Building with AI in 2026: A Long Guide That Ties Funnels, Boring Automation and MCP into One Agency

Why now? A window that is open but narrowing

Over the last year the AI business story swings between two extremes: “everyone gets rich overnight” and “the market is saturated, it is too late.” The truth sits in the middle. Demand is real because every business wants repetitive work off its plate; trust is thin because many attempts are abandoned halfway.

The six videos behind this guide show the same tension. One spends six figures a month on funnel traffic, another earns four figures a day from a plain spreadsheet and a polite email. The common thread is not a shiny demo — it is a loop that runs the same way every day.

Funnel truth — $250k traffic and 84% margin

Greg Isenberg's guest Niels is the founder of Perspective, a funnel builder. His claim is blunt: about $250k a month goes into driving traffic to funnels. A funnel here is a single mobile-first flow that guides the user step by step, branching by answer. Perspective builds it with AI — “AI Block” for variants and “AI funnel” for copy.

The 84% margin figure is not standalone proof; read it as internal accounting. What matters is the shape: traffic → funnel step → AI personalization → booking. People drop at every step, so scaling spend without measuring the leak just scales the leak.

Enlarge

Boring work pays — the $1,200 loop

Michele Torti's “so boring I'm embarrassed” story is the most honest: a friend runs one automation that watches a sheet of unpaid invoices and sends a polite nudge when one goes overdue. That's it. Not exciting, but it pays every month because it replaces work a human is already paid to do.

The nine automations in the video share that logic: inbox triage, content repurposing, cold email, lead-gen agent, speed-to-lead, lead reactivation, AI receptionist and reporting. Each is work nobody wants to do but every business already pays someone to handle.

The $1,200 a day claim should be read as a single-case proof, not a universal. A loop proven on one customer type may not transfer to another. Prove it on one niche first, then measure.

Enlarge

MCP — give Claude every tool

Xavier Mitjana's Spanish walkthrough builds the technical spine: MCP (Model Context Protocol). The idea is simple: let a model like Claude use tools on your computer — files, Blender, Adobe — from one chat. Behind “use any tool” there are three layers: Host, Client and Server.

The Host is where you chat — Claude desktop and n8n in this demo. The Client lists which tools exist. The Server is where authority lives — which file can be read, which command can run. The video deploys it on a Hostinger VPS via n8n, through a secure tunnel.

The flow: find a file on the desktop, work on a model in Blender, finish a job on the Adobe side — all from one thread, with human approval at each step. Automation yes, authority stays with you.

Enlarge

n8n MCP Server — Hostinger VPS
- Install n8n, add MCP Server
- Claude Desktop → Settings → MCP → add Hostinger tunnel URL
- Tools: File System (one folder only), Blender (one project), Adobe (one file)
- For each tool: allow = read-only / write-only / confirm then run

The 10-shop myth — one big, nine small

On Make Money Audit, Rob dissects the “10 AI businesses making $50k a month” claim. Guest Christian Periveli opens with “30 to 50 depending on the month.” Rob's audit is blunt: the biggest business carries most of that range and it is 10 years old. So the picture is not “10 equal shops” but “one big business plus nine small bets.”

Rob grades the video D as a “copy this” promise. Fair warning: scaling many tiny bets at once fragments focus. Picking one business and growing it beats building a portfolio of half-baked ones.

Enlarge

Agency from scratch — the $10k skeleton

Johnny's “How To Start A $10k/mo AI Outreach Agency (Full Course)” is the skeleton of this guide — about 17k words, step by step. Promise: first client in 30 days. Method: pick niche, package, list, outreach, deliver.

Niche is not “sell to everyone” but “pick three niches that spend now.” Packaging is not “we do everything” but “one outcome, one price” — e.g., “15 meetings a month + AI follow-up.” Listing mixes AI personalization with manual check: 20–40 qualified leads a day, each with a short context note.

Outreach runs on email, LinkedIn and a short call; follow-up is the job. Delivery is reported via a tool like Nusifi AI. The course has 20+ sections, but money lives in those five steps repeated daily.

Enlarge

• Pick 3 niches today, write one package line for each
• Find 20 leads per niche, personalize the first line with AI, approve by hand
• Do 1 hour of outreach daily — write, send, log
• Send a one-page report after every call

Outreach opener skeleton (adapt, don't copy-paste):
"Hi {name}, {observation} at {company} caught my eye.
 We set up {method} for {niche} to get {outcome}.
 Open to a 15-min look?"

The math is the same simplicity: 4 clients × $2,500 = $10,000. Close one, write the process, then multiply to four. The course length may scare, but the first 30 days only need that loop.

5 ideas for beginners

The “5 Best AI Business Ideas For Beginners In 2026” video gives the cleanest starter list. First idea is not surprising but proven: generate language lessons with AI and upload them to YouTube. Same format, AI voice and visuals, daily uploads — views compound.

The rest are the boring loops we already know: sheet to mail, content loop, text-to-video. Examples show Veo 3; for Europe where access is limited, Kling is suggested. Earnings claims range from $10k to $80k — the range itself says “not in every niche.”

Enlarge

The lesson: tools change, the loop stays. Today Veo 3, tomorrow another model — but “produce one format daily and ship” stays the same. So niche and language choice beat tool choice.

Tying it together — one stack, one panel

Now tie the pieces. Six videos side by side reveal a three-layer stack. Top is traffic — a Perspective-like funnel tested small and measured. Middle is outreach — AI writes the first line, a human closes the call. Bottom is delivery — MCP via n8n, files and boring automations on the same VPS.

The beauty is one log. Where the list came from, which email was opened, which funnel step dropped, which invoice was nudged — all visible in one place. You cannot improve what you do not measure, so measurement comes before automation.

The most common mistake is building all three layers at once. Rule: one niche, one funnel and one boring automation first. Do not add the third until the first two work. As the stack grows, narrow permissions and widen logs.

Enlarge

Model comparison — where to start? Model Setup First client Main risk Funnel (Perspective) 1 day · $47+ 2-4 weeks Ad cost Boring automation 2-4 hours 1-2 weeks Narrow niche Outreach agency 3-5 days 30 days Rejection rate MCP + n8n 1 day VPS 2-3 weeks Permission error Ranges from video claims + live pricing pages; varies by niche.

Sample split — how $10k breaks down 2 clients × $2,500 $5,000 1 client × $3,000 $3,000 Boring pack $2,000 Example split — start with one pack, then multiply. · beauty v2.6 · limeGrad+glow · premium v7 · 4K · click to enlarge

The hard part — saturation and trust

The hard part is not technical but market. SFAI Labs warns that more than half of AI agencies operating in May 2026 will be closed, sold cheap or pivoted by Q4 2027 — a structural call. Trillet puts it similarly: the opportunity is real and the bubble is real, at the same time.

Why? Model and token prices fall, yet hourly rates on Upwork compress. Staying alive on a generic “AI agency” label with no defensible edge gets harder. Add reputation risk: one bad delivery burns the niche fast. So narrow niche and measurement beat scale early.

6 rules for good results

Good results do not need a complex plan, just six rules repeated daily. The shared success pattern across videos is the same: little but consistent.

• One question, one loop. “Which single job will I deliver the same way every day?” — answer in one sentence.
• 8–12 steps, no more. Longer loops err more; keep short, do not start next step before finishing the prior.
• Ask for proof. Does the funnel step truly exist, does the sheet truly fill — ask for logs, not screenshots.
• Measure and write. Write the same three numbers daily: outreach sent, opens, meetings booked.
• Least privilege. Give each MCP tool the smallest permission, log every run.
• One, then four. Prove on one customer type, write the template, then multiply.

Daily 60-minute skeleton:
- 20 min: find 20 leads + AI opener + hand approve
- 20 min: check funnel steps, fix the drop
- 20 min: watch boring loop (any stuck invoice?)

Toolbox and a 30-day start

The toolbox is smaller than you think. For traffic, Perspective (Base $47/month, up via Grow/Expand/Scale), for automation n8n + a sheet, for delivery an MCP server on the same VPS. For language or video, AI voice and Veo 3/Kling — but that is step two.

Cost is the same simplicity: start small. First month $300–500 for funnel tests, $10–20 for VPS, pay-as-you-go for AI credits. Big spend only makes sense once the small loop is measured. Otherwise bigger budget just scales loss.

Starter checklist (copy and tick):
[ ] Picked 3 niches, narrowed to one
[ ] Wrote one-line pack (outcome + price + time)
[ ] Built 1 funnel draft in Perspective
[ ] Tested sheet → mail on one customer
[ ] Allowed one folder in MCP, turned on logs
[ ] Set up table to write 3 numbers daily

In my view the strongest claim is also the one that needs the most careful reading: boring automation really pays. Large-company cases — Microsoft on MCP posture, GE Appliances with 800 agents on the factory floor, ABB with a full FTE-year regained across four departments — make the single-story videos look less anecdotal. Steelmanned: if a job is already one a human is paid to repeat daily, turning it into a loop that takes the same input and produces the same output is the lowest-risk automation. Invoice nudging is the cleanest example.

What is missing is methodology and durability. $1,200 a day, $250k a month in ad spend and an 84% margin come from one source and one period, with no detail on niche, season, refunds or support cost. The SFAI Labs warning — “more than half will be gone in 18 months” — and Trillet's “opportunity real, bubble real” balance the optimism in the same videos. On MCP, the security surface is barely discussed in the video, yet Microsoft and the IETF draft list seven failure modes for the same protocol.

Incentives and verifiability matter as well. Perspective pricing and the AI credit system live on the vendor site and change monthly — check live at decision time. The “first client in 30 days” promise in the outreach course assumes one hour of outreach a day and a narrow niche; widen the niche and the time stretches. The $10k–$80k range for language lessons and Veo 3/Kling depends on format and language, not every channel repeats it. Do not take a number from one source and reuse it in a different context.

My practical take: for a beginner the lowest-risk start is one boring automation — prove the sheet-to-mail loop on one customer type in one niche, log it, then copy to a similar niche. Funnel and MCP come second; outreach adds the human layer on top of both. Do not build all three at once; measure one, then add the next. Prices and niche response shift monthly, so I would re-check every figure from the videos at decision time.