AI will blur the lines between many functions over time because you can now begin to do things either higher or lower in the stack, or expand to other adjacent functions. A very obvious area is that PMs should almost always be showing up with functional prototypes.

At Box, we spend a lot of time testing Box AI with new models on unstructured data to see what they perform well at in real areas of knowledge work. As we've seen from the benchmarks, GPT-5 offers a meaningful jump in capability over GPT-4.1 in reasoning, math, logic, coding, and other areas of work. Here are a few examples of where those improvements come into play in the real world: *GPT 5 contextualizes information better. When doing data extraction like the final USD amount on an invoice without currency labels but with an address in London, GPT 5 correctly responds needing a conversion rate from USD to GBP. In comparison, GPT 4.1 saw the final bill and returned it, assuming the currency (incorrectly). * GPT-5 delivers better multimodal analysis. For a public company’s annual filing, GPT-5 is asked to isolate a cell in a table from an image showing changes in company’s equity components. The top of the table clarifies that all share amounts are in thousands, and GPT-5 clearly states this conversion, whereas GPT-4.1 does not, getting confused given the table says stock and the legend says shares. * GPT-5 performs better with high levels of prompt and data complexity. When doing data extraction on a resume for all job start dates, job position names, and employer names, GPT-5 was able to pull out every piece of data while GPT-4.1 seems to get overwhelmed and didn’t extract the same fields given the size of the prompt and complexity of the document. * GPT-5 is much more clear and explicit in its answers. In an outsourcing agreement with 6 different services explicitly discussed, when asked about “the 5 specific services in the contract,” GPT-5 will return the first 5 and ask if it was intentional that the sixth wasn’t asked about. In comparison, GPT-4.1 simply returned the first 5 without any further caveats, which can lead to downstream confusion for the user. * GPT-5 is better at data interpretation in complex fields. For a flow cytometry chart, typically used in immunology, GPT-5 correctly identified a high proportion of dead cells and gave plausible root causes that might lead to the situation while GPT-4.1 gave minimal reasoning, needing further confirmation to have any guesses from raw data. * GPT-5 is better able to identify inconsistencies in code. When asked to identify problems in a given python code file, while both GPT-5 and 4.1 can identify real errors that lead to malfunction, only GPT-5 was able to infer more subtle issues, like printing the incorrect variable when that would not make sense in the context of the program. These improvements in math, reasoning, logic, and quality of responses in longer context windows are incredibly helpful to end-users in daily work, but they will show up even more with longer running AI agents, especially when there's no human in the loop to verify the information at each step. It's awesome to see these improvements keep coming in the latest crop of AI models as this will lead to AI agents able to be used in incrementally more mission critical areas of work.

It's clear that we're on a trajectory right now of AI models continuing to improve in capability across math, reasoning, logic, tool calling, and various domain-specific tasks that will get better as more training data continues to get generated. While there will be debates about how much these advancements will show up as major changes in the daily use cases that a consumer has, they *will* have a major impact across many categories of knowledge work. They will incrementally unlock new use-cases in healthcare, legal, financial services, life sciences, etc., where models can reliably perform progressively more critical tasks. On a recent podcast with Alex Kantrowitz, Dario Amodei had a great way of framing this, which is that if you improved the capability of an AI model to go from having an undergrad in biochemistry to a having a graduate degree in biochemistry, a small percentage of the consumer population would notice the impact, but the enterprise use cases for a company like Pfizer would go up meaningfully as a result of this. We should start to anticipate that this is now the era that we are in with AI. So, how does this begin to show up in the real world? It will show up through AI agents going after applied use cases. AI agents for coding, legal work, medical scribes, data extraction, insurance claims processing, pen testing, and so on. The opportunity right now is to build AI agents for verticals and domains with a deep understanding of that space. This is where the impact of context engineering, a deep understanding of the workflows, connections into enterprise data, and specialized user interfaces (that allow users to deploy, manage, and orchestrate these agents) will start to matter a ton. It will also mean building up distribution that aligns to that particular vertical or domain. It will probably mean some form of forward deployed engineering to not only help customers implement the agents, but also quickly learn which workflows the agents are optimized for and bring that back into the core platform. Ultimately, these markets will be won by the players that can best bridge today's enterprise processes (which are often messy and were not designed for automation) to a world where agents are integrated into these workflows. This is the era of AI we are now in.

We’re nowhere near the optimal point for any stabilization in AI capability, but ironically there are lots of commercial use cases for AI that just get frozen because of how quickly the tech advances. This is why future proof architectures matter so much.

Core vs. context is a critical concept to think through when figuring out what people will rebuild themselves with AI. Companies bring in “core” functions that differentiate them. This is what their core product or service is, how they sell to customers, things that drive their culture, and so on. Conversely, they outsource the “context” that is table stakes to get right, but only offers downside in getting wrong. An easy rule of thumb to think through is would a customer ever notice if the company did that function directly themselves or not. Enterprise software is almost always “context”. These are areas like their CRM or HR systems, infrastructure, data management, and so on. These are necessary to operate a business at scale, but rarely are you advantaged in trying to roll your own. Only a few exceptions exist, and it’s almost always because you need a solution to serve your “core” that no vendor offers (like if you needed custom software for a vertically integrated supply chain). No matter how a company starts, they eventually almost always separate work and value between core vs. context over time. It’s the only way they can stay competitive and eventually allocate resources to the optimal areas. So even if a company *could* rewrite their enterprise software with AI, they basically just wouldn’t. The version updates, security, regulatory features, bugs, SLAs, the professional services necessary, etc. just all would make it ROI negative. As bucco points out, the real risk is better versions of these tools that are AI-first. That’s what to watch out for from a disruption standpoint.

Great thread. Whatever an AI agent is capable of doing, it can also be tricked into doing. You should assume if an agent can access data, that a user can eventually get that data too. Agent security, access controls, and deterministic guard rails will be critical.

Lots of conversation on what the future of software looks like in the enterprise. Here’s how I think it plays out. For deterministic workflows where the cost of getting something wrong is high, enterprises will have a tendency to pick core platforms for their most common, important, and repeatable functions in the organization. Think payroll, ERP, CRM, ITSM, customer support, ECM/document management, and so on. These are areas where you want something done the same way, every time. Each of these platforms will have to be AI-first by design, which means that they will have user interfaces that become tuned for interacting with the workflows and data via AI, and be fully designed for AI agents to operate in the platforms. Over time, we can expect usage on these systems to bias far more to AI agents than even people. The seat model remains for the users, but consumption becomes the model for agents. Some incumbents will make it to the end state, but others will not adapt quickly enough and die off. There will then be a new crop of effectively Agent-only companies that are purpose-built for automating specific types of work (and especially for non-deterministic work). Their business models will tilt even more consumption. Think Claude Code or Devins (likely with some UI layer for managing the agents) but for various job functions. We will likely see hundreds or thousands of these emerge over time. Pen testing, coding, bug finding, compliance reviews, financial analysts, and so on. This is a huge space where startups will do quite well because there will tend to not be any software incumbents in these categories. We will interact with these various agents from a mix of the software platforms that they are tied to (like Box AI, or Agentforce), via APIs in other systems, and horizontal workflows systems that stitch together agents across platforms (like ServiceNow, IBM Watsonx, Google Agentspace, and so on). And of course, users will often consume these agents via horizontal chat experiences (like ChatGPT, Claude, Perplexity, Grok, Copilot, etc.) via MCP or other types of direct connections. Users will commonly work in these horizontal chat systems, pulling in the agents, data, and workflows from the various Agentic platforms as needed. When relevant, l they will hop into the core platforms to complete workflows, review information, etc. There will also be a longtail of experiences where users can generate micro apps on the fly when they need quick applications or use cases automated, when there’s no obvious piece of software to do that with. This may happen directly in the horizontal chat systems, a tool like Replit, Lovable, or in workflow automation tools, etc. I’d expect this is more for the power users where they need glue between multiple systems or where no software yet exists. The net of it is that software becomes only more important over time, even if the modalities where we’re interacting from change and expand. Similar to how we hop between our phones and desktop computers with ease, even though they could easily converge, the future will offer a mix of ways of interacting with software.

While fascinating, the idea of AI generating every UI on the fly is probably less likely than people think. The benefits of hyper customization likely won’t outweigh having to re-learn an app each time you use it or the risks of things breaking in unexpected ways.

AI Agent division of labor will be critical for maximizing the impact of agents. We've long had a division of labor in organizations because having individual experts handing off tasks to each other is more effective than a bunch of generalists trying to do things a different way each time. AI Agents present the same dynamic. For AI Agents to work, you need just the right amount of context about the task that they're trying to complete. This means a deep domain understanding, set of knowledge to work off of, clear instructions, and set of tools to use. Too little context and the agent will fail. Yet, equally, as more of this information enters the context window, we know that the models can become suboptimal.  For a complex business process, if you put all of the documentation, description of the workflow, and instructions into the context window, we know that this can eventually lead to context rot, which leads to worse results. The logical architecture then in the future is to divide agents up in atomic units that map to the right types of tasks and then have these agents working together to complete their work. We're already seeing this play out effectively in coding agents. There are more and more examples emerging with people setting up subagents that all own specific parts of a codebase or service area. Each agent is responsible for a part of the code, and there is agent-friendly documentation for the code. Then as work is needed in that relevant area of the codebase, an orchestrator agent coordinates with these subagents.  We could see this pattern likely applying to almost any area of knowledge work in the future. This will allow AI Agents to be used for far more than task-specific use-cases and extend to powering entire workflows in the enterprise. Even as AI models improve to be able to handle larger context windows, and the intelligence levels go up, it’s not obvious that this architecture ever goes away. It’s likely that the role of each agent expands as capabilities improve, but clear lines of separation between subagents may always lead to better outcomes.

Initially the thought was that single AI agent would handle arbitrarily large workflows. Instead, the pattern that seems to be working is deploying subagents that have specialization by task to avoid context rot. AI agent division of labor may be the future.