The primary narrative around AI has been labor arbitrage. That AI tools are getting good enough to do work that used to require people, so functions will shrink, headcount will come down, and likely the next few quarters will be defined by who cuts fastest. We've seen this play out too in many organizations, but this is an incomplete picture.
The better picture, and I dare say the more complete one, is about capacity. AI isn't most powerful when it replaces what people already do. It's most powerful when it lets people do things they couldn't do before or augment them to do it faster and/or better. This distinction is critical. It changes almost everything about how an organization should be thinking about AI.
A lot of "AI replaces X" arguments share a common move. They reduce a discipline to its most visible output, point out that the output is now cheap, and conclude the discipline is finished. Design becomes screens. Writing becomes paragraphs. Analysis becomes charts. Engineering becomes code. Anthropic recently released Claude Design, and the immediate reaction was that product designers are next on the chopping block!
From Fixed UI to Fluid UI
For decades, software was built around what you might call Fixed UI. Static dashboards. Rigid forms. Pre-set reports. The designer's job was to anticipate every reasonable user need in advance and impose a structure that worked across all of them. Some users got exactly what they needed. Most got something close enough.
In an AI world, that's flipping. We're moving from Fixed UI to Fluid UI. Interfaces that adapt to what the user actually needs at that moment. We saw this firsthand while building our Policy Auditor for compliance teams in financial services. The feedback we got was almost impossible to address through traditional product roadmap items, because every user wanted the output shaped slightly differently. An advisor needs a quick summary. A compliance officer needs a liability check. A COO wants a workflow handoff.
In the old world, that would've been a backlog, plus a serious discussion around which ICP we need to optimize the most for. In the AI world, it's a single system that interprets intent and shapes output accordingly. That doesn't make design less important. It makes it harder.
The designer's job is no longer to impose structure on the user. It is to build systems fluid enough to interpret intent and respond to it well. The discipline is more valuable, not less. It just looks different from a Figma file.
This is true beyond design. The places where careful thinking is needed are multiplying, not contracting. Every system that talks to a person, every agent that hands off to another agent, every tool that exposes capabilities to a model; all of it needs someone deciding what to expose, in what order, with what affordances, what fails gracefully, what requires a human in the loop. The surface area of decisions worth making well has grown, not shrunk. AI allows you to expand that capacity for your team and replacing them would be the sub-optimal move.
Grunt Work versus Judgment
I've written before about vibe coding. It's the ability to describe what you want in plain English and have an AI assistant produce working code. It's real. It works. The barriers to building custom software have collapsed in a way that genuinely matters.
But I've also learned the hard way where it breaks down. Vibe coding gets you 80% of the way there. Core logic. A basic interface. Something you can demo. The remaining 20% — polish, testing, deployment, integrations, error recovery, audit trails, compliance is where real context engineering discipline lives. This is the right shape for thinking about AI generally, not just coding.
A seasoned professional with AI is operating at a scope that wasn't possible before. Decisions that took a week now take an hour. Variants that took a month now take a day. Their judgment now reaches further than it ever could on its own. A less experienced person with AI is also more productive, but produces output of less reliable quality, because their expertise and judgment is still building. The gap between the two has actually widened. AI lowered the floor a little and raised the ceiling a lot.
The most valuable thing in any organization right now is the judgment its people already carry. Decades of accumulated taste, pattern recognition, context, and decision-making instinct.
The operational question is: how do you encode judgment?
Most organizations have judgment trapped in places where it can't scale. In a senior person's head, in a side letter, in a procedure document no one updates, in the email thread where a decision actually got made? The work that follows from that judgment gets done manually, by people, every time. It's not that the work is hard. It's that the judgment hasn't been put into a form a system can use, and hence the belief that the human is integral to the process.
Something interesting happens when you try to fix this. A COO of an $8B firm we work with put it well: deploying AI actually forces you to figure out your own process. When you have to describe a workflow in plain language so a platform can build the agent, you end up articulating what actually happens today. Who does what? In what order? What are the exceptions? What does success or failure really mean? What's the feedback loop? Where is human intervention not only the most critical, but also the most appropriate? That description becomes your documentation, often for the first time.
Firms that try to deploy AI before they understand their own processes usually don't get far. The ones that treat it as a forcing function for process clarity end up with both the automation and the operating model written down. The encoded judgment then runs continuously, leaves an evidence trail, and gets refined as the firm learns. The person isn't replaced. Their judgment is more valuable, it comes at the right moment, and their capacity multiplies.
What We're Building
This is the thinking behind what we do at PitCrew. We work with wealth managers and the broader financial services ecosystem. These are places where judgment is especially dense and especially hard to scale. Reconciliation logic that lives in a senior ops person's head. Fee exceptions buried in side letters. Policy interpretations that show up in audit responses but aren't written anywhere. Drift between what a firm says it does and what its systems actually show.
Our platform encodes that workflow and judgment into purpose-built agents. A Cash Availability Checker that flags shortfall risks before invoice debits go to the custodian. A marketing content auditor that doesn't sample, but reviews ALL content against regulation (both internal and external rules) before it publishes. Pre-built Skills cover the patterns common across financial services. Our Agent Factory lets firms build their own Skills against their own rules, in plain language, with full evidence trails.
Production-grade reliability with audit-ready outputs. Documentation that holds up to scrutiny.
What firms get isn't fewer people. It's the same people, freed to do the work that requires their judgment and expertise. Agents handle the laborious grunt work that requires consistent execution against rules the firm has already decided. You now have capacity within the firm to grow.
This is a moment to build, to expand scope, to take on the work that wasn't possible before. The people and organizations who treat AI as leverage will look back on this period as the one where they got a great deal more ambitious. The ones who treat it as a cost lever will spend the same period producing more of what they were already producing, and wondering why it stopped feeling distinctive.
The story of this decade isn't going to be written by what got replaced. It'll be written by what got added.
This article was originally published on LinkedIn. Read it on LinkedIn →