Gen AI is moving at a breathtaking pace, making it hard to distill down today’s reality of AI capabilities vs future promises. It’s not surprising, given incumbents are launching new services with fancy product marketing videos selling a future vision and VCs building the hype to fuel the $27B that was deployed into AI startups in 2023.

After spending a good few months diving deep, with an engineering builder mindset, we have come to understand the capability limits of AI technology much better. We are hoping to summarize some of our learnings to serve as a guide for business leaders, CMOs, and CXOs as they think about using Gen AI in their organizations.

Where do we see strong PMF today?

Current Gen AI technology has product-market-fit within a few core use cases like copywriting, rapid prototyping of visual concepts, and co-creation at scale,

1. Gen AI excels at copywriting - It’s safe to say the areas with the strongest product-market-fit for business use cases with Gen AI are centered around text-based formats. Emails, landing page copy, and SEO-focused content generation are the areas with the largest impact. So, it makes sense to see the success of companies like Copy.ai, Jasper in this space. LLMs are very well suited for text generation and can be fine tuned fairly easily with minimal effort.

2. Rapid prototyping – Getting the first design or ad concept realized is a perfect use case for Gen AI. The technology makes it quick and easy to visualize and iterate, enabled by a very low cost to generate each variant (~$0.05).. You can look at the success of tools like Galileo.

3. Co-creation campaigns at scale – Large brands like Coke, Burger King (see the AI-generated burger below), and Heinz have created these immersive co-creation experiences to drive social and community engagement. This is a great use case for Gen AI today.

   

But marketing-approved visual assets are still a challenge

Currently, Gen AI is not suited for marketing activities like non-text based advertising, visual product marketing, or offline advertising. Let’s walk through a real example to better understand the limitations.

The Ridge case study

Imagine Ridge wallet, an SMB seller and a highly active Facebook advertiser, delivering very high quality ad creatives week over week.

Let’s take a look at what happens if we try to recreate their ads via AI tools available today, straight out of the box.

Dall-E - Open AI’s Image GPT model

Context: We gave ChatGPT a website as a grounding reference
Prompt: create a Facebook ad with an offer for 15% off

Issues:

• The product is not accurately represented in the image

• The photo is not realistic enough to deliver personalized value for users

• Text on the image is not readable and gargled

Off-the-shelf Stable Diffusion Turbo

Context: Manually download a product image, replace background, and edit via prompt
Prompt: create a Facebook ad with an offer for 15% off

Issues:

• Background turned out well but the copy is complete nonsense

• Lack of creativity in terms of objects or other visual elements to land a specific concept

Fine-tuning Stable Diffusion

Context: We fine tuned the Stable Diffusion XL model using 20 high quality ads for Ridge wallet.
Prompt: create a Facebook ad with an offer for 15% off

Issues:

• Product is not represented accurately

• Logo position is not aesthetically pleasing

• Logo is not accurate to the brand

Prompt based editors like Adobe Illustrator

Context: Manually edit product imagery using prompt-based Gen AI within tools like Illustrator.
Prompt: create a Facebook ad with an offer for 15% off

Issues:

• Learning curve for these tools is very high

• AI prompts are able to effectively change localized objects but struggle when changing the entire image

What are the limitations we see today?

As you can see, using these tools out of the box to produce digital ad creatives will lead to low performance and the outputs are unlikely to be deployed mainstream by advertisers.

At the core, the following issues remain with the existing technologies:

1. Image or Video models are not brand aware – they don’t respect font color, style, logo, or voice for the brand. Advertisers spend a ton of time and money to make sure the brand is representative of their business, so ads that don’t honor brand guidelines are a big problem for them.

2. Text overlays are not readable – As you can see from our test images above, text is generated by diffusion-based models as graphical elements, which is problem. With just minor hallucinations, the text becomes unreadable (or, at best, is grammatically incorrect).

3. Product images are not represented accurately – Product images can’t be generated accurately, and can’t compete with high-cost and high production-quality photoshoots. Current models unfortunately don’t let you protect or preserve the product images in the generation process.

4. Legal compliance – Many advertisers are beholden to legal guidelines which are not easy to represent in a model. Sure, there is negative prompting but it’s not guaranteed to work. This restriction will prevent ads from launching as is.

You can read similar explorations, and their learnings, from leaders at Stitch Fix and Klaviyo.

What does the future look like?

The capabilities of Gen AI may be limited for visual assets today, but we're excited for what's to come. Here are some of the innovations we are looking to see in the next few years, enabling Gen AI to hit the mainstream for business use cases.

1. Deep learning models to get trained on business data at scale – A ton of business data like ad creatives and print catalogues are still buried within enterprise databases, or stored with incumbents. This will change if a consortium of businesses and agencies get together to build a data sharing and training arrangement. There are a some examples of this happening already, namely Trillion Parameter Consortium and a data sharing consortium of pharmaceutical companies.

2. New model architectures will better preserve generative context – It’s hard to say whether a new architecture and set of models could come together in a unified pipeline to help enable fine-grained controls of asset generation with the ability to preserve context. If this happens, it will enable advertisers to preserve product images and logo assets, while generating the remainder of the asset.

3. More tools to democratize ad creative production - We are seeing AI tools democratize access to marketing analytics and generated content by removing the knowledge and resource barriers for small teams. Tools like Jasper AI and Funnel are good examples of AI's capability to automate workflows and generate marketing insights. The next step is to unlock the capability to generate accurate, branded, and compliant ad creatives.

Overall, very exciting times ahead with Gen AI applications for businesses. If you are interested to chat or have thoughts we would love to connect. DM us on Twitter! I personally reply to all messages.