The Judgment Gap

For most of history answers were hard to come by. You had to track down books, find experts or learn through experience. Scarcity gave knowledge its value.

Now answers arrive instantly. They're everywhere, basically free. And when answers cost almost nothing, they're not what sets you apart anymore.

The bottleneck shifts to the question itself. What you ask. Why you're asking. Whether you can tell when an answer is good enough or when it's misleading in ways that aren't obvious.

In this new world, edge goes to people who can define problems clearly, spot what's missing and push past the first explanation that sounds right.

When answers are abundant, what matters is judgment.

How Enterprise AI Creates Value Beyond the Interface

It’s tempting to talk about enterprise AI as if it always shows up behind a prompt box or conversational interface. Many examples do, but that’s not how a lot of enterprise systems actually work. In practice, AI often runs in the background, embedded into workflows, triggered by events, or invoked only at specific moments. Users may benefit from it without ever interacting with something that feels explicitly like “AI.”

As a result, the user experience is changing in quieter ways. Instead of asking systems for help, users increasingly encounter work that has already been partially completed, pre-validated, or queued for review. Interaction becomes less about issuing instructions and more about supervising, correcting or confirming outcomes. That distinction matters, because it changes where value is actually being created.

AI Doesn’t Have to Look Like AI

In many enterprise environments, intelligence appears indirectly. A form might already be filled in with the most likely values. A recommendation might appear only when uncertainty is high. A task might be resolved without user input at all, surfacing only when human judgment is required.

From the user’s perspective, the experience feels less like using an AI tool and more like working in a system that anticipates what needs to happen next. The interface remains familiar, but the interaction model shifts. Users spend less time asking for assistance and more time responding to decisions the system has already proposed.

This is one reason visual sameness can be misleading. Two systems can look identical on the surface while offering very different experiences in how and when they involve humans.

Whether AI is exposed through a chat interface, a form, an API or not exposed at all, there is still a layer that determines what information matters in a given moment, which systems can be accessed safely, how actions should be ordered and when the system should act autonomously versus pause and involve a human.

That layer quietly shapes the human-agent relationship. It decides whether users are constantly interrupted or only engaged when necessary. It defines whether AI feels helpful or intrusive, predictable or opaque. In many cases, users never see this logic directly, but they feel its effects every day. As AI systems take on more responsibility, this layer becomes more important than any individual interaction. The interface becomes a checkpoint, while trust is built through consistency rather than conversation.

Earlier generations of enterprise software focused on optimizing user interaction. Systems of record captured data. Systems of engagement improved collaboration and usability. Many AI-driven systems shift attention away from interaction altogether and toward outcomes.

The best experiences are often the quiet ones, where work progresses without friction and human attention is reserved for exceptions rather than routine decisions.

AI changes the role of interfaces. A place for oversight, review and intervention rather than constant input. The quality of the experience depends more on whether the system involves humans at the right moments.

None of this is settled. Models are improving, tooling is evolving, and expectations around autonomy are still forming. Some enterprise AI systems will remain conversational, others will move toward agent-driven automation, and many will blend the two depending on context.

What seems consistent is that value increasingly accumulates in the parts of the system that manage this balance between human judgment and machine action. These parts are often hard to see and harder to demo, but they define whether AI feels like a burden or a natural extension of how work gets done.

As the technology continues to change, so will the interaction patterns. What matters most is designing systems that respect how people actually work and when they want to be involved at all.

When to Use RAG in Enterprise AI Systems

RAG (Retrieval-Augmented Generation) has become one of those terms that shows up in almost every enterprise AI conversation. Need access to internal documents? Add RAG. Need the model to be “grounded”? Add RAG. Over time, it’s started to sound less like an architectural choice and more like a default checkbox.

RAG is a useful pattern however it introduces trade-offs in reliability, cost, latency and system complexity. Used in the right place, it quietly does its job. Used in the wrong place, it creates systems that are hard to debug, hard to trust and expensive to operate. Understanding where that line is matters more than knowing how to set up a vector database.

What RAG Actually Solves

At its core, RAG exists to solve one specific problem: large language models don’t have access to your private or up to date data. Retrieval is simply a way to inject that data into the model at the moment it needs to answer a question.

This works well when the knowledge you care about lives in documents, changes over time and needs to be referenced rather than memorized. Policies, internal wikis, product documentation and legal text fall neatly into this category. In these cases, RAG acts like a just in time reading mechanism. The model doesn’t need to “know” your documentation. It only needs to read the relevant parts before responding. That distinction is important, because it also defines the limits of the approach.

RAG performs best when the system’s main job is to surface and explain existing information. If the expected output looks like a well-written summary, explanation or answer grounded in source material, retrieval adds real value. It allows the system to stay current without retraining, makes updates operational rather than technical, and enables traceability, something enterprises care deeply about.

Another signal that RAG is a good fit is when answers are allowed to vary slightly in phrasing but not in substance. The model is synthesizing not deciding. When users want to see why something is true and where it came from, RAG aligns well with that expectation.

Problems appear when RAG is used to support logic heavy or decision critical systems. Retrieval is probabilistic. Chunking, embedding quality, ranking and context limits all introduce uncertainty. That uncertainty is manageable when the model is summarizing a policy, but it becomes dangerous when the model is deciding eligibility, pricing or risk.

In those cases, the system is failing subtly. The model may retrieve most of the right context, miss one key clause and still produce a confident answer. From the outside, it looks grounded but under the hood, it’s inconsistent. RAG may also not be the best choice, when the knowledge base is small, stable and central to the product. If the entire business logic fits on a few pages, introducing embeddings and retrieval layers often creates more surface area for errors than value. A well structured prompt or deterministic logic will usually outperform a retrieval pipeline in both reliability and maintainability.

Latency and cost are another quiet tax. Every RAG call adds retrieval time, additional tokens and more infrastructure. In high-volume or real-time systems, these costs compound quickly and are hard to claw back later. One of the most common arguments for RAG is that it “prevents hallucinations.” In practice, it doesn’t prevent them it changes their shape. Good retrieval reduces the chance of the model inventing facts, but poor retrieval produces answers that sound authoritative and cite the wrong context. That can be worse than a visible hallucination, because it gives users false confidence. 

What to Use Instead (or Alongside)

Many enterprise use cases don’t need retrieval at all. Carefully designed prompts, examples and constraints can go surprisingly far, especially for workflow assistance, content generation and structured reasoning tasks. This approach is easier to debug, cheaper to run and often more predictable.

Fine-tuning is another option, but it solves a different problem. Fine-tuning is about shaping behavior. Tone, style, reasoning patterns, not about injecting fresh knowledge. It works best when the information is stable and the desired output needs to be consistent. When facts change frequently, fine-tuning becomes operationally expensive and brittle.

The most robust systems tend to be hybrid. Hard rules and structured logic handle decisions. Retrieval provides reference material. The language model focuses on explanation, synthesis, and interaction. This separation keeps critical logic deterministic while still benefiting from the flexibility of natural language generation. For structured or numerical data, traditional databases and APIs remain the right tool. Vector search is not a replacement for SQL. Let the model call tools instead of guessing over text when precision matters.

In short, RAG is infrastructure and users only care if answers are correct, consistent and fast. The strongest enterprise AI systems are usually the ones that resist architectural fashion and choose the simplest setup that can reliably meet those expectations.

When Software Learns to Move

I've been spending time lately learning about the intersection of AI and robotics. There's this whole world where AI isn't generating text or summarising documents. It's navigating buildings, flying aircraft, and making split-second decisions in places where there's no internet connection and no human nearby to ask for help.

The more I read, the more I realise how different the problems are.

Intelligence under constraint

When AI runs on a drone or a ground vehicle, the margin for error is physical. A wrong decision produces a crash instead of a weird written paragraph. The system has to perceive its environment through cameras and sensors, figure out where it is, plan a path and act on it. All in real time, all on hardware with limited power and compute.

One technology that caught my attention is GPS-denied navigation. We take GPS for granted, but it doesn't work indoors, underground or in environments where the signal is jammed. So autonomous systems have to build their own understanding of where they are using whatever sensors they have. There's a technique called SLAM, simultaneous localisation and mapping, where the system constructs a map of an unknown space while tracking its own position within it.

Companies like Shield AI have built their entire approach around this problem. Their aircraft can fly and coordinate autonomously where GPS doesn't work. 

The thing that really pulled me in is multi-agent coordination though. Not one machine doing something smart, but dozens of them sharing what they see, dividing tasks and adapting as a group without a central controller. The coordination has to be decentralised because you can't rely on a single point of communication. If one link goes down, the group keeps going. Each agent makes its own decisions while staying aligned with the others. 

It's easy to think of this as purely military technology and a lot of the current momentum is coming from defence. But the same capabilities apply to inspecting infrastructure that's difficult or dangerous to reach, monitoring agriculture, running search and rescue in disaster zones or handling logistics in remote places with unreliable connectivity.

The common thread is simple. Move the intelligence closer to where the work actually happens instead of keeping it behind a screen.

The questions being worked on in this space feel foundational. How do you build something smart enough to act on its own and reliable enough to trust? Where do you draw the line between autonomy and human oversight? What does trust look like when the thing making decisions is a machine in the field?

I'm going to keep learning about this. If anything interesting comes out of it, I'll share it here.

Empathy in the Age of Algorithms

AI can read a thousand conversations and still completely miss what it was like to actually be in one. It can pick up on tone, flag emotions, even write an apology that sounds convincing. But it has no idea what it means to actually care about someone.

Empathy ≠ spotting patterns in what people say… It's being there. It's the pause before you respond because you're trying to really hear what someone means. It's caring more about understanding them than being understood yourself.

The danger is though when we start to believe that AI has empathy... and let it draft the condolence message, handle an important customer complaint or respond to the hard conversation. And the more we do that, the less practice we get at the real thing.

When Everyone’s a Creator

The idea that everyone can create sounds amazing. Until you realize that when everyone is creating, most of it might just becomes noise.

What we need is better judgment about what's worth paying attention to, that is original, creative and valuable.

The people who matter creatively will be the ones who can tell the difference between what's good and what's just there. Curation > Generation.

The Disappearing Line Between Thinking and Doing

We used to have to think things through before we could do them. Now we can just start. AI lets us build while we're still figuring it out. Every half-formed idea can become a prototype in minutes.

That's exciting, but it's also risky. When execution is that easy, thinking starts to feel optional. Why sit with a problem when you can just generate five solutions and pick one?

But the whole point of technology is to give us more time to think better. The people who get that will still slow down sometimes, because good judgment still takes longer to develop than quick execution.

Beyond Efficiency

Every time technology changes, we tell two stories. One about what we're losing. One about what we might gain.

AI is already taking over a lot of routine work, the templated emails, the data entry, the meeting notes. And the optimistic take is that this frees us up for deeper work. More time for creativity, strategy, human connection. The stuff that actually matters.

I want to believe that. But I don't think it happens automatically.

Because here's what we've seen before; productivity tools were supposed to give us time back, but we usually just filled that time with more work. Email made communication faster, so we sent more emails. Spreadsheets made analysis easier, so we analyzed more things. Every efficiency gain tends to get absorbed into doing more, not thinking deeper.

So yes, AI could give us back human attention. But it depends on what we do with it. Whether organizations see automation as a chance to do better work, or just a way to do more work with fewer people. Whether we use the freed-up time to actually focus, or just pack in another meeting.

AI is not the end of human contribution. What happens next is still up to us.