Introduction

My name is Sahil. I'm the CEO and founder of Solari.

My goal is to help non -technical teams sort of get their heads around AI agents and get the AI assistance that they need.

So I'm going to be giving a little bit of a sort of framework for how I would encourage non -technical folks to think about this unique opportunity, this unique moment of time that we're in.

So just to introduce myself, I actually started my career in non -technical. I went to a liberal arts school, and after an internship in Silicon Valley, I sort of saw just how valuable it is to have at least a few technical skills,

started sort of forcing my way into the space, started with data science, then in product management. While in product management I started teaching myself how

to code, was doing probably what is now called vibe coding, but I was doing it with like the very first version of ChatGPT and that product that I built as a side project got me into the YC Winter 24 batch which is when I sort of

of started my startup career so since then I've been helping non -technical teams figure out what are the agents that they need and understanding sort of one of the blockers and one of the points of friction that prevents non -technical folks from getting all the upside from agents that they could so

this is a talk that is sort of I'm a summary of all of those learnings so I'll be kind of going through things a little bit quickly but feel free to catch me after the talk I'll be hanging out for a little bit and don't worry about taking pictures of the slides if you want them again just come find me I

Why you’re not behind: empathy, FOMO, and the reality of AI adoption

can send them over so I want to start with just like a moment of empathy and hopefully optimism for the non -technical folks in the room has everybody here

heard of open claw or put your hand up if open claw sounds familiar to you how many of you built something with open claw or played with it yeah so there's still a lot of hands not not up yet for folks that don't know open claw was a

github sorry open source project that went viral on tech Twitter a while ago that made it really easy to whip up your own personal assistant. This is the GitHub repo of how to get started.

And if you are technical, if you are comfortable working in Terminal and with CLI, integrating with API keys, even buying a Mac Mini to run an assistant on, I don't know if you've seen, there are no Mac Minis in any Apple store in a major city right now, then OpenClaw was very, very exciting.

And that was a lot of the decision that went behind OpeningEye's decision to acquire them. But if you are not technical, this looks very, very confusing and scary.

And I think that this leads to a bit of a feeling of FOMO and anxiety around this whole wave is passing me by and I've missed the wave entirely.

So first I want to just acknowledge that I think that's a very normal thing to feel. You are not alone. But I also wanted to give some optimism.

Has anybody seen this graph? This went viral on tech Twitter in the last week or so. You, okay, anybody else?

Any guesses, so the folks that have seen it, please don't say anything. For the folks that haven't, any guesses as to what this represents? Good one, population?

Okay, so every dot, like it says, is around 3 million people. And what this shows is roughly what is the state of AI adoption in the world today?

So what it shows is that just under 84 % of the world, so around six billion people, have not had their first AI adoption event ever. They have not used any tool that we would consider agentic today. never used ChatGPT, never used Claude.

They've So you've got about 16 % remaining, just over. Of that, 16 % of people are only using free versions of ChatGPT or Claude.

So they're not paying for anything, they're just using the free tools that are out out there in their everyday life.

And of the remaining, 0 .03 % are paying roughly $20 a month for something like the pro version of these chatbots or for something like Cursor.

And of the remaining, 0 .04 % are actively using tools like Cursor, paying a lot for it, and Cloud Code. So if you feel like you're behind,

I would stop, take a breath. there is a lot of opportunity left in the space the analogy I like to give people is when the Internet came out and it had this level of adoption so roughly

15 16 percent of the world was using the most basic version of websites it was 2005 so if you think that you're behind I would encourage you to just take a moment take a breath it's as if it's 2005 and the Internet came out think think about how much opportunity there was after that.

That's not to say that the people who were up to speed early were not heavily rewarded for it. Of course they were.

But it would be equally naive and I think factual, or non -factual rather, to say that the opportunity has passed you by.

So I just wanted to start with a moment of empathy that I think that algorithms and tech Twitter can make it feel like everybody using AI and all the good products have been built already, but the data proves out that there is a lot of time

and a lot of opportunity left to be captured.

Talk roadmap: three frameworks for non-technical teams

So what I'm hoping to go through now for the non -technical folks is really three frameworks or three sections. The first is how I would think about on -ramping onto agents.

I'm going to go through some of the basically the five options that I see in the market and I don't want to prescribe one over another. I want to give the pros and cons

that I've seen working with teams for about a year now and the idea is you can kind of graft your team circumstances on to the pros and cons and decide what works for you then I want to talk about once you've gotten onto that on -ramp how

can you effectively integrate those agentic tools into your team so that you avoid what I call the prototype wasteland where you've got a very good assistant that kind of works that nobody uses and then lastly I want to touch on on some advice for the aspiring builders and makers in the room.

Not very long ago, I was in your position. I was non -technical, looking at the startup space from afar, figuring out how do I get myself in there. And so I just wanted to give some words of advice to you as well. Great.

Framework 1: On-ramps to AI agents

So first is the on -ramps. I see five on -ramps on two agents. This is ways that your teams can get started using something agentic.

So you've got vertical agent products. products, these are companies that have taken a team or an industry and they've decided, I'm going to build the agent -first product that makes life easier for those folks.

You've got traditional chatbots, so this is like the front -facing, user -facing product behind a lot of the model providers, so famously ChatGPT and CloudCode, or sorry, just Cloud.

Three, you've got the no -code builders, so things like NaN, and Gumloop, and even less sort of AI agent -first products like Make .com and Zapier. Want to give a shout out to Vibe Coders, Base44. I'm sure the team at Base44 will touch on that in their talk.

And then also just using freelancers, contractors, automation agencies.

On-ramp #1: Vertical (industry-specific) agent products

So to go through each one, starting with vertical agentic products.

Pros are these are extremely, extremely simple for folks to get started with structurally they're very similar to the kinds of vendors you would have been evaluating and using already they're usually typical b2b SAS products with a monthly subscription or have some kind of usage based credit

based system so structurally it feels very similar to what you've been using before the only difference is under the hood the teams are doing a lot of work to figure out what are the best ways that we can use AI and agents to make life easier for our customers one advantage that comes with that is that

that they are gathering best practices from hundreds, thousands, millions of teams that are just like yours. So while they have figured out the core problem to be solved that will definitely solve your problem first,

you might also identify opportunities that they have identified that you might not have thought of, because they're talking to many, many teams that are just like yours.

Cons to think about. The big one is the ability to customize and tailor your agent performance as a quick sort of flash forward to the rest of the trend that you will notice with the rest of the on ramps is there is a relationship

between how customizable your agent is and how much you'll have to sort of roll up your sleeves and learn a new tool or learn a new domain very simple tools have very limited ability to be customized if you want to customize a lot generally you will have to start being more comfortable sort of playing around with with a tool that you you might not know just yet so obviously

these vertical agent products are because they're very simple with that comes sort of limited ability to sort of look under the hood and tailor how the

agent works the second con that I notice is data fragmentation across tools you are adding sort of yet another tool to your stack that's one more place that data can live that is not living in other tools that's already a pain that I I hear a lot from my customers and folks that I speak to about this.

And so you are just adding just yet another tool that kind of adds to that problem.

And with that, I'll call out that I think you're leaving one of the big gains from agents on the table, which is that I think that agents represent an opportunity to not be sort of siloed, walled garden kind of products.

Agents do have the ability to be more interstitial and work across the tools that you're already using and kind of solve that data fragmentation problem we've seen from traditional SAS products.

So in summary I would say that if you're looking for something you know very simple you're looking for very limited sort of mental load these are really really good ideas but if you want to sort of optimize your performance if you want to get all the gains that agents can offer that are that are possible I would encourage you to evaluate other options.

On-ramp #2: Traditional chatbots (ChatGPT/Claude for teams)

The second is is the traditional chatbots, so this would just be using ChatGPT for Teams.

Pros here, very, very cheap. I think ChatGPT is like $20 to $50 a month, depending on your size. And they're very easy to use, especially

if you're already using ChatGPT in your personal life. The form factor is identical. So there's a very, very limited learning curve.

We'll also call out that if you are a little bit more creative and you're scrappy in the way that you think about these products, you can squeeze a lot more out of these tools than you might think.

I think there's actually another MindStone talk about how you can use the ChatGPT personas feature to whip up personas within ChatGPT that have dedicated domain expertise, even domain expertise within your team.

So I've seen teams whip up sort of co -pilots for each of the, let's say, five distinct workflows that their their teams need help with, and use the personas

to act as if you've got a marketing lead, or a data lead, or an engineering lead within those individual chats. So I would say people like to scoff at it, but I think if you're, again, looking

for very limited mental load, you can get a lot more out of these tools than you might think.

The other thing I'll call out is that as these products are being built, there is very good coverage across what I would call very standard use cases. If you are working on something that you can safely assume

that 80 % of the teams in the world are doing, data analysis, document creation, there's really, really good coverage of these tools in these products already.

If you've used ChatGPT or Cloud already, you would have seen, you can ask it for a document where it used to just give you the copy of the document.

Cloud Code will now give you a doc that you can download readily, and it's formatted very well. so there's very good coverage for those like standard workflows cons though the

chatbot form factor does come with limits and similar to sort of the vertical agent products you are leaving a lot of gains from agents on the table things like scheduled tasks there's a really there's a lot of advantages to

having agents work on some kind of scheduled basis I know that Claude co -work has scheduled tasks on paper but I think it requires that your desktop is actually open at all times if you want the scheduled task to run so that's an ideal for a lot of teams also sort of the ability to react to other things

going on in your ecosystem so let's take a very simple example of a customer support ticket comes in a user needs help with your product very common Asian form factor that I see people asking for is I want that ticket read I want it

triaged for who needs to be looped in to address this ticket, either help the customer through it with our product functionality already or address what we need to add to our product to solve this problem and then so identify those next

steps and then execute on them in terms of updating slack channels, creating calendar events, writing tickets, whatever it might be. You would lose that kind of reactivity if you're sort of limited to these sort of model specific chatbots.

second point on the note of model specificity there's a lot of model lock -in if you are using sort of the chatbots from these big model providers and if you have paid attention to like the news over the last two years in

terms of all the models that companies are putting out you would have noticed that at any given moment there might be a model from another model provider that is very very good at the use case you are targeting whether that's creating

code creating videos or pictures or even having access to smaller models that can run more quickly to handle smaller less complex tasks you are losing a lot of models that could be very helpful to you if you are relying on just one model

provider and then the third is you still got that data fragmentation problem I mentioned earlier this is just yet another tool that product that data can can live in that doesn't live anywhere else.

On-ramp #3: No-code builders (Zapier, Make, Gumloop, n8n)

So the third on -ramp is these no -code builders. These are tools like NaN, Make .com, Gumloop, Zapier. So these are tools that have done a very, very good job of abstracting the need to write code

and still letting non -technical folks configure very complex workflows and build agents that can do a huge, huge variety of things. So that comes to pro number one.

one, high, high degree of customizability with zero need to write code, literally zero. If you look at sort of Gumloop or Zapier or NNN, if you are willing to roll up your sleeves and get your hands

dirty, you can kind of make anything. These tools are very, very, very powerful. One reason they're powerful comes to point two,

excellent integration across all the tools that you might imagine. These platforms have done a very, very good job of making sure that they've got good integrations with all of the tools you

might be using from your CRM to your email client to your customer success management tool whatever it is agents are only good as the tools they have access to and these platforms give very very good access to tools cons though

one thing I will call out is just because something doesn't require you to write code does not mean that it does not take a slightly technical mind one One thing I notice is that folks who are builders, folks who are more engineering -oriented,

can look at a blank canvas and a set of tools and a goal. And they are very good at and even encouraged by the requirement to wade through all that uncertainty and build what they need.

But I think if you're less technical and you are more interested in very quickly getting all the upsides that you can, that is a huge learning curve, which doesn't always feel like the best investment of time.

So I would encourage you to think about do you have the mental bandwidth and the time required to learn what is a pretty powerful tool. You'll notice if you, I'll

touch on this in the sort of agency or freelancer point, if you go on Fiverr and you look up AI automation, a lot of the contractors and agencies you find will be using tools like this to build your automation.

So this is not as simple of a of a tool, as I think people think that it could be, just because you don't have to write code, does not mean it doesn't take a bit of a systems -oriented,

first principles -oriented kind of mind to use them. So that gets to point two. There is a significant learning curve.

While you can do essentially anything that you want, it does take really rolling up your sleeves and the ability to deeply understand a new domain and a new tool in order to get all those benefits.

And the third point that kind of comes as a 1A, 1B to that is I think that sometimes teams can underestimate the bandwidth it requires from somebody to make these agents useful. Any of the engineers in the room will know that your first version of a product is very, very, very far from the useful version of your product.

So if you're going to say, we're going to on -ramp via these no -code tools, it is not enough to have a one, two -day sprint where people are playing around with these workflow tools and then you leave it there. 1You do need somebody dedicated, I would say, on a very long -term basis to monitoring the effectiveness of these agents and then iterating from there. So you'd have to consider, do we have that kind of bandwidth in terms of energy and time to have somebody dedicated to that?

Vibe coders, I'll let Base44 folks talk about that, but this is a very, very, very powerful tool or set of tools that I would highly encourage folks to consider.

On-ramp #4: Freelancers, dev shops, and automation agencies

And then the last one is these sort of dev shops and emerging AI automation agencies. These really remind me of sort of when 2016 or so when SEO was really, really popular. You had a bunch of SEO agencies coming out. This AI automation space feels very similar.

so pros they're highly customizable with zero need to code because you can very very cleanly outline all your requirements add them off to a contractor or an agency and leverage all of their deep domain expertise to to make that happen and if you as a company are familiar with and very comfortable with working with third -party vendors whether that's contractors

freelancers or agencies this fits very cleanly into that kind of working model and the third point is you can leverage all of their strong domain expertise to make all of those ideas that you have come to life without needing to faff around with okay how do I actually what do I need to learn about to make this happen you can

sort of offload that mental bandwidth entirely but the cons I would highly encourage you to think about is the cost and the incentive structure these dev shops and agencies are businesses themselves they need to make a margin off of your project so they will inevitably be charging you more than the

product might the project might be worth in terms of like how much how much it actually cost them to build it and with that comes misaligned incentives if you had somebody in your team dedicated to building these agents they are highly encouraged to make it as effective as they can as quickly as they can versus

most of these agencies are charging on a per project level so they are are encouraged to take as many swings at bat as they can and I think you will see this again if you go into Fiverr and you look up AI automation freelancers the

cheapest tier will be will make you a workflow with seven nodes and it seems very cheap but then the next tier is much much more expensive and it's 20 nodes or plus and that's because they know that it usually takes more iterations than customers think to get the kind of automation and performance that they want.

So those are the on -ramps. I'm going to pause for questions at the end. So if you have any on these, hold them for just a second.

Framework 2: Integrating agents without falling into the “prototype wasteland”

Now I want to talk about how to effectively integrate. I've put it all in one slide just in case folks want it at the end, but I'm going to walk through these point by point.

Set expectations: risk tolerance, commitment, and time-to-value

So first point, I would encourage you to honestly gauge your team's risk tolerance and commitment and set expectations amongst yourselves about how much time and bandwidth do we have to make these agents as effective as we

need them to be just like I brought up the internet example at the end I think that thinking about previous tech waves is a very useful way to think about this space I think that it can remove a lot of the hype and set expectations in a

very grounded manner so the one I think about is the cloud when I was interning and between 2016 and 2020 every company on earth was thinking about how can we most effectively get on the cloud and get a clean data lake and the companies

that won were the ones that said this will be more complicated than we think it will be this will be harder than we think it'll be it'll take longer than we think but we're just gonna bite down on our mouthpiece push through it because this is important on a strategic level I would encourage you to have a similar

approach with agents if you're deciding to go down this route it will definitely take longer and be harder than you think to get agents to be as useful as you're expecting them to be and I see a lot of teams say my manager or my board of

directors wants us to use agents so I'm gonna buy some kind of tool or I'm gonna do a two to three day sprint where everyone has access to some no -code builder we'll see what people can build we'll have a day of great prototypes and then nothing happens they don't get used at all and they stay in that prototype

wasteland I think that's because it takes longer than folks think to build effective agents or use agentic tools effectively. The second thing, once you've

Start with a real workflow: automate the most painful internal step first

sort of gotten those expectations set, is I would start with a workflow that a human being is currently doing and address one part of that, ideally the most painful, repetitive, and internal facing part of that workflow.

So let me tease out sort of each of those those elements.

One thing I hear a lot which I find very ineffective is I have a backlog of projects or features or things that I want my team to be doing let me buy or make an agent that does this backlog I think this is a bad idea if using or getting on board with agents

for something you're already doing is like building the plane while you're flying it then trying to do that for someone something that a human being is not doing already is like trying to build something while you're flying it And you're also deciding, is it a plane, is it a helicopter, is it a hot air balloon, is it a rocket ship?

You don't understand the workflow yet, and so you can't effectively optimize it. It sounds simple, but I see a lot of people approaching agents this way. I have not found it to be very effective.

If you have, please talk to me after. I would love to learn more about your story. But I have not found it to be very effective.

The second is I would look at the most repetitive internal facing parts of that workflow and I would set the expectation that the agent is going to be more of a co -pilot than owning the workflow in its entirety.

Usually these workflows are more complicated, involve more domain expertise, institutional knowledge, intuition, micro decisions than people think.

Especially high -level managers that don't do the job on a day -to -day usually think things are easier to automate than they are so I would take a small part of it I would encourage the

human who's doing that job to be the human in the loop and be make it more of a co -pilot approach at least to start with a lot of people hear that and they say well I wanted to you know I wanted to work on this automation project to have this project be automated and get all these time and cost savings and

while that's true and you're not going to get as many of those savings as you might have gotten if it was fully automated you will never get those those gains if you don't go at it step by step engineers I think are very good at breaking large projects up and making it more of a piecemeal approach and so I

would very very very much encourage folks to have that sort of approach as well so I would start with one part of the workflow I would stick a human on it, have them monitor the output, and see if you can make their life easier by automating one part of that workflow.

That way the human who is in the loop and understands this project deeply can look at, okay I've been using this for a week, what is the range of outputs that I'm seeing, what looks good and what looks bad.

Raise the floor: evaluations and guardrails before expanding scope

The second step I would go for is develop evaluations and guardrails that limit the catastrophic errors to essentially zero and raise the floor of your agent performance to the point where you can be comfortable using

Example: lead sourcing and personalized outbound (with a human in the loop)

it in the long term you have a question yes so the example the other question was could could we give an example I'll touch on one very very quickly not one that I've done but something that I've built for folks a very common use case

with Solari is grab a bunch of leads from the internet that could be a good fit for my product or service get all the contact information get their website get their LinkedIn get recent news about them and give me some hyper personalized content and then send that out the most painful part of that

is usually all of the aggregation of those artifacts a lot of sales and GTM leads want the whole workflow everything from putting in a service or an industry to gathering the leads gathering the artifacts writing the content and then sending the emails to be automated then they try and do that the and then the

the output contains maybe some competitors that they don't want included maybe the marketing copy isn't the style that they want maybe it's a little bit verbose maybe it's got the sort of chat GPT style copy that I think

we can all we've all started to to recognize em dashes yeah yeah or that like this this and this and honestly this you know what I mean it's just like structure that is like a human didn't write this so in that case I would start

with the aggregation of content and then using that for hyper personalization I would stick your sales let your sales rep as the human in the loop I would have them look at the type of copy that the agent is producing then I would

build evals around are they mentioning competitors what's the proximity to a reference that we think is really really good if it's too far away from a reference let's have it stuck in a human review sort of section but if it's very close to a reference let's have it sent out that's the kind of workflow I would

I would encourage you to go through then from there so now that you've raised the floor of performance and you've limited to the catastrophic errors to zero I would then start working on the prompts adding references and evaluations to to increase the likelihood of outputs that you consider good.

So now that nothing catastrophic can go wrong, you're now increasing the likelihood of good things happening. And then now that you've gotten all of that set up

for the tiny part of the workflow that you're working on now, I would then expand agent coverage from there. So that's just a rough mental model that I would highly encourage folks to consider.

And then the last thing I'd like to bring up

Make it measurable: define success/failure and run test vs. control

up is very specifically aligned on what does success and failure look like for automation for this team ideally push yourself to have a measurable metric I

think that a lot of these decisions are based on vibes and any time that you're introducing something new to a team vibes can be bad vibes can be a little bit be a little bit of friction so having a metric can really encourage you

or really push you to be honest bonus points if you can give the agentic tool to what part of your team and then not for the other part of your team you have a very built -in test and control and have the experiment run probably longer than you think it will and revisit those metrics and compare them on a regular

basis last thing before I leave you a note for the aspiring builders I think

Framework 3: Advice for aspiring builders in the agent era

this is a very very very unique time in history where the friction to building has never been lower but I think it also does require a little bit of a reframing of how you think about building.

So what I would do is instead of learning like how to code I would think about what are the tools that are available to you and then what is the high -level system architecture that contributes to the product you want to build and what the solution for your particular problem is.

In doing so I would develop a sense of what is your personal product style and taste.

There's a good quote I like to say which is that taste is sort of the athleticism of arts and broadly making things and as the barrier to entry for making things goes down the requirement of things being built with good taste and good style goes up that becomes a more important

consideration and then the last thing is I always always always see non -technical folks posting on LinkedIn that they've got the certification or they've done this course that is very very good for base literacy but I would highly encourage you to start building projects.

Doing the hard thing will always be worth it and there are many many many learnings that you are leaving on the table from not just building things to start with.