The AI tools we use to run a marketing studio, and what changed in October 2026
AI does the repetitive work in our studio. People do the thinking. Here is every tool we use, what we use it for, the five things we never let AI do, and what changed this month.
In short
- We use AI in five jobs: research, writing and checking, images, video and motion, and building websites, automations and dashboards.
- Claude Code is our workbench: it builds, connects to ad accounts read-only through MCP connectors, and runs the checks that stop bad work before it reaches anyone.
- Five things AI never does here: redraw a logo, repaint a product, invent a fact, send or spend without a person's approval, or generate a real person's face.
- New in October: Meta made video from static images generally available in Advantage+ creative, and Razorpay announced ChatGPT ads support for Indian brands. OpenAI's own Ads Manager page now lists India for self-serve.
- Pick tools by the job and the evidence, and test every new one on a small piece of real work before you trust it.
Written from the work RAAY Creative's AI Lab does every week. Launch details come from each company's own announcement, linked in the sources.
The five jobs we use AI for
We do not pick a tool and look for work for it. We start from the job, then use whichever tool does it best, and check the result before anyone sees it.
- Job 1 · ResearchFind what is trueCompetitor ads, platform changes, what people search and ask. Every finding keeps its source.
- Job 2 · Write and checkSay it, then prove itBriefs, ad copy, articles and reports, each checked against twelve gates before it leaves the studio.
- Job 3 · ImagesMake the sceneRooms, light and props for product ads, with the real product placed in unchanged.
- Job 4 · Video and motionMake it moveAI video shots, voice, and motion graphics written as code so every frame is exact.
- Job 5 · BuildMake it workWebsites, automations and dashboards that pull live numbers from ad accounts.
- Every jobA person approvesNothing is sent, published or paid for without a person saying yes.
The tools, and what each one does for us
| Tool | Made by | What we use it for |
|---|---|---|
| Claude Code | Anthropic | Our workbench. It builds websites (including this one), automations, dashboards and video pipelines, and runs our own systems |
| Claude | Anthropic | Long writing, research synthesis, briefs, and checking work against our rules |
| Nano Banana Pro, in the Gemini API | Product scenes and ad concepts. The real product photo goes in afterwards, unchanged | |
| Gemini | Reading reference videos frame by frame, so we can break down why a reel works | |
| Google Flow | Animated shots, now with Gemini Omni video models | |
| Higgsfield | Higgsfield | AI video shots from many video models in one place |
| ElevenLabs | ElevenLabs | Voice for films and explainers, when a real voice is not recorded |
| HyperFrames and GSAP | HeyGen and GreenSock | Motion graphics written as HTML and code, rendered to video frame by frame |
| MCP connectors | Meta, Google, Windsor.ai and others | Letting Claude read ad accounts, Search Console and analytics, read-only first |
| Apify | Apify | Collecting public data for research: reels, ads and comments |
| Make | Make | Automations between sheets, WhatsApp, LinkedIn and Instagram |
| Figma | Figma | Design files, which Claude can read and edit through Figma's connector |
Why Claude Code sits in the middle
Claude Code runs on our own machines, so it can read files, run scripts and use connectors. That makes it a workbench rather than a chat window. We use it to build a site, record a website frame by frame, cut a video from a brief, or audit an ad account through a connector, and every one of those jobs runs through the same checks.
Two of those checks are now products. Quality Checker runs twelve gates on any draft before a person sees it: every fact must come from your material, no placeholder text, no banned words, the exact format, working links, nothing added or removed from the brief. Fix It Once writes down every mistake with its cause, and turns a cause into a rule when it comes back a third time. Both are on our products page.
Five things AI never does here
- Redraw a logo. A brand's logo is its identity. We use the official file every time and check it before anything renders.
- Repaint a product. The pack, its artwork and every printed word are real pixels. AI makes the scene around them; the product photo goes in unchanged. In one 100-ad batch we checked, 49 of the 70 defects were misprinted pack text from AI images, which is why we stopped letting AI touch the pack at all.
- Invent a fact. Every number in our work traces to a source we can name. A figure the AI recalled, rather than read, is removed.
- Send, publish or spend without approval. AI drafts the email, the post and the budget change. A person approves each one.
- Generate a real person's face. Real people appear only in real photos and footage they agreed to.
What changed in October 2026
Two announcements from the first week of October, one change now confirmed on OpenAI's own pages, and two late-summer launches we already use.
What it means: one good static concept can now become a video without a shoot. Test it as a new format of the concept, not as a new concept, and keep checking that the product and its text stay exactly right in every frame.
What it means: Indian brands can now buy ChatGPT ads directly. There are no keywords: write for the task the person is doing. Our ChatGPT ads guide covers the setup and a 30-day test.
What it means: a payments company is now offering to run AI ads for the merchants it already serves. Ask any partner how it measures results against your own orders before you hand it a budget.
Also recent, and already in our work
- Gemini Omni 1.1 Flash (Google, August 2026). Google's video model in Flow now extends a scene in 10-second steps up to 40 seconds, takes a first and last frame, and exports at 1080p or 4K, with cheaper 360p drafts first. Its videos carry Google's SynthID watermark and Content Credentials.
- Meta ads AI connectors (Meta, open beta). Meta's ads MCP server lets an AI assistant such as Claude read and manage ads, and since 16 July 2026 portfolio admins can set rules that limit what an AI agent may change. We connect read-only, and keep spending actions blocked.
How to choose your own tools
| Ask | Why it matters |
|---|---|
| What job does it do better than a person, or faster? | A tool without a job becomes a subscription you forget |
| Can it show where its output came from? | Untraceable numbers are a risk, not a feature |
| What can it change, and can you block that? | Read-only first; spending actions always behind approval |
| What happens to your data? | Read the terms before you connect a client account |
| Does it pass a test on real work? | Demos are chosen to look good; your work is not |
A weekly routine that keeps it under control: on Monday, read the week's numbers through read-only connectors; mid-week, make the creative and the content, each checked before review; on Friday, log anything that went wrong and why, so the same mistake does not come back.
If you would rather have us run it, tell us what you want to grow.
Questions people ask
Which AI tools are worth using for marketing in 2026?
Pick by job, not by hype. We use Claude and Claude Code for writing, checking and building; Gemini's image models for product scenes; Google Flow and Higgsfield for video shots; ElevenLabs for voice; HyperFrames and GSAP for motion graphics in code; and MCP connectors to read Meta and Google ad accounts.
Can AI make product photos for ads?
It can make the scene. We generate the room, the light and the props, then place the real product photo in unchanged and check every word on the pack at full size. In one 100-ad batch we checked, 49 of the 70 defects were misprinted pack text.
Is it safe to let an AI tool change ad campaigns?
Only with guard rails. Connect read-only first, prove that a blocked action is refused, and keep anything that spends money behind a person's approval. Meta's ads MCP server lets admins block agent actions per ad account.
What changed for marketers in October 2026?
On 6 October Meta made video from static images generally available in Advantage+ creative, and on 5 October Razorpay announced it will help Indian brands run ChatGPT ads. OpenAI's own Ads Manager page lists India as available for self-serve, which opened at the end of August.
How do you test a new AI tool before using it on client work?
Run it on one small, real piece of work with a known right answer, compare its output line by line, and only then widen its use. A tool that cannot show where its numbers came from does not touch client work.
Sources
- Meta: Advertising Week New York 2026, new AI capabilities to guide campaigns and reach customers (about.fb.com, October 2026)
- Meta for Business: Introducing Meta Ads AI Connectors; Meta for Developers: Ads MCP Server overview
- OpenAI Help Center: Ads Manager Availability (India listed as available, checked 14 October 2026); OpenAI: A milestone in expanding access to AI
- Inc42, 5 October 2026: Razorpay partners with OpenAI to help Indian brands advertise on ChatGPT
- Google: Introducing Gemini Omni (May 2026); Gemini Omni 1.1 Flash lets you build with more control (August 2026); Nano Banana 2 (February 2026)
- Anthropic: Claude Code documentation; Model Context Protocol
- RAAY Creative: 100-ad pack text check for PLB Snacks (August 2026)
Facts checked on 14 October 2026. Platforms change often: check the date on any setting before you rely on it.
Keep it
Download this guide
The tool map, the five rules, the October launches and the weekly routine, on six printable pages.
Download the PDFRather have us run it?
Start with a free sample ad
We make and run Meta ads, content and websites for food and consumer brands. For PLB Snacks: 3.16x return on Meta ads, April to October 2025.
Written by
Anisha Dogra
Anisha runs performance marketing and AI systems for brands at RAAY Creative, the AI Lab and marketing studio she co-founded in New Delhi. She writes these guides from the work the studio does every week, and checks every fact against its source before it goes up.
Talk to us about this
