I use Claude, OpenAI’s tools, Google’s AI products. Every day, all day.
After two years of living inside these systems, I have a pretty solid handle on how to spin up a workflow, automate a process, even write the bits of code that occasionally need writing.
That is not the position most business people are in.
Most of them are watching a new AI tool launch every Tuesday and trying to answer real questions: what should I use for sales? For customer support? For the bookkeeping I do on Wednesdays? How am I supposed to fit all of this into the way I already work?
The chaos is real.
The promise is real too. AI genuinely can change how a small business operates. But the gap between the promise and the actual work of making it land in your day is wider than anyone in tech wants to admit.
The result is a quiet, growing anxiety: everyone else seems to be “doing AI”, and you cannot tell whether you are behind, on track, or being sold a story.
Closing that gap is what this newsletter exists for.
How do real business people, not engineers, not researchers, actually adopt new AI products without dread?
So for this issue I decided to play the role honestly: behave like an average business owner who has not gone deep into the technical weeds, pick a problem every B2B founder eventually faces, and see what “the AI” could actually do about it.
In a single afternoon I built something I would happily call a personal SaaS, the B2B Autopilot Distribution Engine.
I’ve used Hyperagent for this.
What every business actually needs is an AI environment shaped around the way it already operates, not the other way around. Hyperagent, from what I have seen, is built exactly for this. And it delivers.
The Hyperagent team reached out and offered me the chance to put the platform through its paces. I am grateful for the offer, and the return I want to give them (and you) is a careful, honest walkthrough of what an average operator can build in a few hours. Step by step. No hype. No anxiety.
What I Built
Four agents working in coordination, orchestrated through Airtable as a shared state bus.
Onboarding captures a founder’s Ideal Customer Profile and outreach strategy from a single URL. Five conversational turns. Zero typing beyond the URL.
Hunter runs hourly, autonomously. It searches the public web for prospects emitting buying signals: recent funding rounds, new leadership hires, product launches, public complaints about competitors. Enriches each lead with a verified business email.
Copywriter drafts a personalized message per lead. Four opener styles, varied across each batch. Twelve-point self-check before any message is written, banning marketing-speak, mail-merge templates, and em dashes.
Postman sends through Gmail at timezone-optimized windows (Tuesday through Thursday, 7:30 to 8:30 AM local), monitors replies, and updates the pipeline state.
I tested the system on two real businesses. My own Next Big App, specifically the Tap Grow sales suite, targeting VPs of Sales at SMBs scaling fast. And Haven Day, an AI receptionist product for real estate brokers and small business owners.
Two completely different audiences. Same architecture. Same agents. Different ICP definitions plugged in.
Within hours: 15 qualified leads. 80% had verified business emails. Every lead came with a named decision-maker, the specific signal that flagged them, the source URL, a priority score, and a drafted message ready to send.
How It Was Built
The build process is what made me write this issue.
I opened a fresh thread in Hyperagent and described the system I wanted at a high level. Four roles, one shared state, fully autonomous.
Within minutes, the platform had proposed an Airtable base with two linked tables, walked through the four-agent decomposition, and suggested status fields as the baton between stages.
Here is the moment that changed my mental model. The platform did not just execute. It pushed back.
When I described the Hunter’s job, it proposed ten enhancement ideas I had not asked for. Pattern-interrupt opener styles. Send-time optimization per prospect’s inferred timezone. Dead-lead resurrection, where Cold and Skipped prospects get re-scanned for fresh trigger events. Compound deduplication across three identity dimensions, so I never email the same person twice even across renamed companies and reformatted profile URLs.
Three of those ideas I had not considered. All three made it into the production system the same hour.
Hyperagent is not a code generator. It is a design partner.
The difference is enormous. A code generator does what you ask. A design partner reads your spec, finds the holes, and proposes the thing you would have asked for if you had thought of it.
By the end of the session, I had four agents drafted and saved, two linked Airtable tables, scheduled invocations ready to activate, and a clear go-live sequence.
I had written zero lines of application code. The longest thing I typed was a description of the kind of personalization I wanted in the outreach.
What the System Actually Produced
Here is one drafted message, verbatim from the live Airtable. The recipient is a newly-appointed CRO at a ServiceNow AI activation partner backed by growth-stage private equity.
Name redacted; everything else is unedited.
Subject: scaling without the headcount
Most ServiceNow partners hit a wall right about now. New CRO, big growth mandate, but the outbound team can’t hire fast enough to match.
We built Tap Grow to fix exactly that. It runs 5,000+ sales calls an hour at about 1/40th the cost of a call center.
Worth a quick look, or bad timing?
Selim
Look at what is happening in that message. The opener names a category-wide pain (“Most ServiceNow partners hit a wall right about now”) before naming anything specific to the prospect’s company.
Establishes domain credibility without sounding like mail-merge. The bridge maps the pain directly to the product with a concrete stat. The close is a single yes/no question, which paradoxically increases reply rates because the ask is low-stakes.
That message was drafted automatically. So were 14 others in the same batch. Each tied to its lead’s specific signal. Each with a varied opener style across the four available (Observation, Provocation, Insider Intel, Signal-First Hook). Each one passed through a twelve-point deliverability self-check before being written to the database.
To my knowledge, no system on the market today writes outreach with this kind of grounded personalization at this volume. Most “AI personalization” tools fill in {first_name} blanks. This one reads a press release, identifies the operational angle, and writes a sixty-word note about it.
That is a categorical difference.
The Real Product is the Loop
The agents are impressive on their own. But they are not the actual product. The loop is the product.
A few hours into the build, I read the first batch of drafts with my own eyes, the way a real recipient would. I found two problems the agent had missed.
The English was too polished. Words like “infrastructure orchestration” and “at a fraction of the cost” read like marketing copy, not the quick founder note I had asked for. Worse, every message opened with the same template structure (“Name, observation”), the dead giveaway of mail-merge that any sales leader scanning their inbox for five seconds would recognize and delete.
In any other AI tool I have used, this would have meant manually rewriting each of the 15 messages.
Instead, I described both problems back to the Copywriter agent in one paragraph. The agent extracted its own current system prompt to a workspace file, added two new non-negotiable writing rules (sixth-grade reading level with specific word swaps; ban the templated opener pattern), pushed the revised prompt back, and redrafted all 15 messages with the new rules in a single parallel burst.
From “this feels spammy” to a permanent rule change plus 15 rewritten messages: under a minute.
That is the loop. That is what makes Hyperagent feel categorically different from “AI assistant” tools.
Most AI tools are smart at one task. This one is fast at the entire feedback cycle. Spot a problem in production output. Describe the fix in plain English. The system rewrites its own permanent operating rules and regenerates the affected batch.
Closer to tuning a system than doing manual work.
The Pattern Adapts to Anything
I built this for B2B sales. But step back from the use case for a moment.
The B2B Autopilot Distribution Engine is not a sales tool. It is a URL-triggered relationship discovery and outreach engine. Drop a URL in. The system understands what you do and who you need to reach. It finds real-time public signals of those people. It delivers personalized outreach at scale, multi-channel, lights-out.
The same four-agent architecture works for anything where (a) you can describe your business via a URL, (b) there is a category of person you want to reach, and (c) those people emit detectable public signals when they become relevant to you.
Which means the same system can do:
Recruiting: drop your careers URL, find talent showing job-search signals (recently announced role transitions, layoffs at competitors).
Partnerships: drop your product URL, find companies launching complementary products or publicly seeking partners.
Investor outreach: drop your company URL, find investors funding similar companies in the last 90 days.
PR: drop your URL, find journalists covering your beat with articles published this month.
Customer advisory: drop your URL, find existing customers or community members showing engagement signals.
Same agents. Same Airtable backbone. Same iteration loop. Different ICP definition. The system adapts because the agents are defined in plain English.
The Founding 500
While I was finishing this review, the Hyperagent team announced something that fits the thesis of this issue almost exactly. The Founding 500.
They are putting $10 million in inference credits behind 500 agent-first companies. Founders building from scratch on agent-native infrastructure. Operators rebuilding existing companies the same way I am rebuilding mine.
The first 500 qualifying candidates get $20,000 in inference credits for $200.
The thesis they published alongside the offer:
Most companies are still treating AI as a feature, a chatbot here, a summary there. A smaller group is going deeper: rebuilding workflows from the ground up and putting agents at the center of how the business actually operates. The companies built agent-first will have structural advantages their competitors cannot retrofit later.
That is the right frame. If you finished this issue thinking “we should be building this way,” the offer is exactly that.
Apply to The Founding 500 → Submit by May 31st.
Verdict & Sign-Off
I spent thirty dollars. I built in an afternoon. The system is now running on its own while I write this newsletter.
A few years ago this would have cost hundreds of thousands of dollars and a six-month build. Most founders simply could not get it. The SaaS vendors offering something similar today charge tens of thousands a month for systems that are, frankly, less sophisticated than what I ran for $30.
This is what personal SaaS, built right, actually unlocks. Not just one custom tool. A reusable pattern.
![1.mov [video-to-gif output image]](https://substackcdn.com/image/fetch/$s_!992i!,w_1456,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F47a03bfa-fe39-4ebe-ad84-4df0d8b6f1ac_800x500.gif "1.mov [video-to-gif output image]")
The long version walks through the build session step by step, with real screenshots from inside Hyperagent, the actual ICP records pulled from my live Airtable, three drafted messages with reasoning panels explaining why each line lands, a head-to-head comparison against the existing sales-tech tools, and a complete friction file so you know exactly which rough edges you would encounter.
If you want to build your own version (for sales, recruiting, partnerships, or anything that fits the URL-triggered pattern), Hyperagent is giving away 1,000 credits to the first 1,000 Next Big App readers. Create an account to claim yours here.
Airtable as a shared state bus, and a feedback loop that overrides the system prompts themselves. This is exactly the agent architecture that works on an operational level. Decision-making takes place within the agent, data storage remains in Airtable, and orchestration is handled externally.
Great! Now build some for me! LOL!