We Spent $3,300 on AI Agents – Here’s What Architecture Firms Should Know

What are we spending today on Codex?

And then last month we spent, even after putting the limits in, we spent about 3,300 more.

Surprise, double yes.

I dropped this big spreadsheet and I said, I just need a little bit of advice.

And it's just, it ate my limit for the day and I was sad.

Don't be sad.

For example, why not put together, you know, last 17 years of your company history proposals, categorize them as proposals you didn't win to proposals you actually won and they turn into contracts and then build the agent that specifically works with your team to create proposals and get you 80, 90% there to win the contract.

Foreign.

Welcome to Design Under Influence show here with you host, Alex Osenenko.

I'm in charge of AI labs here at ARC it.

Why we call it labs because it's frontier work.

We're learning something new.

We're deep in IT right now, integrating IT into our own operations and into our customer operations too, which is we serve architecture, design, engineering companies.

That's who we focus on.

I'm joined by my co host, Boris Rapport.

He's a founder CEO of Ark it.

Welcome to the show.

Hello.

Hello, Alex.

Long time.

It's been a long time.

What we want to do for the next set of episodes is really to help in inform our audience on how to actually deploy AI.

What is actually working today, what may work tomorrow is an important question, but it's today and how they should treat this new technology, how to adopt it correctly.

And so I'll just open quickly with what the adoption stages are on a very sort of high level and then you and I will dig in.

But today, you know what we're going to talk about, Boris?

What are we going to talk about?

Alex?

You already know what we're going to talk about.

I have no idea.

Okay, yeah, well, we're going to talk about money, money, money, money, money.

And so we're running tons of experiments and actual workflows for both ARC IT and our client environments and some other businesses that we integrated into this as a pilot and our spend is skyrocketing.

And so today we're going to help you take control of that or at least know how to think about it, because that's something going to, that's definitely going to come up down the line for your company.

So let's go back and say Stages of adoption, the way they should be is your people first begin to experiment in their own personal accounts and eventually as a leader or is a thought leader in a company, you realize we might as well take control of this and deploy AI company wide.

What is AI going to do?

Is it going to replace humans?

No, it's not going to replace humans, but it's going to give us the freedom to do things we love, like building buildings, like designing buildings.

Right.

And spaces and whatever it is that you signed up to do and all your schooling and what you love to do.

But then all the admin stuff, that is real work.

And for us, for the IT company, for architecture space, you know what we want to do, we want to consult our clients on strategy, we want to deploy new tools, we want to test time saving technologies with our clients and we don't want to do all of the minutiae and admin stuff as well.

So that's how we're doing this ourselves.

We're deploying this inside of our company.

And so the first stage, as I said, is folks gonna start experimenting.

Then a company leadership realizes we should probably put some guardrails around it and this is what you call us or somebody like us and say, hey, we need help putting AI policy.

Yeah, you can prompt ChatGPT and say, hey, just write my AI policy.

What kind of a job will it do, Boris?

It'll do good enough to fail job, let's put it that way.

Good enough to fail.

Exactly.

It will read like the most knowledgeable document you ever laid your eyes on and you'd be so impressed.

And then you run it by your legal counsel and you know, and the rest of the people and then just be ambiguity and stuff like that.

So with AI, I repeated that a lot of times.

You ask a stupid question, you get a stupid answer.

Now this question is not stupid.

Hey, help me create a policy for my company.

The problem is it doesn't know the concept, doesn't have anything.

What's already in your head and what should be in your head because you're not experienced with IT tool set anyway.

You hire people, firms like us who specialize in this sort of work because we are IT and technology focused company.

So we understand how to keep your company secure, but also give you the freedom innovate, operate and so on.

So you deploy the policy, you start the tooling rollout, let's say whatever you want to approve for the company.

ChatGPT, Claude Copilot and then you go through some training, your team begins to experiment and there'll be some hiccups, but for the most part, you will love the outcome of it.

Now, the next phase from there is agentic phase.

So the next phase is you got your basics in place and you're ready to actually have AI take care of some of your tasks.

Right.

So it's not just a chat where you're asking a question and getting an answer or asking it to create a document.

Here you're maybe creating an agent that is your chief of staff, possibly, or maybe your administrative assistant, where it does things either on schedule or in the background on your behalf and then gives you the output without you prompting it every time.

For example.

Go ahead.

For example, I have one that runs every morning.

Actually, no, I have one that runs every two hours.

It looks through all my emails, it looks through all my teams messages, it looks through all of our company public teams messages, it looks through all my fireflies notes from the meetings and basically outlines every decision that was made in any of those threads and creates a little table for me.

And then on top of that, it records it into a log, basically.

So by now I have a history probably of four months of decisions made by our employees or us working with clients together in a lot so that we can reuse that data later for some other things, like putting together sales meetings and business review meetings with our clients.

Yeah, that's excellent.

And so that's.

That really aids your executive work and the way you guide the company, which is super important.

But also there's like less grandiose things.

For example, why not put together, you know, last 17 years of your company history proposals, categorize them as proposals you didn't win to proposals you actually won and they turn into contracts and then build the agent that is specifically works with your team to create proposals and get you 80, 90% there to win the contract based on all this knowledge that you currently just virtually impossible to access because it's sitting on service somewhere.

And so for that using a chatbot, that's not going to work.

Right?

You need to build agents so that agentic future.

And don't be afraid of word agent.

I used to be afraid of it.

Agent Smith, you know, Matrix.

That's sort of like.

That's always my association with the word agent.

So it's not, it's not an elegant word, but it is transformational once you start building it like we do.

However, what comes with Agentic Era or Agentic sprint into new innovation for the company?

I'll tell you, Boris, amazing things happen, okay?

Amazing Things you can do within your organization.

Okay.

Amazing things.

Think of it as.

The way I think about it, is I have a room full of talented developers who can code anything in any language for anybody and design things.

But I really need to know what I want and just have a very clear thinking.

So as a human, I drive and control this, and if I put in garbage, garbage comes out.

But there's another big aspect with agentic things.

And you'll find yourselves, ladies and gentlemen, we will find yourself there.

And that is what Boris.

Okay, so what are we.

Let's just say, because this is just.

Because it's probably gonna change a week or two.

But what are we spending today on?

Codex.

What's our monthly.

Because we limited ourselves, didn't we?

Yeah.

So we have some spending limits in place, and it's only two of us have access.

Right.

The rest of the company does not yet have access to these agentive tools.

And then last month we spent.

Even after putting the limits in, we spent about 3,300 more.

Surprise.

Double.

Yes.

3,300.

$3,300.

Yes.

Correct.

Okay.

And that's just between the two of us.

And we're.

We are a small business, but these things can get away from you fast.

I think one of the.

At one point, I felt like I was addicted to talking to an agent, which was kind of interesting in retrospect, because you're like, oh, you can do this for me and can do this for me.

And that's where the costs actually build.

Then you get the bill.

So we'll help you optimize and think through this.

But even if your organization's still in the first stage, which is policy, tooling, rollout, or training, experimentation, or just starting to build agents, you will get there and you will need to know what we.

What we already are experiencing.

Okay.

And so what that is is tremendous efficiency gain.

So, for example, let me give you one simple example that I asked you today, boys, on our one on one, and you answered, you have a little agent that runs the bonus calculations for you.

Right.

Okay, so what does it take you to do manually, which was like, what, three months ago?

Two months ago?

You were doing it manually three months ago?

Yeah.

Yeah.

All right.

How long did it take you?

Two hours.

Two hours.

So copilot does it now perfectly executes it perfectly, every single time automatically.

And how much does it cost us per execution?

Per payroll report, I'm hoping that it.

Costs between five and ten dollars.

But we'll see.

This is something that copilot is now with copilot charging.

Now we need to identify what the cost is because I haven't run it in a paid environment yet.

Yeah, we're assuming it's six bucks somewhere along those lines.

Because part of it is because it's a skill that's been already created and saved as a skill.

So Boris, maybe you can talk a little bit about that optimizing so.

Because if I upload, let's say, oh, by the way, so that's to add to our Codex spend we have now, Cowork spend.

Okay, so it used to be all free, but now the free is over.

But I guess help us line up, help people understand that when you just do it.

And let's say you do a prototype work, right?

So I'm going to upload a big spreadsheet, I'm going to tell it to go do a couple of things, I'm going to see what it finds, it could find.

That's all.

I call that experimentation.

Right.

And that's very costly, right?

That's very expensive.

How do we think about putting framing that into a lower cost process?

Once you conduct an experiment and you like the output,.

There's a few methods, right?

But one, like always hanging fruit, is making sure that as soon as you've put the process together, you've kind of given the instructions, you've iterated on the instructions, you've iterated on what the outputs are.

You finally have a set of instructions that works to produce the outputs that you were looking for and the output looks great.

You want to turn that into a skill.

So every agented platform, whether it's Cloud Cowork, whether it's Codex, whether it's Microsoft Copilot, Cowork, even the Google, I forget the name of it.

It used to be Anti Gravity.

I don't know what it is now, they always change it.

But even the Google platform that you can run agents in basically allows for skills.

Skills is an open source format.

So basically any of the platforms, you can interchangeably export and import skills between them.

So you'd want to, once you have that session where you got the right inputs and the right outputs, you'd want to create a skill.

And there's a very easy way to do that, at least in Codex, Copilot, Cowork, or in Cloud Cowork, you basically just tell the system to create a skill.

You just say, based on everything we've just done, go ahead and create a skill for this task.

Name it this and then it will go in and create a skill for you and you'll be done.

Let me give you a simple example so all of you can connect with it.

So, for example, you want codecs to help you, or codecs or any other agentic tool to help you change a page on your website or design a new page for your website.

Let's say you have a new project that you just stood up and you want that to be showcased on your website.

Okay.

Used to be to have to call a web designer or somebody internally who does that for you, spec it all out.

It's a little bit of a pain in the process.

Now you can just tell Codex, hey, look at my previous projects and how they're displayed on my website.

Here's the new one, here's the link to all of the pictures and documents.

Go ahead and put together a prototype of how this project will look on my website.

So it's gonna go do that and it's just gonna go work.

You're gonna go back, go get some ice cream, whatever, take your kids out for a movie, come back and it's done.

Your prototype is stood up.

Now, that costs you about 40 bucks.

Could be 20, could be 10, could be 50, but okay.

But it also probably suggested a page design, changes to layout, design changes for your overall project, to feature this project, whatever the case is, whatever you told it to do.

So you've done it with the prototype.

Next thing you do, you say, okay, this looks great, let's change this.

Okay, let's put a contact us, let's do this.

You make these changes, they're incremental and small.

They're not very costly at all because all the scaffolding has been stood up, design has been made.

So all these changes are virtually a nothing burger.

And then as a cost from a cost perspective.

And the next thing you do, you tell it, Boris, what?

Put it into skill.

You have it all done and you like the output and you've made all the changes.

You tell it to create a scale, create a skill.

So next time, Now, I can't quote you exact costs because that's very difficult to do.

We'll be running lots of experiments on these shows in the future, on our shows, by the way, because Boris and I are gonna start digging deep, deep, deep, deep, deep into a lot of the AI related stuff and how they work for our clients and for us.

But suffice it to say, it will be a lot less next time you have a project.

And so next time, let's Say you stood up another project and you're ready to publish it on the website again using this.

Just an example for you.

Small example, you'll say, hey, can we add project for Main Street 2 to 2 onto our website?

It will go and say, okay, let me start thinking.

And it's going to go, see?

Okay, I see.

I have a skill for putting projects on a website.

I'm going to go ahead and use that skill.

Boom.

It's going to be like five bucks maybe, right?

Don't know.

Don't quote me here.

I don't know for sure.

Depending on the kind of stuff you have, if it has API already established to your website, it doesn't need to do a lot of the engineering to get the solution out.

It's just repeated steps.

But the cool things I want to point out is, number one, I've never actually went and named the skill.

Boris, you have probably because you think like an engineer.

I think a little bit differently.

I think more on creative side.

I can get engineering done, but I prefer my minions to do that, which is codex and cowork and all that stuff.

Anyway, I don't name them, but I invoke them via prompt.

Do you name your skills?

Not often, no.

Usually it does a pretty good job of naming the skill.

Sometimes you want a more accurate name, so I probably have to do it once or twice.

And basically after it creates a skill, it says, hey, I created the skill.

Here's what I named it.

You're like, okay, this is either a little too long or a little too non descriptive.

Or it could clash with another skill that you may have so you can tell it to update the thing.

So with this example wrapped up, here's the idea, the big idea, big takeaway from this show.

Okay?

You will be here, you will get to this phase and this phase will be cost optimization for your AI deployment.

Okay, again, spending $7 to save Boris 2.

I mean, it's probably cost the company, I don't know how much thousand dollars an hour or something like that for him to do this Payroll.

Maybe more, because opportunity cost.

Even if he goes out for two hours and walks his dog.

He doesn't have a dog, but walks his kids and just drinks coffee.

Like, he'll be so much better for the next client meeting, right?

He'll be so much better for the next project.

This is stupid work.

Payroll.

I'm not calling payroll people.

I'm talking about the actual taking numbers and putting them in different columns.

Yeah, Basically evaluating a spreadsheet and putting data into Another spreadsheet, right?

Nobody should do this.

We should all do more important work.

And so for six bucks we'll take it all day long.

But you will come to a stage where a lot of your team members will experiment and you should give them some space, but definitely set a budget and then you'll start coming up with cool things for the company.

And Boris, talk about a little bit about skill sharing and how you see being deployed here at ARC IT and other companies that you work with.

So as I said, Skills is an open source format, so it's very easy to actually export and deploy to either other tools or share with other people.

So at ARC IT right now, so we have a library of shared skills to where if somebody, one of our people or myself, creates a skill that we think can be used by more than one person, we just upload it to that folder.

It's basically just a teams folder.

And so we can ask the AI to export exports it into a zip file.

That zip file contains everything the Skill needs.

And basically we can then another person can just grab that zip file and ask their agent to import it and it's available to them.

So it's very easy to use and share.

So yeah, having a library, I think as a best practice, having a library of skills that you can share across your organization is very important.

Similarly, having a library of prompts may not, may not be cost optimizing, but definitely like prompts that can deliver results consistently and multiple people can use them.

You should a library of those as well.

So yeah, so this is, I would compare this very similar to an organization who organizes SOPs best practices, standard Operating Procedures, which is SOPS and their process.

And it's trainable and it's repeatable and each job description understands, each individual understands their job description and they actually have instruction on how to execute what the process is for the company to deploy certain things.

I think Skills is the future of that.

And this is one significant cost saving method that you can deploy once you go agentic for your AI using your company.

But there's more stuff we do ourselves here, or at least experiment with to do some cost savings and that is more advanced.

So if you want to listen to it, you can keep listening.

But I think we're going to go into the forest a little bit for a few minutes here.

Boris, we talked about open source models.

We talked about putting compute inside of the organization, having a computer or multiple of them crunch the data.

For AI, we talked about also building infrastructure for data using existing tools and only Letting AI interpret it instead of having to build the database every time.

Can you talk about those things, how you think about them right now?

Yeah.

So I.

So the very easy thing to do, it's not very easy, but it is an easier thing to do once you have your kind of codecs or copilot cowork set up with connectors is just use that for everything.

But because you can just prompt it and it goes and does whatever you need it to do and then you can schedule those workflows and it goes and does whatever you need it to do.

Now again, that comes back with costs, real financial costs, and those costs can get away from us.

So what we're helping, what we're doing internally and helping our clients as well.

I think one of the steps is, or the first step is, hey, can we use other tools to actually automate some of these workflows, as Alex said, and then have AI just review, interpret, get contacts, provide contacts to these workflows and then the workflow then takes it on again.

And let's say you ask AI to create a Word document.

So you ask AI to go fetch some data and then based on that, create a Word document of some sort that's an agentic work that make us some money.

Depending on the plan that you're using, we can then basically create an automation that will do that work.

Except AI is not going to do most of it.

There's going to be an automation that goes, gets the data, interprets the data, then sends it to AI to actually do more interpretation and more detailed work and create an output format.

And then there's another automation that doesn't cost you anything that then grabs it and puts it into a word format.

There's a lot of ways to kind of engineer this in a way to where it starts costing you less, but achieves the same results.

Right.

In order for that to work though, you first need to kind of go through the initial process of, hey, can I automate this fully with AI iterate?

And then we can convert that workflow into a different automation.

So in my example of proposals, Proposal Helper, for example, if I have seven years of proposals, I want to integrate into Decision Assistant for my next proposal, Boris, how do you recommend?

So in other words, you can ask AI to go review all of those proposals, pull out variant components like, hey, here's the client type, here's the proposal type, here's the size of the building, blah, blah, blah, whatever the metrics are relevant, apply them, and apply to this particular quote that you're building proposal and then having AI do all that work will be expensive.

How would you think about building that agent that function with minimum AI involvement for cost savings?

So I think maybe we need to kind of go into and explain for the audience why that would be expensive.

Right.

So if we kind of go back to the basics, what we're doing is we're using tokens.

And what tokens are basically, in kind of layman's terms, is basically a word.

So a token is award.

So token of input and a token of output costs you some money.

In this scenario that you're describing, where basically AI is going to go, is going to take all those seven documents,.

Is going to look seven years worth of documents.

I'm sorry, seven years worth of documents.

Maybe for simplicity's sake, we just say there's 10 documents, right?

It's going to look at those 10 documents, is going to interpret every word and then it's going to process it.

Right?

So let's say each of those documents is a thousand words.

So you're basically interpreting 10,000 words, which is basically 10,000 tokens as an input that you give to AI.

Then it's going to provide an output of, I don't know, thousand to fifteen hundred words.

So that's fifteen hundred tokens that you are outputting.

With some of these models, you're basically, you're paying like $50 per million tokens, like the latest cloud models, which is hugely expensive.

So you can see how this is just kind of a baseline example or very easy basic example.

But if you're doing the same workflow, even if you're using skills and you're doing the same workflow over and over and over again, you can see how it's still going to be the same amount.

The cost adds up.

Yeah, the cost adds up because it still has to.

It still has to review every word, which is going to be 10,000 words in our case.

Correct.

But if we can put an automation in place where those 10 documents are stored somewhere, goes in and indexes them and puts them in a database that lives that you're hosting either within a Microsoft or AWS or Google environment.

We prefer Microsoft because they already have the tools in place to make it easy.

You basically index it, store it in a database.

So all you're really paying for is the minimal amount for storing that data in a database.

And now that data is already what is called vectorized or easy for AI to understand.

So it's not really words, it's actually just math.

And it's a lot more efficient for AI to look at your documents that way than it is to just interpret your words.

So you're saving then a lot of time or a lot of, you know, tokens on data interpretation.

The outputs are still going to cost you the same, but the inputs are now going to cost you a lot less.

That is very cool.

So in example, where I put a spreadsheet that has tens of thousands of entries, maybe hundred thousand entries, because it's a full catalog of some manufacturer for some tiles or whatever.

And I ask it, hey, here's my specs for this particular design I'm doing.

Can you pick top three tiles, colors and variants and sizes and thicknesses and.

And so that's going to be what, like 150?

That's going to be expensive, right?

Yeah, ultra expensive.

That's the thing.

So it's the down and dirty way to do this would be to go to their catalog, pick out only the colors that you're considering, right?

With only the sizes of the tile that works for your project, extrapolate them and then put it into chat and then ask it to evaluate the colors and give you a recommendation.

That way the subset is much smaller.

And that's.

As a human, you've made decisions to make it this smaller and the output's gonna be this much better than giving it this hundred thousand entries Excel spreadsheet that not only has tile, but that manufacturer does cement tiles and all the other things that don't matter because you were.

You want a quick answer?

That's where the cost runs up.

Huge quick answer.

And I've learned that the hard way.

I mean, the hard way.

I've done exactly what I just told you.

I dropped this big spreadsheet and I said, I just need a little bit of a device.

And it's just.

It ate my limit for the day and I was sad.

Don't be sad.

All right.

I think this is pretty much all the time we have for today.

I really want to get into with you because I'm actually.

Dude, I'm learning.

I'm learning as much as our audience learning because you're so deep into it and I am as well.

But it helps to elevate and talk to each other even though we expose audience to our musings.

But I hope it's useful for everybody.

I am definitely learning and that was an interesting one.

But I want to next next shows, I do want to get into some of the experimentation stuff you're doing with localizing compute and using potentially open source models, LLMs.

Of course there are security concerns.

Security is number one priority for us as an IT company.

So I'd love to quiz you on that, but that's.

I think we'll leave that to the next show.

So I appreciate everybody watching.

Boris, any parting comments?

No.

Thank you everybody.

It's good to be back and yeah, looking forward to providing more wisdom to everybody.

More valuable wisdom.

Yeah.

And thanks a lot for sharing.

I know your day is busy, but we couldn't replace this with a five dollar spend on Codex.

You actually have to show up on these things and talk and share your knowledge well.

So hopefully Codex can do the other work that I don't like to do because I love doing this.

Yeah.

Awesome.

All right, thank you very much.

You guys need our help.

Get Arcit.com we'll help deploy AI with your company.

We help manage the costs and we definitely take care of your it.

See you next time.

Thank you for watching.

See you.