Build Custom AI Agents That Actually Work with Jeff MacPherson

Full video transcript

Jeff MacPherson joined the session to share his framework for custom AI agents. He opened by noting how much overwhelm exists in the AI space right now — new tools, new technologies, constant noise — and said the goal with any client is to stop chasing shiny objects and instead build custom AI agents around the business that’s already there.

He ran a short exercise: close your eyes, picture jumping up and down, spinning in a circle, and doing jumping jacks, then open your eyes. His point was that nobody in the group did those actions the same number of times or the same way, because the instructions weren’t explicit — and that’s exactly the problem with AI agents. They can’t read your mind. Without explicit, granular instructions, you get hallucination and inconsistent output.

On his background: he’s been self-employed for over 12 years, is 32 turning 33, and grew up playing competitive hockey through the college level. His first business was an e-commerce company importing a product from overseas manufacturers; when he didn’t know how to market it, he started trading product for Instagram shoutouts — what’s now called influencer marketing — 12 years ago. He sold that business and built the first Instagram direct-messaging chatbot, called Direct Heroes, sending about a billion messages and accumulating roughly 500 million followers in aggregate across clients including Tony Robbins, Dean Graziosi, and other large names. Facebook eventually shut the platform’s messaging access down; after 8–10 months of legal back-and-forth, he lost the company overnight while Facebook effectively relaunched the capability through its own partners. That experience shapes his current warning to businesses: build your own foundation rather than depending entirely on a platform that can disappear.

He now runs Cloud37 (and previously Moment37), helping service-based businesses learn, adopt, and integrate custom agents into their processes. Client range spans a $100 million-plus company he helped generate $2 million in 10 days, down to a new AI consultant he helped land a $36,000 contract within 14 days of starting.

MacPherson pushed back on the popular quote “AI won’t replace you, someone using AI will.” He prefers: “AI won’t replace you — it will expose what shouldn’t be there already, including you.” His point is that many businesses delegate tasks before they understand the underlying process, and AI will amplify whatever is already happening — good systems get better, broken systems get more broken, the same way social platforms amplify good or bad marketing. He described a CEO client who shifted mindset, cleaned up the foundation, and onboarded 15 new clients in two weeks without adding headcount.

He argued there aren’t actually that many real AI experts, because the technology hasn’t been around long enough — everyone in the room is already an expert in their own field, and Cloud37 doesn’t even discuss AI on a first call with a new client. Instead they ask what the client does day to day, how they do it, and who they hand it off to next.

He made a case for “operators over visionaries,” citing a quote that the best organizations have the healthiest immune systems. Operators, COOs, and ops managers who are close to the work tend to drive more AI success than purely visionary leaders, because they understand people and process. His advice: become a better operator yourself, or bring one into the organization if you can’t.

On SOPs versus frameworks: most people don’t like writing or reading SOPs, and long SOPs (like long prompts) overwhelm both people and AI. He compared long prompts to old string Christmas lights — one bad bulb takes down the whole strand — and recommends breaking instructions into microprompts you can isolate and fix individually. A framework with clear guidelines, rather than a rigid SOP, gives an agent (or a person) room to reason and solve problems.

He used Henry Ford’s “faster horses” quote to argue that simply adding more tools or services to a broken system is like asking for a faster horse — you need to rebuild the production line, not just bolt AI onto what’s already broken. He referenced “the economist theory” of how factories struggled to get value from electricity until they redesigned the entire production line around it, rather than retrofitting it into old layouts — the same mistake he sees businesses making by bolting AI onto unchanged processes.

He raised what he calls “Black Friday syndrome”: businesses running many siloed AI tools and agents have to update all of them for every promotion, then revert them, over and over, because nothing is unified. His recommendation is to prototype with tools like n8n, Make, or Zapier, but build the actual agent logic and prompts yourself so the business isn’t dependent on any one platform surviving. He described a 24-year-old client generating over $600,000 in monthly profit entirely on an n8n-built workflow with no testing period before releases — a platform failure could end that business overnight.

He then walked through his four-pillar framework for building any agent: inspiration (the input — a transcript, a website, a sales call, whatever data is feeding the process), transform (structuring that raw input, e.g. stripping filler words from a call transcript, so the agent can actually use it), human-in-the-loop (someone approving or rejecting the output until it’s consistently right, at which point it can be automated), and output (the finished result — ideally with a concrete example, since agents work far better when you feed them an example of what “done” looks like). He said he builds these backward, starting from the output, because if you already know what the finished product looks like, you already know the steps you took to produce it manually — which removes most of the guesswork in prompt-building.

Asked how to tell “signal from noise” — which AI ideas are worth pursuing versus distractions — he pointed to mapping your customer journey: every step from how a customer finds you, through onboarding, delivery, and repeat business, is effectively your production line. Each step is a task someone (or something) performs, and walking through it typically exposes where the real bottleneck is — often onboarding or lead volume rather than fulfillment, which is where businesses instinctively assume the problem lives.

He then introduced his five-part structure for writing the actual prompt: task (the goal), logic and background (the instructions), analyze (style and tone — should it sound like you, your business, or a specific client), modify (dos and don’ts, mostly don’ts — his example being AI’s well-known overuse of em dashes), and examples (real finished outputs baked directly into the prompt, which he calls the most important piece for consistent results). He again stressed breaking work into micro-prompts — for content, separate agents for the caption, hashtags, hook, and image prompt rather than one long prompt doing everything — and connecting a working prompt to a knowledge base (not a database) so the agent gets smarter as it accumulates real reference material, particularly useful for customer support.

He closed with a case for “growth by collaboration”: partnering with fellow builders and communities rather than just accumulating more tools, and positioning yourself as a partner rather than a vendor when working with clients, which he said raises both contract value and clarity of expectations. He also cautioned that most of the many “shiny object” tools on the market are really just selling compressed time — a prebuilt prompt behind a UI — and that once you understand the four pillars and the prompt framework, you often don’t need the tool at all.

In the Q&A, he discussed his two companies: Moment37, the original service-based agency side (named after the AlphaGo documentary, referencing the famous “move 37” moment when Google DeepMind’s AI beat a world-class Go player), and Cloud37, the software platform side where clients build and centralize their custom agents, connect third-party tools, and build their knowledge base in one place. He noted the whole presentation deck was built in Gamma — text in, one-shot generation, then light cleanup and export to Google Slides. Attendees raised related points about “work slop” (AI output that creates more rework because teams lack the underlying business intelligence and systems to use AI effectively) and the risks of building a client-facing AI business without a reliable technical delivery partner, which MacPherson said Cloud37 addresses by producing a scoped “blueprint” for a business before any building begins.