AI automation is misunderstood and it’s an opportunity.

For a while, it was easy to dismiss AI-generated visuals as entertaining but not particularly practical. The videos were awkward, the avatars looked strange, and the slides often felt like they had been made by an overconfident intern with too much animation and not enough judgment. That is changing quickly. Visual asset generation is now good enough to make serious professionals pay attention. Avatars are becoming far more realistic. Slide and presentation tools are improving fast. And increasingly, AI can now produce communication materials that are genuinely usable, not just superficially impressive in a demo.

Most people still look at these tools and think about marketing first. It is reasonable because we use visual polish as a signal of overall quality of service. Real estate is a visual industry, and anything that makes it easier to produce videos, pitch decks, explainers, training content, and client-facing material will matter. But I think the more important point is broader than marketing. The cost of polished communication is collapsing. Things that used to require a designer, a video editor, a presenter, and quite a lot of time can increasingly be done by one person who knows how to direct AI well.

That changes who gets to communicate well. It also changes what good looks like. Once your competitors can produce a decent deck, a good explainer, and a clean video quickly and cheaply, the bottleneck moves away from production and toward judgment. What should be communicated? To whom? In what form? With what degree of nuance, caution, and commercial awareness?

That brings me to the point that matters much more than visual AI: everyone talks about automation, but in many important workflows what we are really doing is delegation.

That distinction matters because delegation is harder, but it is also becoming much easier than it was even a year ago.

We now have much better tools for long-running and autonomous work. It is becoming increasingly easy to schedule tasks, set up recurring processes, connect models to files and apps, and let them work through multi-step workflows with much less supervision than before. In my experience, some of the new cowork-style tools are genuinely excellent at this. They make delegation feel real rather than theoretical (at the same time, some products like Copilot Cowork still underwhelm badly). So the direction of travel is clear even if the quality of specific tools varies a lot.

This is where I think many firms misunderstand AI implementation. They think they are buying software that will simply automate work. In reality, they are trying to teach a very capable but slightly strange junior analyst how to operate inside a very specific workflow. And increasingly, they are giving that junior analyst not just a chatbot window, but a calendar, a file system, connectors, permissions, and the ability to execute tasks over time.

Take financial modelling. In principle, you may want every building to be reflected in the same framework. In practice, buildings differ. Contracts differ. Lease structures differ. Debt structures differ. Risks differ. Sometimes the correct treatment is not obvious. You need to decide how to reflect something unusual in the cash flow, whether an expense should be normalized, whether a covenant should be interpreted conservatively or commercially, whether a risk deserves a scenario adjustment or a different discount rate. That is not simple automation. That is delegated judgment.

The same is true of strategy, underwriting, investment presentations, invoice processing, market research, and portfolio analysis. Even something that sounds routine, like processing invoices, can require a judgment call. Was something coded correctly? Is this a real exception or just an odd but legitimate case? Does this line item belong here or somewhere else? The more commercially important the task, the more likely it is that judgment matters.

If the task involves judgment, the challenge is not merely to automate it. The challenge is to delegate it responsibly.

That means clear instructions, examples, and constraints. It often means training. It always means governance. And it absolutely means verification. The fact that delegation is getting easier does not mean that checking becomes less important. In fact, the opposite is true. The easier it becomes to hand work to AI, the more discipline we need around review. The firms that do this well will not be the ones with the loudest AI strategy. They will be the ones that design workflows that are auditable, testable, and easy to supervise.

This is why I increasingly think the most useful analogy is not automation software but human delegation. Imagine hiring a Nobel Prize-winning physicist to help with property analysis. They clearly have intellectual capacity. But if you want them to do the work exactly the way you need it done, using your assumptions, your context, your formatting, your risk preferences, and your practical constraints, you still need to explain the process properly. Raw intelligence is not enough. Context, examples, standards, and feedback matter enormously.

That is what using AI feels like now. The models are incredibly capable, but capability alone is not the product. The real product is a combination of model quality, tools, interfaces, connectors, permissions, and workflow design.

Even when model behavior becomes a little less impressive on some tasks, overall usefulness often continues to rise. New tools are added. Better interfaces appear. Connectors improve access to files, email, calendars, and internal data. Scheduling becomes easier. Computer-use capabilities improve. Models gain ways of acting, not just answering. So although there may be some decline in how sharp a model feels on a narrow benchmark or a favourite prompt, the total value of the system can still increase.

That is an important lesson for professionals and firms. The question is not simply: which model is smartest today? The more useful question is: which setup lets me get reliable work done inside my workflow? How do I connect the model to the files, tools, and approval steps that matter? How do I make the output easy to verify? How do I build a process that still works when models change, improve, or drift?

I think this is where AI is going over the next few years. Not toward a world where one perfect model automates everything, but toward a world where delegation becomes progressively easier, cheaper, and more normal. The underlying models will improve, but just as importantly the surrounding systems will improve. We will get better task scheduling, better agents, better connectors, better workflow orchestration, and better enterprise controls. In practice, that will matter as much as raw intelligence and often more.

So what should people expect? First, more work will be delegated in pieces before it is ever fully automated. Second, the best results will come from workflow design, not from simply buying access to a powerful model. Third, verification and governance will become core skills rather than optional extras. And fourth, the firms and individuals who learn how to combine model capability with process design will move much faster than those who focus only on whether AI is theoretically intelligent enough.

My practical advice is simple.

Do not wait for a perfect model. Start by identifying tasks where judgment is limited but still present, because those are often the best delegation opportunities. Build workflows where AI can do the first pass, the heavy lifting, the search, the formatting, the synthesis, or the repetitive analysis. Then make review easy and explicit. Freeze sources where possible. Keep approval points clear. Test edge cases. Rerun important tasks periodically. And focus less on hype and more on designing a system you would actually trust in a live commercial setting.

In other words, stop thinking only about automation and start thinking like a manager of AI workers.

That is where the real opportunity is now.

Yes, visual AI is becoming extraordinary. Yes, the models will keep improving. But the bigger shift is that delegation is getting easier. And once firms understand that, they will stop asking whether AI can do useful work and start asking how much of their workflow they can redesign around it. That, I think, is where this is all going.