The Exact AI Stack ADSWORM Uses to Run 12 Indian Brands (Midjourney, Kling, Higgsfield, Nano Banana Pro, ChatGPT)

Every month I get the same DM from another agency owner: "What AI tools are you actually using? Like really using, not the LinkedIn list."

So here it is. The actual stack that runs ADSWORM in 2026. Twelve clients, 30+ creatives produced per client per month, two full-time team members on creative production, and an honest admission of what each tool does and does not do.

If you are a brand owner trying to evaluate whether to hire an AI-first agency, this is the diligence you should be doing on us. If you are an agency owner trying to build your own stack, take what works for you and ignore what does not.

The Six Tools That Run Our Day

That is roughly $180/month + the Nano Banana lifetime. Less than what most agencies spend on a single freelance video editor for one project.

The Workflow That Turns These Into Ad Creatives

Tools matter less than workflow. Here is how a single ad creative actually gets made for one of our derm clinic clients.

Total time from "let's make a new creative" to "ad is live" is roughly 90 minutes per creative. Multiply by 8-10 creatives a week per client across 12 clients and you have why our retainer covers this volume sustainably.

What Each Tool Is NOT Good For

This is the part nobody writes honestly about because everyone is trying to sell you their AI agency. Each tool has clear limits.

Midjourney

Garbage at Indian faces. The default training data skews white-American-coastal. Every Midjourney prompt for an Indian context needs explicit ethnicity terms, specific city descriptors, and even then the output skews lighter-skinned. We have given up on Midjourney for hero shots that show people's faces clearly. Aspirational silhouettes, environments, products — yes. Recognisable Indian faces — no.

Kling AI

Hands. Always hands. About 1 in 4 outputs has a hand morphing into a fifth finger or a wrist that bends impossibly. We run every Kling output through a "hand check" before approval. Anything with text on screen (phone screens, signs, menus) gets garbled. Use static images for those scenes.

Higgsfield Cinema Studio

Slow. A 7-second clip can take 8-12 minutes to render. Plan your day around it. Pricier per output than Kling, so we reserve it for hero shots, not mass production.

Nano Banana Pro

Sometimes too aggressive on the upscale. It will sharpen details that the eye reads as fake. Use the "natural" preset, not "enhance" for ad creative.

ChatGPT

Indian English. It defaults to American spellings (color, organize, optimize). Every output needs a manual pass. Also: it hallucinates Indian client names, project locations, and "case study" numbers if you let it. We never let it write client-specific case studies without us editing every paragraph.

Tools We Tried and Killed

Stack hygiene matters. Stuff we tried in 2024-25 and stopped using:

• Runway Gen-3. Quality dropped relative to Kling 2.0 by Q3 2025. Cancelled.
• Sora. Could not justify the cost vs Higgsfield for our use cases. Indian creative scenarios produced inconsistent results.
• Pika Labs. Decent for memes, useless for premium clinic content. Killed.
• Stable Diffusion (self-hosted). The 4 hours/week of GPU babysitting was not worth the marginal cost saving.
• Jasper / Copy.ai. ChatGPT does this better and cheaper. Cancelled both.
• Predis.ai. Indian SaaS, decent for batch social posts but the output looks templated. Not premium-brand fit.

Hardware and Other Boring Realities

An AI stack runs on hardware. We use 2 MacBook Pro M3 Max machines for the team. Local rendering in CapCut + Premiere is fast enough that we do not need cloud rendering for most Reels. For longer cinematic sequences we push to a Hetzner cloud GPU instance for ~₹800/render.

Internet is the other under-discussed constraint. AI creative requires consistent 50+ Mbps upload to push outputs to clients fast. We pay for an enterprise Airtel Xstream Fiber plan (₹5,500/month) just for this. The day the connection drops we lose 4 hours of production.

The Real Cost of Running This Stack

Less than the cost of one mid-level video editor's salary. That is what makes the AI-first agency model work in India. The cost structure flipped between 2023 and 2026 — a small team with the right tools now beats a large team with traditional production.

What I Would Tell Someone Starting an AI Agency Today

Pick three tools. Master them. Do not subscribe to fifteen things. The agency owners I see failing at this are running a tool zoo and producing nothing. Midjourney + Kling + ChatGPT is enough to start. Add the others when you have a real reason.

Spend 20% of your time on prompt libraries. The agencies that win at AI are the ones with proprietary prompt libraries built from 6-12 months of testing. We have a Notion doc with 340 tested prompts categorized by use case. That is the actual moat, not the tools.

Hire a creative director, not a prompt engineer. The bottleneck in AI creative is taste, not technical skill. Anyone can learn Midjourney syntax in two weeks. Knowing which output is the right one for a Bandra-based luxury wedding catering brand vs a Janakpuri chaat chain takes years of taste building.

Note: tool prices and version numbers in this article are accurate as of April 2026. AI moves fast. Some of this will be obsolete by year-end. The workflow principles will outlast the specific tools.

— Aman