how to build an automated AI video agency (step by step playbook)

in this article i’ll show you step by step how to build an AI video agency with the help of agents to automate the boring parts of it. let’s say you’ve already cracked AI video generation and can make ads that actually look good. now you’re probably wondering how to find clients, and turn all that into an actual business

i’ve condensed a lot of the information i’ve picked up from people who are already doing this, combined it with what i’ve learned through my own experience and what has worked for me in the past. i’ve put everything together into one practical step-by-step playbook so you can see exactly how to approach building this from scratch, without having to piece everything together yourself

quick context on why NOW is the moment:

not even 2 years ago making good ai video meant deciphering comfyui node spaghetti and still getting 6 fingers on every hand. when sora dropped, people were in shock how realistic ai can look, and companies slowly started paying attention. today, just a few months later, every brand is testing ai ads. i talk to a lot of these companies and EVERY single one of them is looking for good ai editors. they have the budget, they don't have the people.

this niche will only grow. it's a blue ocean and the entry ticket is a laptop.

here's the play:

1. use an ai agent to find DTC companies or ones that just raised funding + their emails
2. analyze their business and competitors with claude opus
3. make better ads than the ones they're running
4. use an agent to write a personalized pitch and send the emails
5. schedule the whole thing to run daily on autopilot
6. collect

let's break it down.

1. find companies that just got funded

why freshly funded companies: they just got millions in the bank, investors are breathing down their neck about growth, and the marketing team literally just received a budget to spend. you're not convincing anyone to find money. the money is already there, looking for somewhere to go.

the filter i use:

> raised $2M - $30M in the last 30 days B2B / SaaS / consumer / DTC US / UK / Canada 5 - 100 employees

small enough that your email reaches a decision maker, big enough that they can pay you.

and don't sleep on ecommerce DTC and lead gen brands, funded or not. these guys use ai ads like CRAZY: they burn through creative testing dozens of variants a week, their whole business is fed by paid traffic, and they need new videos constantly. a DTC brand doing $200k/mo on meta is often a better client than a startup that just raised, because they already know exactly what an ad is worth to them.

how to build the agent, concretely:

1. install claude code: one command from the quickstart at code.claude.com/docs (or npm install -g @anthropic-ai/claude-code if you have node)
2. open your Claude app, create a folder on your desktop (you'll see why in step 2) and start working inside it
3. save the instructions below as a CLAUDE.md file in that folder (this is the file claude code reads automatically on every session, it becomes the agent's brain):

> you are my lead sourcing agent. > > 1. search the web for companies that announced a funding round in the > > last 30 days. sources: crunchbase news, dealroom, techcrunch, > > twitter/x funding announcements. > > 2. keep only companies matching: raised $2M-$30M, B2B / SaaS / > > consumer / DTC, headquartered in US / UK / Canada, 5-100 employees. > > 3. for each company find 2-3 contacts: head of marketing, head of > > growth, founder/CEO. find their emails. common patterns: > > first@company.com, first.last@company.com. verify the pattern > > against any publicly listed email on their site. > > 4. if you cannot confirm a piece of data, write "null". never guess. > > a wrong "personalized" fact burns the lead forever. > > 5. save everything to leads.csv: company, domain, round size, date, > > what they sell, contact name, role, email. > > never add the same company twice. check the csv before writing.

type "run the sourcing process" and watch it fill your csv.

if you want it more reliable, plug in an api for the contact part (apollo or hunter both have cheap tiers) and let the agent call it instead of guessing patterns. but the setup above is enough to start today at $0.

manual alternative if you don't want to touch agents yet:

go on facebook and instagram and start interacting with every ad you see: click, watch, engage. the algorithm will think you're the target and start FLOODING you with ads. every brand that shows up is a company actively spending on creative right now. save them to a list. slower, but these leads are pre-qualified by definition, they're already buying what you sell.

and one channel extra: run meta ads for your own ai ads.

flip the whole game, instead of hunting clients, make clients come to you. you make ai ads for a living, so make one selling yourself, put $10-20/day behind it and target founders and marketers. AI ad agencies are already doing this and printing with it:

your ad IS your portfolio. a founder who sees a great ai ad about ai ads doesn't need any more convincing that you can do the job.

spin up a good landing page in Claude Code with portfolio and description of your offer, write a short VSL and record it, connect Calendly, and you're ready to rock

use these resources to make it visually appealing: CLICK

how to find inspiration for your ads? at this point, you should be able to write them yourself, but if you aren't, just do this:

1. go to Facebook Ad Library and search for "AI ads", "DFY ads", and keywords related to this business, take inspiration from ads that have been running for a long time and adapt them to your offer

2. interact with profiles and content in advertisement/ai video niche on Instagram or Facebook, and the algorithm will start showing you similar ads by itself, just like when you're looking for potential DTC clients

3. later on, just run testimonials as your ads

analyze their business and competitors

this is where 99% of cold outreach dies. i can't even count how many times i've been cold dm'd by someone trying to sell me something completely unrelated to my business, so just think about how often those companies get pitched like that...

take each company from your list and run it through claude opus with extended thinking on, and use this prompt:

> research this company like an ad strategist: {company}, {domain} > > 1. their homepage: what they sell, to whom, at what price. quote > > their headline word for word. > > 2. their 3 closest competitors: one sentence positioning for each, > > and the one way each differs from this brand. > > 3. customer reviews (g2, trustpilot, app store, reddit): find a > > phrase customers use themselves, repeatedly. copy it verbatim. > > 4. THE GAP: what is this brand not saying in its marketing that > > points 2-3 suggest it should be. > > 5. one sentence: the ad angle i should build for them. > > do not guess. missing data is "no data", never an assumption.

the human checkpoint (do not skip this):

agents are great at volume and terrible at knowing when they screwed up. so you build one manual gate into the system and nothing gets pitched without passing through it.

put the project folder on your desktop and have your agent maintain this structure inside it:

> ad-agency/ > > leads.csv > > to-review/ <- agent drops every research brief here, one file per company > > approved/ <- you move a brief here = greenlit for an ad + pitch > > rejected/ <- you move it here with a one line note why > > scripts/ <- ad scripts, only for approved companies

add this to your CLAUDE.md:

> folder rules: > > - write every research brief to /to-review as {company}.md > > - never touch /approved or /rejected, those folders are mine > > - only write ad scripts and emails for companies in /approved > > - when i reject a brief, read my note and propose an update to your > > own rules so the same mistake never happens again

do it daily:

> open /to-review

> read each brief

> check the facts aren't hallucinated and the company actually fits as a client

> drag to approved or rejected

every rejection note becomes a new rule in the agent's instructions

"rejected: they sell to enterprises, way too big" turns into a permanent filter

by doing that, your agent will keep improving along the way

make better ads

now you take the angle and produce the actual ad

rules i follow for every 30 second script:

• the hook has to work with no sound and no context,
• the first 3 words decide everything one problem per ad, not three benefits at once
• use the customer language from the research verbatim
• specifics beat adjectives: not "faster", but "from 40 minutes to 4" 70-85 spoken words total

then production: either generate and edit it manually or push the script into crafftai or your own MCP and let it automate the heavy lifting: visuals, animation, voiceover, all generated from the script

the pitch agent

all you need to set up is one AI agent writing short personalized emails built around the ad that already exists

same setup as step 1: it lives in the same claude code project, so just add these rules to your CLAUDE.md under a "pitching" section (the agent already has access to leads.csv and every research file, that's the whole point of keeping it in one folder):

> you write cold emails for my ai video agency. for each company read > > its research file and its ad script, then write a 3 email sequence. > > rules: > > - subject line: lowercase, 2-5 words, no emoji, never "free" or "offer". > > - the first sentence must be impossible to send to any other company. > > "saw you just raised" fails this test. "your homepage says 5 minute > > onboarding but your ads only talk about integrations" passes. > > - mention that i already made them an ad and link it. one link only. > > - every email ends with ONE question answerable in one word. > > - no attachments, no images.

hook the agent up to your sending tool (instantly, or straight gmail through an mcp) and let the sequences go out and follow up on their own. review the first batch yourself before it sends, after that you'll trust the output

put it on autopilot

last step: claude code has two commands built exactly for this, /routine and /loop

/loop makes the agent repeat a task until a condition is met, so you use it to fill the pipeline in one sitting:

> /loop run the sourcing agent from step 1, then research and script > > each new company, until leads.csv has 10 new fully prepared > > companies. skip anything already in the file

/routine turns a prompt into a scheduled action that runs on its own. set up two:

> /routine every morning at 8am: run the sourcing agent from step 1, > > add new qualified companies to leads.csv, flag the 5 most > > interesting ones for me

> /routine every monday: check who received email 1 more than 5 days ago with no reply, queue their follow up

from this point the system feeds itself. your job shrinks to 60 minutes a day: read the report, approve the batch, take the calls

to speed things up, **just copy the link to this article, drop it into Claude Code or Hermes, and tell it to build it for you

===**

**if you want to learn AI content creation from zero and get access to a community full of people building exactly this, everything is inside here: CLICK

full update coming within 2 days**