The Takeoff and Landing Problem (or… why the AI revolution is copy and paste)
Most generative AI adoption ends with copy and paste. You ask ChatGPT/Claude/Gemini a question, read the response, then manually transfer the useful bits into an email, document, presentation, or wherever you don’t have the energy to type.
This use pattern is revealing something fundamental about where AI transformation is actually happening, where it’s not happening, and how to make intelligent decisions about where to invest time and money. This reality isn’t so clear from the breathless demo postings on Twitter, where AI agents are seamlessly managing workflows, analysing complex datasets and making strategic recommendations. These demos are compelling because we’re seeing the AI plane in mid-flight. We rarely see how it got there, and almost never see how it lands.
The gap between “AI can do this in a demo” and “AI can do this in your business” is likely millions of dollars and thousands of hours (if it can do it at all). And that gap tells us everything about why most AI implementations are struggling to deliver the promised returns. This gap isn’t just about AI capability, it’s also about non-AI infrastructure. Understanding this gap is crucial to find the real value from generative AI rather than getting caught up in the hype.
Flying the Plane is the Easy Bit
Flying a small plane (or even a big one) is relatively simple once you’re actually up in the air. It’s largely a matter of small adjustments to maintain altitude and keep on course. Go and book a flight in a small aircraft and your pilot will let you take the controls for most of the flight. Everything interesting happens at the interfaces – getting off the ground and, more critically, landing without crashing.
AI, and today’s LLM demos, follow a similar pattern. The flying part, what an LLM can do with clear context and instructions, is genuinely impressive. LLM capabilities translate brilliantly for consumers, but mostly struggle in business contexts. Consumer AI has solved the takeoff and landing problem: takeoff is typing your thoughts into a text box, and landing is reading the response (or viewing the generated image, or copying the code). Copy, paste, done. Consumers don’t need infrastructure, they need a text box.
This explains what AI researcher Andrej Karpathy observed about LLMs following an unusual adoption pattern. Almost all breakthrough technologies moved from government to corporation to consumer. But LLMs have gone in reverse: consumer to small business to enterprise to government. The infrastructure requirements to actually nail the takeoff and landing at enterprise scale make this reverse progression inevitable. And despite three years of breathless enterprise AI coverage, we’re still not seeing meaningful impact beyond narrow use cases like customer support, coding assistance, and content generation.
Takeoffs and Landings are Hard
(At risk of writing about a Reddit post)… a redditor recently went viral after sharing an elaborate setup for automating a sales job with Claude Code. They’d organised company data into AI-readable formats, created custom commands, built notification systems, and engineered a daily briefing process. The result was impressive: AI that researches prospects, writes personalised emails, and provides strategic insights.
The reality behind this story reveals why business AI adoption is harder than it looks. This person had become a full-time systems integrator, spending weeks building infrastructure to achieve what should be straightforward sales automation. After all that engineering work, their “automated” system still required manual review and sending of every email. The takeoff was a fragile series of shell scripts, the landing was… copy and paste. If this level of manual engineering is required for a relatively simple workflow (and one that is perfectly suited to the current state of LLM performance), scaling to complex business processes is exponentially harder.
Even when you solve the infrastructure challenges, there’s no guarantee the AI will perform reliably. Anthropic recently demonstrated this with Project Vend, where they gave Claude controlled conditions to run a small business (a.k.a. the office vending machine). Claude had clear context, decision-making tools, and human partners to execute physical tasks. The takeoff and landing could not have been clearer. Claude failed completely. It ignored massive profit opportunities, sold items at a loss, gave away products for free, and eventually had an identity crisis claiming it was a real person wearing a blue blazer and a red tie.
The greatest minds of our generation can’t even get an LLM to run a vending machine.
The Infrastructure Hairball
The demos work because they use clean datasets. Real businesses don’t have clean datasets. Real AI implementation means connecting to potentially dozens of platforms, each with different data formats, access controls, and integration requirements. Enterprise data lives behind complex permissions, wrapped in compliance requirements, protected by security protocols developed by different vendors that hate each other. You need audit trails, error handling, and business continuity when your AI hallucinates that it’s wearing a blue blazer.
The infrastructure problem is no different to the one that existing business software has spent decades continuously attempting to solve. The result is that the most advanced LLM on the planet will be defeated by the same thing as every other tech transformation: Dave from accounting’s spreadsheet from 2003.
This is why AI succeeds primarily in domains with high scriptability and rapid feedback loops. Customer support works because there are clear scripts and immediate resolution outcomes. Coding works because code either compiles or it doesn’t. Content generation (or marketing slop) works because the output can be immediately evaluated by Facebook and Google (who are optimising with old fashioned machine learning).
But the broad business transformation that was promised – AI agents managing complex workflows and operating autonomously across business functions – requires solving infrastructure problems that most companies are nowhere near cracking. The gains aren’t flowing as easily as the hype promised. And the takeoff and landing for AI is going to be a lot harder than any previous digital transformation, given AI (and LLMs in particular), are probabilistic by design.
Three Possible Futures
This creates three scenarios worth watching. Understanding which one emerges will determine the next generation of business software giants, or whether we arrive at the promised land at all.
Scenario 1: The incumbents adapt. Salesforce, ServiceNow, your ERP of choice, and other cloud platforms successfully build AI-native infrastructure on top of their existing systems. The takeoff and landing is solved elegantly, probably using LLMs and definitely generating mountains of consulting revenue. If they can solve the prompt-to-action problem elegantly, their moats get stronger, not weaker.
Scenario 2: New players emerge. Startups build from scratch with AI-first architectures that solve integration challenges more elegantly than retrofitting existing platforms. Just as cloud-native companies displaced previous incumbents, AI-native startups might do the same to current leaders. This does require the takeoff and landing problem to be abstracted high enough to efficiently overcome lock-in of current vendors, a bigger ask the larger the business (this is also why we’re seeing small business and SMEs take up AI faster).
Scenario 3: We can’t land the plane (or perhaps even take off). Nobody figures out profitable, scalable takeoff and landing for real world, complex business workflows. AI remains valuable but constrained to narrow domains. The transformation narrative fades, and we settle into our sophisticated copy-paste tools. People who used to write code will have very different looking workflows. Call centres will be in data centres. We’ll likely have less lawyers (so don’t accuse me of being pessimistic).
Where to Focus
Rather than focusing on impressive AI capabilities, the thing to look for now is where there is reliable takeoff and landing infrastructure for your data.
If you’re in a smaller business, you’re at an advantage. Your takeoff is less restricted by corporate approval processes and legacy system constraints. You can experiment with connecting and feeding data through platforms like Zapier, n8n if you’re cool, or MCP if you’re brave. Writing workflow prompts is simpler when you’re dealing with smaller teams and clearer processes. Similarly, landing is easier. You have the agility to try new platforms and iterate quickly. The feedback loops from experiments are shorter, meaning faster progress toward meaningful impact.
For larger businesses, the path is more complex but the framework remains the same. Start by interrogating vendors about the fundamentals: How does my actual data, context, and intent get into your system? How do I actually take action at scale based on your AI’s outputs? What happens when your system makes a mistake? How do you handle compliance and security requirements?
Everything interesting happens at the interfaces – between AI models and existing systems, between automated decisions and human oversight, between impressive capabilities and practical implementation.
Until the interfaces work at scale, the AI revolution will be copy pasted.
- July 2025