The AI Ads Stack We Run in 2026

I build ads with AI every day, and most of what's written about AI ad stacks is either a tool list from someone who's never spent money, or a scare piece from someone who's never opened a terminal. This is how the work is actually ordered, and the honest limits at the end. The specific vendors and models we use stay in house, partly because they change every few months, and mostly because the order of operations is the part that matters. Nothing here is theoretical. All of it ships.

The rule the stack is built on

Options first, judgment second. Every stage below exists to generate more options per dollar than a human team could, or to compress grunt work a human shouldn't be doing. Judgment, which options live and which die, stays human at every stage. Get that backwards and you'll ship confident rubbish at scale.

Stage 1: research scrapes before anyone writes

Before a single word gets written for a niche, we pull the most viral organic videos in it and take them apart. Transcripts, hooks, formats, pacing, what the first frame shows. The goal is patterns: what this audience already stops for, in their own words. That research used to take a week of watching and note-taking. Now it runs while I make coffee.

This is the least glamorous stage and the one that moves results most. Most bad ads aren't badly made. They're built on guesses about what the audience cares about, and the scrape replaces guesses with receipts.

Stage 2: volume in, most of it dies

For every concept the model drafts hooks in bulk off the research, and a human kills most of them. That ratio is the point. AI's real gift in scripting isn't quality, it's cheap volume: a pile of genuine options costs minutes, and somewhere in it are a few angles a copywriter staring at a blank page wouldn't have reached by Friday. The taste, deciding which survive, is still entirely the human's job, and the kill rate stays brutal on purpose.

Stage 3: every ad read frame by frame

Every ad we study, ours and competitors', gets pulled apart frame by frame with vision models: what's actually on screen, text placement, product visibility, expression. Those per-frame notes then get synthesised into a verdict pack: what the ad is doing, why it might be working, what to steal and what to skip.

The reason to automate this at all is coverage. A human can study a handful of competitor ads properly. The interesting patterns only show up once you've read all of them, across a whole niche, every month. Analysing everything beats analysing a sample.

Stage 4: an AI agent drives the edit

An agent executes edits on the real timeline: cuts, captions, reframes, versioning. That turned versioning, the donkey work of editing, into a background task. A human editor still reviews every cut before it ships, because the agent is fast and tireless and has no idea when a jump cut feels wrong.

Stage 5: AI avatars, used honestly

We run AI avatar ads. They convert in some accounts and flop in others, so we test them with open eyes like any other format. One rule is non-negotiable: an avatar never pretends to be a customer. An avatar presenting information is an actor. An avatar claiming "this changed my life" is a fake testimonial, and no CPA is worth being the brand that got caught running fake people. The longer version of that argument is in AI UGC vs real UGC and how to run AI ads without the backlash.

The stack at a glance

The honest limits

Here's the section the tool lists leave out, and it matters more than the tools.

• AI can't pick winners. Ask a model to rank ten ads and it will, with total confidence, and the ranking barely correlates with spend outcomes. The auction picks winners. All the stack can do is raise the quality and volume of what enters the auction.
• AI can't feel cringe. It cannot tell you a hook is trying too hard, that a line will make a 34-year-old mum roll her eyes, or that an avatar's smile sits exactly one frame too long. Cringe detection is the most commercially valuable skill in creative right now, and it is entirely human.
• AI can't take blame. When an ad account burns a week of budget, someone has to own it, learn from it, and answer for it. A model doesn't stand behind its output. Whoever signs off owns the result, which is exactly why sign-off can't be delegated to the thing that can't be held responsible.

So the honest summary of the whole stack is one line: the machine does volume, the human does verdicts. Whether AI ads perform at all is a separate question with data behind it, covered in do AI ads work.

If you want to see what this pipeline would produce for your brand, book a free creative audit and we'll run your niche through it.