Good morning.

This spring, the most hyped AI video product on earth got switched off. Sora ran for barely a year, burned about a million dollars a day, and pulled a billion-dollar Disney deal down with it when it went.

The reason it died is the reason most people are still using AI video wrong.

This is the mid-2026 field guide. It covers which tool wins which job and what each costs, the three unglamorous uses with documented ROI, the honest read on AI ads, the faceless-channel truth, and where to go to get good at this.

Plus how to wire the whole thing into a system your agents run, instead of a pile of subscriptions you click through by hand.

— Sam

Let’s get into it.

The whole issue lives inside one fact, so start there.

Sora, the model half the AI-video courses were built on, died of arithmetic, not scandal. It cost around a million dollars a day to run and made roughly $2.1 million in purchases across its whole life. The most-funded consumer AI video product in history could not make the numbers work, and it was gone inside a year (OpenAI).

And it was not just Sora. Coca-Cola and McDonald's walked back AI ads their own customers mocked, and the platforms started demonetizing mass-produced channels. The hype crested and broke.

I run every new AI development through the same three questions before I touch it:

AI video splits that test cleanly. As the get-rich machine the gurus sold, it fails the first question. That was a bubble, and it popped on schedule. As a set of specific jobs inside your business, it passes all three.

Two things this spring changed the AI video map:

The flagship died of its own economics. No competitor killed Sora and no regulator banned it. Generating video with AI is expensive, and the price you paid did not cover it. If your business sits on one hyped model that loses money on every render, the vendor can vanish, get sued into a corner, or price you out, and you find out on a Tuesday. Build on a single subsidized tool and the business you think you own is really someone else's burn rate on loan.

The platform turned on the slop. In July 2025, YouTube renamed its "repetitious content" policy to "inauthentic content" and made mass-produced, templated video demonetizable channel-wide. Then it enforced. The fake-trailer channels Screen Culture and KH Studio got cut, part of a wave that hit faceless AI channels with tens of millions of subscribers between them (Deadline).

The hype got this wrong in both directions, though. YouTube did not ban AI video, and it did not ban faceless channels. Its own Rene Ritchie called the change a minor update. The policy targets inauthenticity, mass-produced sameness with no human judgment in it, not the tool that made it.

You can still use AI video on YouTube and monetize. You cannot run 300 templated slideshow videos through a script and expect a paycheck.

Stop asking which AI video tool is best. It is the wrong question, like asking for the best vehicle before you know whether you are hauling gravel or racing. Ask what job you are hiring it for.

Here is the mid-2026 field, sorted that way.

One honest note before the prices. This category re-prices itself constantly. The figures below are mid-2026 ballparks, so treat them as roughly-this and check the vendor's own page the week you buy.

A scene, a b-roll shot, an establishing image that moves.

For most operators, that means Veo to get a great shot fast, Runway when you need to control it, and Kling when you need volume on a budget.

A person on camera reading a script.

A "real person" testimonial for a DTC product.

Long video into short clips.

Dubbing and localization.

Nobody uses one tool. A working stack is usually a clip generator, an avatar tool, a repurposer, and a dubbing tool, each hired for its one job. The whole decision in one pass:

The tools change every quarter. The skill of directing them does not, and it is worth more than any single subscription.

Most online entrepreneurs skip this, buy three tools, and generate mediocre clips for a month before quitting.

Here is the shorter path, in the order I would learn it.

Start with the manual, not a prompt thread on X. Each tool responds differently to motion, camera, style, audio, and iteration, so generic tips travel badly. The people who get the best output are the ones reading the docs.

There is a smarter move than memorizing any of it. AI Video School makes the case to stop hand-writing prompts and instead train a custom assistant on the official docs for your model, then let it turn a plain-language idea into a model-optimized prompt. It lowers the technical barrier without lowering the output.

Judge a course by what it optimizes for, not its runtime. Three worth the time:

YouTube still wins for seeing it done, if you pick carefully:

A gallery of disconnected clips teaches nothing but one that exposes the prompt behind each clip teaches a lot. YouMind is the one to bookmark, since it is free, updated daily, and shows the full prompt on each result, so you can find something close to what you want and reverse-engineer how the motion, subject, and style were phrased.

The prompt is not the asset. The workflow that turns prompts into finished, on-brand video on a schedule is the asset, and that is the difference between the operators who own a folder of clever prompts and the ones who own a system.

The proven money in AI video sits in three unglamorous jobs where the savings are structural. They do not depend on going viral, they just cut a cost or open a market.

The numbers here hold up better than anywhere else in the category, and none of the three is the cinematic clip everyone spent a year chasing.

If your business maintains a library of talking-head video (course lessons, onboarding, training, product walkthroughs) the old model was brutal.

You booked a studio, hired a presenter, and re-shot the whole thing every time a detail changed. Avatar tools break that. You update a script, not a shoot.

The play. Any info-product, SaaS, or agency sitting on a video library that goes stale should move its evergreen, script-driven video (onboarding, how-tos, course modules) to an avatar tool, and keep the human-presence work (founder story, sales) human.

The mistake people make is feeding an avatar tool a wall of prose. Written for the eye, it comes out stilted. Avatars read scripts, so write a script. Paste this into Claude with your source doc:

This is the single best-documented use case in AI video, and it comes straight from YouTube, not a vendor deck.

Early testers Jamie Oliver and Mark Rober saw international viewership triple, with 25%-plus of watch time coming from non-primary languages.

YouTube rolled auto-dubbing out to millions of creators through 2025 (YouTube). This is the strongest independent proof in the category, and the audience was already there, sitting behind a language wall.

Studio dubbing runs $1,500 to $4,000 per language. AI dubbing runs under $50.

Trivago. Localized campaigns across 30 markets and cut months of post-production. For a brand with one asset and many markets, dubbing collapses a multi-month, five-figure localization job into days.

You already made the video, so dubbing is the cheapest growth lever on this list. It adds no new content, just unlocks an audience that was always there behind a language wall.

The play. Take your best-performing existing video, the one that already converts in English, and dub it into the two or three languages where your analytics show lurking demand.

Do not guess the languages. Let the data pick them:

Never dub the whole library on day one. Dub your winner, prove the lift, then scale the winners.

Opus Clip is the proof this job is worth paying for. Independent financials (via Sacra) put it around $20 million in revenue in 2025 at a $215 million valuation after a SoftBank Vision Fund round, with 10 million-plus users and 170 million-plus clips generated.

People do not spend that on a toy. They spend it because one webinar turning into thirty captioned vertical clips in ten minutes is a real multiplier on content you already recorded.

This is also where the less glamorous workflow tools earn their place. Gling's writeup on AI video workflow is worth reading because it focuses on the unsexy, high-value parts, like ideation, drafting, trimming filler words, subtitles, and adapting one long piece for several channels.

The play. If you produce anything long (a podcast, a webinar, a YouTube upload) a repurposer is the highest-leverage subscription on this list.

One input, a week of channel-ready output. Run it as a fixed sequence so it happens every time, not when you remember.

For each long-form asset:

The captions and copy are where auto-repurposers go generic. One prompt per clip fixes it:

All three win the same way. They multiply work you already did, instead of conjuring net-new content from nothing. That is where AI video pays, reliably, today.

Every entrepreneur I talk to steers to AI video for paid ads. It is the biggest make-money lever in the category, and the one where you most need the honest version, because the hype is loudest and the mistakes are the most expensive.

Start with the one rock-solid number. Meta's own AI-powered ad tools hit a $10 billion annualized run-rate, growing roughly three times faster than its ads business overall, with new ranking models driving a measurable lift in clicks and conversions (per Meta's newsroom).

For AI creative specifically, the most credible case is first-party.

RevenueCat. The subscription-infrastructure company ran its own head-to-head, an AI digital-twin avatar ad against a human creator.

The avatar hit 87% of the human's conversion rate at $20 per avatar versus $500 per creator, with a 31% lower cost-per-acquisition.

A Brazilian-Portuguese dubbed version came in at $8 CPA, 4.5% CTR, and 2.1x ROAS.

AI creative got about 90% of the result at about 4% of the cost, which is why it belongs in the testing seat, not because it beat the human on quality.

AI wins the test, humans win the close.

Across the comparison data, AI UGC wins on volume, speed, and cost-per-variation, so you can generate dozens of ad angles for the price of one human shoot.

But in trust-dependent categories (finance, health, beauty) real human creators still convert better at the bottom of the funnel.

One first-party $100K test found human UGC beating AI on click-through (2.4% versus 1.9%) while AI won on cost-driven ROAS, and the hybrid beat either one alone.

A quick honesty note, because this is where the internet lies to you most:

The "AI ad got 45% higher conversion, 3.4 ROAS" numbers floating around are almost all anonymous agency blog claims with no named brand, so ignore them.

The credible reads are RevenueCat's own test above and one independent academic study (Taboola with Columbia, Harvard, TUM, and CMU) that found AI ads averaged a 0.76% click-through rate against 0.65% for human ones, a real but modest edge measured across real campaigns.

Modest and real beats spectacular and fabricated. Plan around the modest number.

Here is the playbook as a loop you can run, with the decision rules that keep it from becoming an expensive slop machine:

The variations start with the script. Hand the UGC tool a structured brief, not a vague one, and let it spin the angles:

And track the number that matters. Cheap AI video lies to you if you count generations instead of finished, usable, converting video:

Use AI as the wide top of the funnel and the cheap testing engine. Do not ask it to be the trusted face that closes a $2,000 sale. It is not there yet, and the data says your customers can tell.

The ads conversation is about reaching strangers at volume. The job most operators skip is the opposite one, using AI video to make one-to-one outreach feel personal at a scale a human could never staff.

AI video walks straight into the sales workflow here, and three tools show the shape of it:

The mechanism is different from ads. You record one real message, then let AI swap the name, the company, and the first line, so a hundred prospects each get a version that reads as made for them. It is one authentic message scaled, rather than a synthetic person invented from scratch.

The catch runs through this whole issue. Personalization at scale can tip into the uncanny, and a prospect who realizes the "personal" video was machine-assembled can sour faster than if you had sent plain text.

Use it for the high-volume, low-stakes top of the pipeline (first touch, reminder, recap), and keep a genuinely recorded human video for the moment the deal is real.

Every guru selling an AI video course is selling the same dream, a portfolio of faceless, AI-voiced YouTube channels printing passive ad revenue. Let me be precise.

What's real. It works, and there is one clean, independently-reported earner.

Adavia Davis (solid, reported by Fortune). A 22-year-old running five faceless AI channels, grossing about $700K a year, roughly $40K to $60K a month against about $6.5K in costs, for 85 to 89% margins, with videos costing as little as $60 to $110 each.

Note the scale, though. This is one named, audited operator treating it as a content business, a long way from the "spin up 50 channels" fantasy the courses sell.

What's not real. The idea that it is easy, safe, or repeatable at guru scale. Adavia Davis is the one audited operator in a sea of "$10K/mo faceless channel" claims that trace back to people selling the course, not running the channels.

The category grosses nine figures in aggregate, and most of that accrues to a handful of operators who run it as a real content business rather than a script-and-forget farm.

Two risks decide whether it lasts:

The channels that survive have a distinct voice, real editing, and a point of view. The slideshow farms die.

So if you play here, run it as a media business with AI as an input, not an AI business with media as an output. That distinction is the difference between $700K a year and a terminated channel.

Every tool page shows you the magic clip. None of them show you these three, and all three will bite an operator who plans around the sticker price.

You pay per generation; your business runs on finished video. AI video is non-deterministic, so the same prompt gives a different result every time.

That "$1.50, 10-second clip" becomes $5 by the fifth regeneration, as the lighting shifts, the outfit changes, or a hand comes out wrong.

So the number that matters is cost-per-finished, brand-safe video, and it runs several times higher than the cost-per-generation you were quoted. Budget for the regenerations and the editing time to stitch clips together, or your "cheap" video line lies to you.

Coca-Cola's 2025 AI holiday spot reportedly took around 70,000 generated clips and about 100 people to assemble (Futurism). "AI made it" is a long way from "AI made it cheaply or alone."

The models have no memory. Every clip starts from scratch, so keeping the same character, product, or brand look across shots is the unsolved core problem, and multi-character scenes (two people sharing a close-up) still break across every platform, with identities blurring together.

For a business, consistency is the point, which is exactly where the tools are weakest. If your use case needs the same face or product to look identical across ten shots, test that hard before you commit. It is the thing most likely to fail.

Labeling costs conversions, and this is the one that should reshape your strategy. Peer-reviewed research (Journal of Business Research) found that when consumers believe a message is AI-authored, they rate it less authentic, feel measurable aversion, and show weaker purchase intent, even when the content is identical.

Germany's Nuremberg Institute found that merely labeling content as AI-generated lowers ad attitudes and willingness to buy, and about 70% of consumers (Deloitte) worry AI content is being used to deceive them.

You can watch it cost real money. That Coca-Cola spot drew public mockery, and McDonald's Netherlands withdrew an AI Christmas ad, calling it "an important learning" (NBC News).

Brands pay more than you think for AI video and get punished for it by their own customers. The backlash is about impersonality, real enough that big brands eat the sunk cost to pull the work.

Put the cost trap and the trust penalty together and the thesis is simple. AI video is a strong top-of-funnel and production-support tool, and a risky bottom-of-funnel closer.

Use it where volume and cost matter. Keep a human where trust does.

Most operators leave the real money on the table right here, and it is the difference between a tool and a system.

Buying a HeyGen seat and an Opus Clip subscription and generating clips by hand, one at a time, when you remember to, only gives you a faster version of the work you were already doing.

You bought hands. You are still the brain, still clicking every button.

Getting real leverage out of AI video means wiring it into a pipeline something else runs. Same rule as everywhere else in this business.

The video generator does not decide what to make, from what source, for which channel, or which version to run. Your system decides. The tool executes.

Each of the three wins changes shape once you stop treating it as a subscription and start treating it as a step in a loop:

None of that runs on a single magic agent. It runs on the boring, dependable pattern of a workflow doing the mechanical steps, a model making the small judgment calls inside them, and you reviewing the output that matters. Video generation is one component.

This reframe separates the 70-to-90%-autonomous operator from the one who just bought faster tools.

Ask of every AI video tool on your list what loop it lives inside, and what runs the loop. If the answer is "me, by hand, when I remember," you own hands with no brain attached.

Wire the brain on top, and the same subscriptions you already pay for start compounding instead of running faster.

Same field, but your first move depends on what you run.

Your win is avatar video for the unglamorous library (onboarding, feature announcements, in-app how-tos, release notes) on HeyGen or Synthesia, so you update a script instead of rebooking a shoot. Layer a repurposer on your webinars and demos. Keep the founder's face on the sales and story video.

Synthesia customer Infinite Peripherals is the proof. It produced 4 explainer videos in 48 hours against roughly 30 days the old way, driving 40% more views and 35% more meetings booked at a trade show, with Pyne AI reporting similar on HeyGen (both vendor-reported).

The win is speed, collapsing a month-long explainer project into a same-week one.

Two moves. Internally, AI video is margin. Run creative testing at volume for client ad accounts (generate 30 variations, let the platform pick, re-shoot winners with humans in trust categories). Externally, it is a service. Productize dubbing and repurposing as retainers, because they have documented ROI and clients see the output immediately.

One rule is non-negotiable. Work in an indemnified tool for anything you deliver to a client, and get written likeness consent on every avatar. Your exposure is the client's brand and your own contract.

Two firms mark the range. Superside cut client production time 57 to 85% with AI-assisted creative ops, and a commissioned Forrester study put its ROI at 94%. UK agency Favoured scaled AI-UGC variants from about 10 to 15 a month to 50 to 100 on HeyGen, with one team member going from 20 videos a day to 100 in a morning (vendor-reported). The edge is throughput and margin, and Superside raised prices on the strength of it rather than lowering them.

This is your force multiplier. One of you can now offer video localization, repurposing, and ad-variation packages that used to need a small team. Use the avatar and dubbing tools to sell services you could not staff before, and use a repurposer to keep your own pipeline full while you deliver client work. Stay on the cheaper tiers, price per finished deliverable, and never quote a client the per-generation cost. Quote the finished-video cost, regenerations included.

Aim it at growth, not headcount. Dubbing is your highest-leverage lever, the one with the tripled-international-reach proof, and your archive is full of content that only ever ran in one language. Use repurposing to turn long pieces into short-form across channels. Keep a hard human line on anything editorial or fact-bearing, because the trust penalty hits publishers hardest of all, and your credibility is the product.

Sora's death is the story of AI video graduating from a lottery ticket into a tool.

Tools reward the operator who knows exactly what job they are hiring for, and punish the one still waiting for the magic clip.

Take one video you already made that already works (a top course lesson, a converting ad, a strong long-form upload) and run it through the boring, proven play.

Dub it into a language where you have lurking demand, or chop it into a week of clips.

That is where the money in AI video is, in getting more out of what you already have.

Then wire it into a loop something else runs, so the next video gets multiplied without you touching the tool.

Building those loops (the workflow that pulls the video, runs the tool, and queues the output while you sleep) is the work Cortex exists for. Free issues give you the play. Cortex gives you the system that runs it.

The hype sold you a machine that prints content while you sleep. The reality is better and smaller.

Wired into a system, a set of sharp tools makes the work you already do cheaper and faster, and playable in thirty languages you were locked out of.

Pick the job before the tool, keep a human where trust lives, and let the system do the clicking.

The next batch of issues are coming soon:

Click here for more details on Cortex

Talk soon,
Sam Woods
The Editor

.