Nate B Jones argues that AI image generation should be treated as enterprise infrastructure, not a design tool
Solo presentation by Nate B Jones of AI News & Strategy Daily.
Summary
Opening with the observation that Nano Banana Pro reached one billion images generated in 53 days, the presenter argues that the mainstream conversation about this milestone is focused on the wrong things — artistic quality, viral trends, and creative productivity. His central argument is that capable visual AI represents the removal of a long-standing structural constraint on enterprise automation: the inability of AI systems to reliably interpret or generate visual information. He contends that this constraint has quietly limited AI adoption to language-centric functions for years, and that its removal creates a compounding flywheel effect across bottleneck removal, data generation, trust calibration, and workflow integration. He closes by distinguishing between organizations that deploy visual AI as a point solution in the design department — capturing roughly 30% efficiency gains — and those that treat it as infrastructure embedded throughout their systems, which he argues can unlock order-of-magnitude improvements.
Key Takeaways
FULL TRANSCRIPT
The mainstream framing of visual AI is wrong
Nate B Jones: Nano Banana Pro hit a billion images generated in just 53 days. But that's not the real story. The conversation about AI image generation has been captured by the wrong people. If you scroll through any technology publication's coverage of tools like Nano Banana, you're going to find articles about viral trends, comparisons on artistic quality, how to craft a 3D figurine of your pet. They're treating this like a creative tool — something for designers and marketers and hobbyists. And that framing is causing enterprise leaders to systematically underestimate what is actually going on here.
The real story of Nano Banana is not about image quality. It's about what happens when artificial intelligence gains the ability to both interpret and generate visual information. When that capability becomes reliable, fast, and programmable, something fundamental shifts in how organizations can deploy AI across their operations. The constraint that has quietly limited AI adoption for years — the fact that automated systems cannot see and cannot show — that's beginning to dissolve. And when that constraint goes away, the implications extend way beyond your design department.
So this is a story about those implications. Not about which image generator produces the most photorealistic output, but about why the emergence of capable visual AI represents a force multiplier for enterprise AI adoption more broadly. If you're a leader trying to understand where AI creates genuine leverage in your organization, you need to understand this dynamic. The companies that recognize visual AI as infrastructure rather than just a design tool are going to pull ahead of those that don't — and not only in the creative department.
The invisible visual constraint that has limited AI adoption
First, let's look at the invisible constraint that's been holding us back. For the past several years, enterprises have been deploying AI in increasingly sophisticated ways. Language models are drafting communications, summarizing documents, analyzing data, generating code. Intelligent systems can now route customer queries, flag compliance issues, and surface relevant information from enormous code repositories. The capabilities have expanded remarkably, but there has been an invisible fence around what these systems can do, and most organizations have learned to work around it so instinctively that we've stopped noticing it's there. And that fence is visual.
AI systems have been fundamentally blind when it comes to images — unable to reliably interpret visual information, and unable to create it. Consider what this has meant in practice. A customer sends a support ticket with a screenshot attached. The AI system can read the text of the ticket, but a human has to look at the screenshot to understand what the customer is actually experiencing. A market research team uses AI to analyze competitive positioning, but someone has to manually review competitor websites, packaging, and advertising because the AI can't reliably interpret visual assets.
I do emphasize "reliably" because it's not lost on me that we have been trying to solve this problem for a while. Midjourney has been out there building models. ChatGPT has had an image generation model for well over a year now. We have made attempts to solve this problem. But there is a difference between making attempts to solve a core constraint in AI and successfully doing so. Nano Banana is a significant moment because it marks the difference between attempting to solve the visual constraint problem in AI and more or less fully solving it for business purposes.
To get back to our story — if you want to look through your business, don't look at the creative side of things to find those bottlenecks. I just outlined a market research team gap. I outlined gaps in customer support tickets. I could even go into documentation, where if you want to keep your product guides current, every diagram and every annotated screenshot has to be updated. That also hasn't been possible. None of the things I'm describing are edge cases, and none of them are in the design department.
Visual bottlenecks are embedded throughout enterprise operations, and they've been so persistent that organizations have simply designed around them. We staff roles whose primary function is to bridge the gap between what AI systems can process and what requires human visual interpretation. We build workflows that route visual tasks to human queues. We accept that certain categories of work will require human involvement. The business consequence of this invisible constraint is that AI adoption has been systematically limited to text-centric processes, because the earlier image models were not good enough for production business use cases.
The functions that have seen the most dramatic AI-driven productivity gains are therefore those that happen to operate primarily in language — legal document review, customer service correspondence, software development, research synthesis. Functions that are more visual, like marketing, creative, technical training, and customer experience design, have seen AI applied around the edges but not at the core. Automation chains have kept breaking at the visual links in those workflows. That loop is now closing.
How closed-loop visual workflows change automation
This is not a metaphor. I'm describing a specific and consequential shift in what automated systems can accomplish. It forms a closed-loop workflow. Previously, any workflow that required visual understanding or visual creation had to route through a human — at least for a check, and often for the actual work. The human was the bridge between the AI's text-based capabilities and the visual dimension of the task. Now that bridge is no longer required in many, many situations.
Let's get an example going. Say a telecom company's AI system receives a customer complaint about connectivity issues. The customer has attached a photo of their router. In the old model, a human agent has to look at the photo, interpret what they see — which lights are illuminated, whether the cables are connected — and then either resolve the issue or escalate. In the new model, the AI system can interpret that image directly, immediately, and correctly every time. It can identify the router status lights, determine the error condition by doing a lookup, and provide resolution steps live to the customer — or escalate with a visual annotation correctly placed on the image to highlight the relevant detail. None of what I'm describing is a stretch from current capabilities. The human is no longer the interpretive bridge.
Let's try another one. Say a compliance team processes documentation submitted by vendors. Those documents include contracts with tables, forms with signatures, and ID documents with photos. In the old model, AI could extract the text, but humans had to verify all of the visual elements. Does the signature match across documents? Are the tables internally consistent? Does the photo on the ID match the individual on the records? In the new model, they don't have to do that anymore. The AI system can interpret those visual elements directly, flag inconsistencies, and generate compliance reports that include text comparison as well as visual evidence with annotations. All humans have to do is review exceptions.
These examples share a common structure. A workflow that previously broke at visual touch points can now run continuously. The human role shifts from performing visual interpretation and creation to reviewing outputs and handling exceptions. Total human touches on the work decrease. The automation ceiling rises. And the quality of what the human needs to pay attention to also rises — so human engagement becomes more interesting. You're looking at the genuinely weird edge cases. You're directing the flow. Things that simply could not be automated before because they required seeing or showing now can be.
The flywheel effect: how visual AI accelerates broader AI adoption
What's really interesting is that when visual constraints dissolve, the effects compound through a flywheel effect that accelerates overall AI adoption way beyond just creative functions. If we understand this flywheel, we're going to understand how visual AI capabilities matter strategically, not just operationally.
The first part of the flywheel is bottleneck removal. Organizations that could not automate visual-dependent workflows now can. This directly expands the surface area of what is automatable in the business. Customer onboarding processes can now include visual identity verification. Quality control can now include visual inspection of outputs. Training programs that need customized visual materials. Competitive intelligence gathering that involves analyzing visual assets. These categories were all fenced off from serious automation efforts, and there are so many more like them across the business. That fence is now down, and the immediate effect is that more organizational processes are available for AI-driven efficiency gains.
That leads to the second stage of the flywheel: data generation at scale. Every generated image, every interpreted image, every visual interaction produces data that can be used to improve subsequent performance. When a system generates a product visualization and a human approves it, that approval can teach a good enterprise vision system what "good" looks like. I'm not saying you have to use Nano Banana Pro without any tuning or adjustment. It is already a great image model as it stands, and many small businesses are just going to use it as is. But you can pass it modified system prompts, you can adjust many things about Nano Banana in the API, and you can do so based on continuous feedback from your system.
For example, if you discover that Nano Banana Pro is consistently producing excellent results except on a certain model of router, you can write a specific business rule in the Nano Banana API call for your router image interpretation widget that you're building into your business workflow. You can say: when you're interpreting a router image, please be aware that these two models are easily confused and this is the thing to distinguish them. You don't have to build your own image model. You can use known adjustment techniques and known workflow definition techniques that we've developed working with text-based AI agents to deliver the same kind of efficiency gain with image-based agents.
We're also going to see improvements in Nano Banana as a whole, driven by the very rapid usage it's getting across the population. Earlier I said a billion images have been generated in 53 days. That is a massive training signal that Google can use to improve Nano Banana further in subsequent generations.
So we have the first two stages: removing bottlenecks and generating data. Think of Nano Banana as both a tool that an LLM can call — Claude Code can call that tool, Codex can call that tool, many agents can call Nano Banana — and yet it is also conceptually something of an agent itself. This is a multi-agent system, because when you write to Nano Banana, you can give it instructions as if it is reasoning and thinking, not just an image generator. There are multiple levels here. The more we understand the capabilities of the system, the more we're going to be able to knock down visual barriers that are hampering automation inside the business.
The third piece of this flywheel is trust calibration. One of the persistent challenges with AI adoption is that humans struggle to verify whether AI outputs are correct. When the output is text, verification requires careful reading and domain expertise. But when AI can show reasoning visually — generating a diagram of a proposed solution, creating a visualization of a data pattern, producing an annotated screenshot highlighting the evidence it used for its conclusion — that makes verification so much faster and more intuitive for humans. Humans can look at a visual output and quickly assess whether it makes sense. This accelerates trust-building in a way that has genuinely gated deeper AI adoption, because people just don't intuitively trust all the time.
I don't use the news example lightly. I have actually checked AI's ability to pull news for me by looking at Nano Banana visualizations of longer news presentations that I've had generated in Perplexity. It's a single prompt — you take the Perplexity news piece, paste it into Nano Banana, and say "please make an attractive infographic highlighting this week's news," and it just does it. You can immediately see: does it match the real headlines? Is it hallucinating something? And by the way, it's pretty accurate. You can also digest the news more quickly because we're visual people. So having the ability to create images, ironically, has human effects that accelerate AI adoption in our organizations. We are calibrating trust more deeply with AI because we can use images to scale trust mechanisms across text-based workflows.
The fourth part of this flywheel is workflow integration. Once visual AI capabilities are proven in particular applications across the business, they become connectable components — Lego bricks in larger systems. Image generation capability ends up connecting document production capabilities to customer communication capabilities to analytics capabilities. Implementations start to feature really interesting bidirectional information flows where you can easily translate across and say: can you please draw a graph for me of the customer tickets that you triaged, and because you have image generation capability, can you show the product team on the page where customers are having trouble right now? That sort of thing is new. It highlights how images form a kind of universal connector that helps link information flows across the business and accelerate workflow integration.
The four stages start to compound. More automatable surface area generates more data. More data lets you build trust faster because you're engaging with the data flows that drive the business. All of that in turn enables humans to drive deeper workflow integration, and that exposes even more automatable surface area as we discover parts of the business where we didn't realize there was a visual component. I have never seen, for example, a product team receive a report from a customer success team where they triage the last thousand tickets and say "let me visually identify the parts of the page that hurt." That is now possible for the first time. There is additional automatable surface area that wasn't even accessible before that these technologies make possible.
Where visual AI creates the most leverage — and it's not marketing
If you accept that visual AI capabilities are maturing rapidly — and I think most of us do — the strategic question becomes: where in the business do these capabilities create the most leverage? The obvious answer that everyone gives is incorrect. The obvious answer is marketing and design. Creative teams will be able to produce more assets faster at lower cost. Absolutely true and valuable. But I would argue that's not where the primary leverage for image generation lies.
Creative functions are already staffed to handle visual work. They already have budgets for it. And in practice, what I have found is that when they get the ability to generate a hundred times more visuals, they are not well trained and not well supported to pivot into an editing function — into selecting the right image, into going from generation to something new. That is absolutely a role they need to take. It is part of the organizational creative journey that creative departments are on. It is important, but it is not necessarily transformative to the entire business.
The primary leverage for image generation in AI lies in functions that have been artificially constrained by their inability to work with visual information at all. Functions that process information, make decisions, communicate with stakeholders, and coordinate activities, but have had to route around visual elements as they use AI rather than through them.
Let me get more specific. Customer operations is a key example. Support interactions increasingly involve visual information because customers expect it. Customers will send you a screenshot of errors. They'll send you photos of defective products. They'll send you images of interfaces that don't make sense to them. When AI systems can now interpret those visual signals and respond with correct visual outputs — an annotated guide to solving the problem, a generated diagram showing the correct versus incorrect way to plug in your router, a visualization of what the resolution to your software problem should look like — you are multiplying the resolution time savings you can get because it can happen in real time. Human agents can then focus on genuinely complex cases rather than routine visual interpretation.
Product management is another one. Product managers spend an enormous amount of time on communication artifacts — roadmap presentations, competitive analysis decks, feature spec documents, stakeholder updates. These artifacts are heavily visual because that's how product information is often most effectively communicated, and because that is how the product is often shipped. What does it look like matters in product management. When AI systems can generate these artifacts programmatically, pulling from product databases, rendering competitive landscapes visually, creating spec docs with actually working real diagrams, PMs are going to spend less time on artifact production and more time on the strategic decisions that the artifacts are meant to support — which is really where they need to be.
Training and enablement is another good example. Employee onboarding, customer training, partner enablement — all of these functions rely heavily on visual materials that are expensive to produce and expensive to maintain. When product interfaces change, training materials become rapidly outdated and often a liability. When processes evolve, onboarding documentation falls behind. I don't know how many times I've been in onboarding and they point you to the wiki and say "but it's outdated." Organizations that solve this problem typically do so by either accepting outdated materials or by dedicating a lot of headcount to maintenance. Visual AI offers a new path. Materials can update themselves as underlying systems change. Personalized visual explanations can be generated on demand. Training content can adapt to the learner with visual explanations rather than requiring the learner to adapt to static materials.
In all of these cases, the common thread is this: the value proposition is not making existing visual work more efficient. It's enabling visual communication that is extraordinarily effective in contexts where it was previously just unviable. Organizations don't just produce the same materials faster. They can now produce visual materials reliably and in an up-to-date way in situations where they previously relied on text alone because visual production was too cumbersome.
The 30% vs. 300% distinction: point solution versus infrastructure
This brings us to a distinction that separates organizations capturing modest value from AI from those capturing transformative value. I call it the 30% versus 300% distinction. It's the same distinction that separates organizations achieving modest productivity gains overall from those achieving order-of-magnitude improvements as they ship AI.
If you're using visual AI and you're a 30% organization, what does that look like? Typically, you're deploying it in the design department. Designers are using it to generate concepts faster, you're producing variations more efficiently, the design team probably becomes more productive overall. There are a lot of role changes that have to happen, and they may become significantly more productive. But the impact tends to be bounded within the design team's existing footprint. If you make the design team 30% more efficient, you're not changing the story for the business as a whole.
What do 300% organizations do differently? They treat visual AI as infrastructure. They recognize that visual generation and interpretation are capabilities that can be embedded in systems throughout the enterprise — not just creative tools. They build pipelines where visual AI capabilities are components in automated workflows. Sales systems can now generate pitch materials dynamically from CRM data. Customer support systems can interpret incoming visual information and respond with visual explanation. Product systems can maintain their own documentation as features evolve.
The difference is not really sophistication. It's about where you place the capability in your architecture. A point solution lives in a department. Infrastructure lives in all of your systems at once. Point solutions improve the productivity of the people that use them. Infrastructure changes what your systems can do as humans design, supervise, and handle edge cases.
Let's walk through an example. Say you're an e-commerce company and you previously employed a team that produced product photos. They would receive product samples, photograph them professionally, and edit the photos. Now, with visual AI as a point solution, the same team can generate those product photos. Their productivity improves and everyone seems happy. That's the 30% story.
The 300% story: if visual AI is infrastructure at this e-commerce company, product photo generation is embedded inside the catalog management system. When a new product gets added with basic specs, the catalog system automatically generates appropriate product photos, sizes them for different displays, and populates the catalog without human involvement. Human reviews are there to flag exceptions, and the photography team can actually redeploy toward higher-value creative work — brand work and other things that also use visual AI.
Five questions leaders should ask to unlock visual AI in their organization
Let me close with the questions I think leaders should be asking to unlock this in your organization. If visual AI represents a force multiplier for enterprise AI adoption — and I believe it does — then we need a framework to assess where to invest, and it will look different depending on what kind of organization you are.
The first question is: where in your organization do visual communication bottlenecks slow decisions? Where are you taking a week to produce something that has a visual component? Where are customer-facing materials out of date? Where is technical documentation always behind? Look for the bottlenecks, and more importantly, look at each bottleneck as a place where faster visual communication could improve decision quality and execution speed — not just take less time.
The second question is: which workflows currently break because they require human visual interpretation? Look in quality control, customer support, document processing, onboarding. There are automation boundaries that we may have assumed were permanent that are just not there anymore. The question is which of these boundaries, if removed, would unlock the most upside — and which of these give people new things they can do that are more interesting? Because I would not want to stare at images of routers being plugged in and out.
Third: what would change if visualization were instant and programmatic? Could you personalize customer materials at the individual level rather than the segment level? Could you test fifty visual variants of a campaign rather than three? Could you brief executives with real-time visual dashboards rather than static weekly documents? Could you maintain documentation continuously rather than in periodic update sprints? The answers to these questions reveal where visual AI might take you well beyond efficiency gains. The highest-leverage investments are typically in net-new capabilities, not in cost reduction.
Fourth: where are you building visual dependencies into human roles that will become bottlenecks as you scale? If your current growth plan assumes that certain visual tasks will always require human involvement, now is the time to revisit that. It's much better to think about your vision system as part of your infrastructure and to start looking at how humans — who are creative, who have incredible visual and artistic capabilities — can focus on next-level creativity with visual materials, including AI as a tool, than to assume that humans are just going to be part of a predictable workflow that includes visual processing. That kind of work is very AI-able at this point.
Fifth and most important: are you thinking of AI as a department tool or as organizational infrastructure? If you are buying visual AI capabilities and thinking of it as three seats on the creative team, you are only going to capture point-solution value. If you are building visual AI into the product catalog system, the customer support platform, and the documentation pipeline, now you're starting to capture infrastructure value. That distinction will determine whether visual AI contributes to marginal efficiency for you or whether you truly start to unlock transformative capability expansion.
The window of first-mover advantage is open now
There is a window during which visual AI infrastructure is a new thing. It will not be new forever. The fences have come down and not everyone has realized they're down yet. That window of first-mover advantage is just not going to be there forever. Right now, organizations that recognize visual AI as infrastructure can build systems that competitors can't match. You can get the flywheel going, start to generate learnings, and get ahead in a way that organizations starting in a year or two years are going to have trouble catching up with — because in two years, the capabilities we're talking about are going to become table stakes. Integration patterns will be well documented and shared out. What represents a real competitive advantage now, at the beginning of 2026, is going to be very basic operational capability by 2028.
The question is not whether your organization will eventually get there. It's whether you will be among the leaders who shape how it's deployed and who are able to drive those learnings to generate sustainable competitive edges because you have been working with images longer as a business.
Let me return to where we started. The conversation about AI image generation has been captured by the wrong people and it's being framed around the wrong question. The question is not which tool produces the nicest-looking outputs. It's not which tool has the best 4K image generation. It's not which platform has the best prompt engineering features. It's not whether Midjourney or Nano Banana Pro is best at faces.
The question is: what becomes possible when your organization's AI systems can see and show, when they could not do so before? When the automated chains that currently break at visual touch points can now run without a break. When the workflows that currently route to human queues for visual interpretation can now process autonomously.
This is the correct frame for visual AI. Not as a creative tool that makes designers more productive, but as an infrastructural capability that removes constraints around organizational AI deployment much more broadly. Companies that get this are going to invest accordingly in their image generation capability as true infrastructure, and those companies are going to get ahead.