change your process to match the software, right?

Mike (42:41)

Right.

Andrew Sweet (42:42)

Right.

Mike (42:43)

Customize yourself.

Alon (42:44)

That’s right.

Pete Reilly (42:44)

Yeah, right, right. Exactly. You need to adjust,

right? I think this completely inverts the equation.

So if I were going to sort of bring home what I think of the key points I heard, one, just the whole build versus buy equation is being completely rewritten. That affects not only what I attempt to build myself versus what I buy, also maybe even where my development team is and how do get them closer to business and who product managers are. And there’s a bunch of sort of interesting things that will be driven from all of that.

Number two was, I think, Alon, what I saw in a lot of what you and Mike and Mike, what you guys have done is designing much more around the human and an assistant for the human trying to get a job done as opposed to maybe feeding the system, which I think has been maybe historically how a lot of these CRMs have seemed to me. Context is king. I think we said that a number of times and it’s the fuel for the agent and to the degree that you can then capture that and we didn’t show this, but

doing the email from the application, doing the phone call from the application. And in all that context, you sort of immediately available as key. And then once you’ve done that, the CRM, depending on your business, can sort of become this front door to the system of record because that context has value at the beginning of the relationship and ongoing as I continue to evolve my relationship with the customer. Did miss any?

Alon (44:10)

Yeah, that’s good stuff. think the buyers, as the equation moves, you’re looking to buy flexible building blocks more than a finished product because you expect to be able to customize your finished product to your needs.

Pete Reilly (44:25)

Well

guys, this has been awesome. Appreciate you sharing that with us. I think it’ll be great for our audience.