There’s a particular kind of misery that anyone who has worked in an office knows well. It’s 4:47 PM on a Thursday. You have a meeting in thirteen minutes. And somehow, between now and then, you need to pull last month’s sales numbers from a spreadsheet, cross-reference them against a CRM report, paste the relevant rows into a slide deck, format everything so it doesn’t look like a ransom note, and email a summary to six people with different opinions about what “brief” means.
You are not thinking. You are not creating. You are not doing anything that required your degree, your experience, or frankly your consciousness. You are a very expensive, very stressed human copy-paste machine.
This is the dirty secret that no one talks about when they discuss the future of work: an enormous chunk of what knowledge workers actually do every day is shuffling information from one digital container to another. Moving data between systems. Reformatting things. Filling in fields. Generating documents that follow the same template as the last forty documents. Sending follow-up emails that say, in seventeen different ways, “just checking in.”
We built the most powerful information-processing tools in human history, then spent thirty years using them to manually recreate the administrative busywork that existed before computers. The irony is almost too much.
But something is shifting, and it’s shifting faster than most people outside the AI industry have clocked. We are entering a period where the manual digital workflow, as a category of human labor, is beginning to genuinely collapse. Not slowly. Not hypothetically. Right now, in real companies, with real tools, the layer of human effort required to move information through digital systems is thinning in ways that would have seemed implausible even three years ago.
This isn’t a story about robots taking jobs. It’s more interesting than that. It’s a story about what happens when the scaffolding of modern office work, the endless clicking and copying and formatting and routing, starts to disappear. What do we do with the time? What do we lose? And perhaps most pressingly: are we actually ready for the version of work that comes next?
Before we can talk about what’s ending, we need to be honest about what we built and why it persisted for so long.
The modern knowledge worker’s day is, in large part, a workflow tax. This is the invisible toll you pay to keep information moving through the systems your organization depends on. Every time you export a CSV from one tool and import it into another. Every time you update the same information in three places because your systems don’t talk to each other. Every time you generate a weekly status report by pulling numbers from six dashboards and manually compiling them into a document that someone will read for ninety seconds.
The workflow tax is invisible because it’s so normalized. Organizations don’t measure it. Employees don’t complain about it directly because it feels like “just part of the job.” But researchers who’ve actually tried to quantify it have found numbers that should make any executive flinch.
A 2023 study by Asana found that workers spend, on average, roughly 60% of their time on what the researchers called “work about work” — status updates, searching for information, switching between apps, attending meetings that exist primarily to synchronize information that, in a better-designed system, wouldn’t need synchronizing. McKinsey’s research around the same period found that employees spend nearly two hours a day just on email management. Not responding to substantive emails. Just managing the inbox.
Part of why this happened is structural. Most organizations didn’t design their digital infrastructure, they accumulated it. A CRM here, a project management tool there, a data warehouse that someone’s brother-in-law recommended in 2019, a Slack workspace, a shared Google Drive with folders inside folders inside folders named things like “FINAL_v3_ACTUAL_FINAL.”
These systems were each, individually, defensible choices. But together they created what you might call the three-system problem: any piece of information your organization cares about probably lives in at least three different places, and keeping those places synchronized requires ongoing human effort. Someone has to be the connective tissue. Usually, that someone is you.
The tragedy isn’t that we built these systems. It’s that we hired human beings to be the glue between them. We took people with expertise, judgment, and creative capacity, and we assigned them the job of being organic middleware.
Now, to be fair, this wasn’t for lack of trying. The automation industry has been promising to fix this problem for the better part of two decades. Tools like Zapier, Make (formerly Integromat), and UiPath’s RPA (robotic process automation) suite have genuine enterprise adoption. If-this-then-that logic has been available to non-programmers for years. So why didn’t it fix the workflow tax?
A few reasons. First, traditional automation is brittle. It works beautifully when inputs are perfectly consistent and the process never changes. The moment someone renames a spreadsheet column, or a vendor updates their API, or the email format coming from your invoicing software shifts slightly, the automation breaks. And then someone has to fix the automation. Which is, itself, manual work.
Second, traditional automation can’t handle ambiguity. It can move a clearly labeled invoice from an inbox to a folder. It cannot read an email from a client that says “hey, can you update that thing we discussed?” and figure out what needs to happen next. Humans are extraordinarily good at operating in ambiguous, context-dependent situations. Software wasn’t, until very recently.
Third, building and maintaining these automations required a level of technical skill that put them out of reach for most of the people who actually needed them. The people drowning in manual workflows weren’t, generally speaking, the people with the engineering time to build Zapier chains or configure RPA bots.
So the workflow tax persisted. It became an accepted feature of office life, like bad coffee and passive-aggressive reply-all emails.
Here’s where I need to be careful, because this is exactly the kind of topic where it’s easy to veer into hype. The AI industry has a long and not entirely honorable history of promising to automate human work and then quietly failing to deliver on the specifics. So let’s be precise about what has actually changed, and why this time the dynamics are genuinely different.
The critical shift is not that AI has gotten smarter in some abstract sense. It’s that AI has gotten competent at the specific cluster of skills that manual digital workflows require: reading unstructured text and extracting structured meaning from it, operating software interfaces that were designed for humans, making reasonable inferences when instructions are incomplete, and maintaining context across multi-step tasks.
These four capabilities, individually, existed in weaker forms before. What’s new is having them bundled together in systems that can be deployed without a PhD in machine learning.
Think about what a huge percentage of modern knowledge work happens in a web browser. Email (Gmail, Outlook Web). Documents (Google Docs, Notion, Confluence). Project management (Asana, Linear, Jira). CRM (Salesforce, HubSpot). Analytics (Looker, Tableau). Finance (QuickBooks, Expensify). Communication (Slack, Teams, both of which now live partially in browser tabs).
For most of the automation era, software systems could only talk to each other through formal integration channels, APIs, webhooks, structured data exports. But a browser-based AI agent doesn’t need an API. It can look at the screen the way a human looks at the screen and operate the same controls a human would operate.
This is a bigger deal than it sounds. It means that every piece of software with a web interface, including all the legacy enterprise systems that were never designed with automation in mind, suddenly becomes automatable without requiring cooperation from the vendor. The agent doesn’t need Salesforce to build an integration. It just needs to be able to see Salesforce and click things.
Google’s Project Mariner, Anthropic’s computer use capabilities, and a new generation of agent frameworks are all building toward this same vision: an AI that can operate your digital work environment the way a skilled contractor would operate your kitchen. It doesn’t need custom-built tools. It works with what’s there.
The other shift that matters is moving from rules-based automation to reasoning-based automation.
Here’s a concrete example. Old-school automation: “When an email arrives with the subject line containing ‘Invoice’ from a sender in this approved list, extract the attached PDF, upload it to this folder, and log the sender and timestamp in this spreadsheet.” This works until the subject line says “FW: RE: Invoice follow-up question” or the PDF is inline rather than attached.
Reasoning-based automation: “Here are my invoicing responsibilities. Handle incoming invoice emails.” The system reads the email, understands what it’s asking regardless of formatting quirks, figures out the appropriate response, and executes it. When it’s uncertain, it asks. When it encounters something genuinely novel, it flags it for human review rather than silently failing.
The difference isn’t just technical. It changes the relationship between humans and automated systems. Instead of writing a precise specification that the machine executes exactly, you describe an intent and the machine figures out how to fulfill it. That’s a fundamentally different kind of collaboration, and it’s one that opens up automation to tasks that were previously too variable, too ambiguous, or too dependent on context to automate with rules.
OpenAI’s Operator, released in early 2025, is a real-world example of this in practice. Users can instruct it to handle tasks like booking reservations, filling out forms, or navigating multi-step web processes, and it reasons through the steps rather than following a predetermined script. Early adoption data suggests completion rates on complex multi-step tasks that would have been impossible for RPA systems two years ago.
Let’s talk about agents, because this is the concept that ties everything together, and also the concept that’s been most thoroughly muddied by marketing departments.
An AI agent, in the meaningful sense, is a system that can pursue a goal across multiple steps, using multiple tools, over an extended period, without requiring a human to supervise every action. It’s not a chatbot that answers questions. It’s something closer to a junior colleague who you can hand a task to and trust to figure out the intermediate steps.
The word “agent” has been so thoroughly overused in the past eighteen months that it’s worth anchoring it to something concrete. When Klarna announced in early 2024 that its AI assistant was handling the equivalent workload of 700 customer service agents, doing two-thirds of all customer service chats in its first month of deployment, that was an agentic system in operation. It wasn’t answering predetermined questions. It was reading customer issues, pulling account data, making decisions about resolutions, and executing those resolutions across multiple systems.
It’s one thing to describe agentic AI in the abstract. It’s another to look at what’s actually happening in specific industries right now, because the reality is both more mundane and more consequential than the breathless press releases suggest.
The mundane part: most of the early wins aren’t dramatic. They’re not AI systems replacing entire departments or autonomously making high-stakes decisions. They’re AI systems doing the stuff that nobody wanted to do, reliably, quickly, and without complaining about it. The consequential part: when you add up all that stuff, it’s a significant fraction of what many organizations pay people to do.
Law firms are an instructive case study because they have a very clear, very expensive version of the workflow tax problem. A substantial portion of junior lawyer and paralegal time goes to document review, contract analysis, due diligence research, and the production of first drafts of documents that follow established templates.
This work is not intellectually trivial. It requires understanding legal language, recognizing relevant precedents, flagging anomalies, and maintaining accuracy under the threat of serious consequences if you get it wrong. For decades, it was assumed that this kind of work required trained human judgment and couldn’t be automated.
What’s happening now is that firms like Allen & Overy, through their Harvey AI partnership, and Linklaters, through Luminance, have deployed AI systems that handle substantial portions of this work. Not as a curiosity or a pilot program, but as operational infrastructure. Harvey processes thousands of legal documents, drafts contract clauses, conducts jurisdiction-specific research, and flags issues for attorney review. The attorney still makes the final call. But the time they spend getting to that decision point has compressed dramatically.
The practical effect: a first-year associate who might spend forty hours reviewing a set of contracts for a due diligence process can now review the AI’s analysis of those contracts in a fraction of that time, focusing their attention on the edge cases and judgment calls that genuinely need a trained human brain. The workflow tax on that process has dropped by, depending on who you ask and how you measure it, somewhere between 60% and 80%.
What nobody quite predicted was what lawyers did with that recovered time. Some firms expected productivity to increase in a clean, linear way: same lawyers, more work, more revenue. What actually happened in several documented cases is that the nature of the work shifted. Lawyers started spending more time on client strategy, on the interpretive and advisory functions that clients actually value most, because those were the tasks that had previously been crowded out by the sheer volume of document processing.
This is a pattern worth holding onto, because it shows up across industries: when you remove the workflow tax, you don’t just get efficiency. You sometimes get a qualitative upgrade in the work itself.
Finance is another sector where the early evidence is striking, partly because the workflows are so clearly defined and the value of accuracy so high.
Month-end close processes are, in most organizations, a special kind of organizational hell. Teams spend days reconciling accounts, chasing down discrepancies, generating reports, and consolidating data from systems that were never designed to talk to each other. The work is repetitive, error-prone precisely because it’s repetitive, and consumes significant hours from people who are, theoretically, employed to provide financial insight rather than to manually reconcile spreadsheets.
Companies like Workiva and BlackLine have been building automation into their financial close workflows for years, but the new generation of AI-augmented tools is doing something qualitatively different. Instead of automating specific, predefined reconciliation steps, these systems can analyze variance, identify anomalies, suggest explanations based on historical patterns, and draft narrative commentary for financial statements. They’re handling the cognitive work of interpretation, not just the mechanical work of calculation.
JP Morgan’s COiN platform, which the bank has discussed publicly, analyzes commercial loan agreements and extracts relevant data points in seconds that previously took lawyers and loan officers roughly 360,000 hours annually to review. That’s not a typo. 360,000 hours, reduced to seconds.
Marketing teams live in a particularly dense thicket of manual workflows. Campaign setup across multiple channels. Ad copy variations for A/B testing. Performance report compilation. CRM segment updates. Email sequence management. Social scheduling. SEO brief generation. Each of these is its own workflow, touching multiple tools, requiring human coordination at every handoff.
The AI impact here has been both faster and messier than in legal or finance. Faster because marketing content is lower stakes than legal contracts or financial reports, so experimentation happened earlier and more aggressively. Messier because the quality questions are harder to answer: Is this email copy good? Depends who you ask.
What’s emerged is something like a new operational model for marketing teams. The humans define strategy, set creative direction, make brand judgment calls, and focus on the work that requires genuine intuition about audiences and culture. The AI handles execution: spinning up the variations, drafting the briefs, populating the templates, scheduling and reporting, and managing the connective tissue between tools.
HubSpot’s AI features, Jasper’s campaign workflows, and the newer generation of tools like Clay for prospecting automation have collectively shifted the ratio of strategic to mechanical work for marketing operations teams in ways that are measurable in headcount. Smaller teams are running larger campaigns with more personalization than larger teams could manage two years ago.
Here’s where most technology articles take a sharp right turn into either utopian optimism or dystopian handwringing. I want to try to stay on the road instead.
The honest version of this conversation requires holding two things simultaneously: that the automation of manual workflows genuinely liberates people from tedious, unrewarding labor, and that the transition creates real disruption, real anxiety, and real displacement that shouldn’t be minimized.
When people who study meaningful work, researchers in the vein of Mihaly Csikszentmihalyi or Cal Newport, describe what humans find genuinely satisfying about their professional lives, manual digital workflows are conspicuously absent from the list. Nobody reaches the end of a forty-year career and reflects fondly on the time they spent copying data between spreadsheets. Nobody lists “managed high email volume” as a peak experience.
What people do find meaningful: solving genuinely hard problems, creating things that didn’t exist before, building relationships, making decisions that matter, teaching others, and exercising expertise in ways that require real judgment. Roughly speaking, all the things that the workflow tax crowds out.
There’s real evidence that removing administrative burden improves both performance and wellbeing. A Stanford study of remote call center workers found that giving them tools to handle administrative tasks faster led to measurable improvements in job satisfaction. Microsoft’s research into Copilot adoption found that users who relied on AI assistance for routine tasks reported spending significantly more time on the work they described as their most important.
This is genuinely good. We should say so clearly and not bury it in qualifications.
The liberation argument, taken too far, becomes a way of avoiding an uncomfortable conversation about what happens to the people whose jobs are primarily composed of the tasks being automated.
Not everyone in a modern organization is being held back from meaningful work by the workflow tax. For some people, particularly those in administrative, coordination, and operational roles, the workflow is the job. Data entry. Document processing. Report compilation. Invoice handling. These aren’t tasks that someone does in between doing their real work. For a significant segment of the workforce, these tasks are the whole job description.
The International Monetary Fund estimated in 2024 that roughly 40% of global jobs have significant exposure to AI automation, with the proportion higher in developed economies and among white-collar roles. That’s not a precise prediction of displacement, and the IMF was careful to distinguish between exposure and displacement, but it signals the scale of the question.
The pattern from previous waves of automation gives us some guidance, though not as much comfort as optimists would like. Automation historically eliminates certain categories of tasks and creates demand for new ones, but the transition is often painful, geographically uneven, and slower to resolve than the theoretical models suggest. The coal miners who were told their jobs would be replaced by clean energy and service sector growth weren’t wrong to notice that the promised alternative jobs didn’t materialize in their towns on a useful timeline.
There’s a subtler version of the disruption problem that I think gets less attention than it deserves: what happens to skill development when the entry-level tasks disappear?
In many professions, the junior work, the document review, the first draft writing, the data compilation, isn’t just boring work that needs doing. It’s also how people learn. A junior lawyer who spends two years in document review is building a mental model of how commercial contracts are structured, what normal looks like, what anomalies look like, and how deal terms vary across industries. This is tacit knowledge that’s hard to transfer any other way.
If AI handles document review, where does that learning happen? If AI drafts the first version of the financial analysis, how does the analyst develop the instinct for what looks right? If AI manages the customer service queue, how do the next generation of service leaders build their understanding of customer psychology?
This isn’t an argument for preserving tedious work for its own sake. But it is an argument for being thoughtful about how professional development models adapt to an environment where the traditional apprenticeship path, which ran through entry-level execution work, is being shortened or eliminated.
Some organizations are already grappling with this. Law firms, in particular, have started having explicit conversations about how junior associates develop professional judgment in an AI-augmented practice environment. The answers are not obvious, and the firms that figure it out first will have a meaningful talent advantage.
All of the above, the agentic systems, the sector transformations, the workforce implications, rests on an infrastructure layer that most discussions of AI automation gloss over. And that’s a problem, because the infrastructure layer is where most of the implementation actually breaks down.
AI agents that operate across your digital workflows are only as good as the data those workflows contain. And the data that lives inside most organizations’ systems is, to put it charitably, a mess.
Duplicate customer records. Inconsistent field formats across systems. Historical data that was entered by humans who had their own idiosyncratic interpretations of what a field meant. Integrations that broke silently six months ago and have been recording nulls ever since. Spreadsheets that someone has been “maintaining” by overwriting the formulas with hard-coded values because it was faster.
When you send a human worker into this environment, they bring context, intuition, and the ability to make reasonable guesses about what corrupted data probably meant to say. When you send an AI agent into this environment, you get highly confident errors at scale. Garbage in, garbage out, but now the garbage comes out faster and in more places.
Organizations that are genuinely succeeding with AI workflow automation are, almost without exception, also investing heavily in data quality infrastructure. They’re standardizing how data is recorded. They’re building validation layers. They’re doing the boring, unglamorous work of cleaning up the systems before pointing AI at them.
This is not the part of the AI story that makes for exciting conference keynotes. But it’s the part that determines whether any of this actually works in practice.
Alongside data quality, there’s a governance question that most organizations are handling by simply not handling it. When an AI agent takes an action on behalf of your company, who is responsible for that action? When it sends an email to a customer, updates a record, approves a workflow step, or generates a document that gets sent externally, the accountability chain gets murky in ways that traditional software automation never created.
Traditional software does exactly what it’s programmed to do. If it does something wrong, someone programmed it wrong, and you can trace the error back to a decision. AI agents exercise judgment. When that judgment is wrong, the failure mode looks different and the accountability is harder to assign.
Regulated industries, finance, healthcare, law, are navigating this with varying degrees of success. Healthcare organizations deploying AI for administrative workflows have to balance efficiency gains against HIPAA obligations, liability questions, and the very high cost of errors. Financial institutions using AI for customer-facing processes have regulatory obligations around explainability and auditability that current AI systems struggle to meet cleanly.
The organizations that are getting this right are building what you might call human-in-the-loop architectures: AI handles execution and first-pass judgment, but defined categories of decisions require human review before action is taken. This isn’t a perfect solution, and it adds some of the friction back that automation was supposed to remove. But it’s honest about the current state of AI reliability in ways that “just automate it” is not.
Here’s a quiet irony in the whole story of AI-driven workflow automation. One of the biggest obstacles to deploying AI agents effectively is the fragmented, poorly integrated software environment that created the workflow tax in the first place.
Agents need to be able to read data from your systems and write actions back to them. They need authentication credentials. They need to understand the data schemas. They need reliable APIs or, failing that, stable web interfaces to operate against. The messier and more fragmented your technology stack, the harder it is to deploy agents against it, and the more brittle those agents will be when they’re deployed.
There’s a real tension here for mid-market companies in particular. Enterprise organizations with resources to invest in clean infrastructure and dedicated AI engineering teams can navigate this. Very small teams using a handful of modern SaaS tools can often find off-the-shelf agent solutions that connect those specific tools. The middle ground, organizations with five to fifteen years of accumulated legacy systems, custom integrations held together with duct tape, and limited engineering capacity, faces the hardest path.
This is where I’d push back gently on some of the most optimistic timelines being floated for AI workflow automation. The technology is ready in a way it wasn’t three years ago. The organizational infrastructure, in most companies, is not. And closing that gap requires investment, time, and frankly organizational willpower that doesn’t come from anywhere on the AI vendor’s roadmap.
Most organizations approaching AI workflow automation are making a fundamental conceptual error. They’re treating it as an efficiency layer on top of existing processes. Find a manual workflow, automate it, move on to the next one. This is understandable, it’s the path of least resistance, and it does produce measurable gains. But it’s roughly equivalent to putting a faster engine in a horse-drawn carriage.
The organizations that will pull significantly ahead are the ones asking a harder question: if we didn’t have to design work around the limitations of human attention and manual execution, what would we actually build?
This is a different kind of question. It requires stepping back from the inherited architecture of how work gets done and being willing to dismantle things that have worked, not because they’re broken, but because they were designed for a set of constraints that no longer apply.
Before you can redesign a workflow, you have to understand it. And in most organizations, deeply understanding how work actually gets done is surprisingly difficult, because a large portion of institutional process knowledge lives in people’s heads, not in any documentation.
The way invoices get processed, the unwritten rules about which approvals are genuinely required versus which ones are performed out of habit, the workarounds that someone built three years ago because a system limitation that has since been fixed, the judgment calls that happen informally before anything enters an official workflow, all of this is invisible until you try to change it.
There’s a whole discipline emerging around what some are calling process intelligence: using data from existing tools to map how work actually flows through an organization, as opposed to how org charts and process documents say it flows. Tools like Celonis and Process Mining platforms from SAP and IBM have been doing versions of this for enterprise clients for years. What’s new is the combination of process intelligence with AI-driven redesign recommendations.
The most forward-thinking organizations are using this approach not to automate their existing workflows but to first audit them, then question them, then rebuild them with AI capabilities as a given rather than an afterthought.
A workflow designed with AI capabilities as a given looks different from one that was designed for human execution and then had AI grafted on.
Consider how a contract review process might be structured. The traditional human-designed version: contracts arrive, get assigned to a reviewer based on whoever is available, the reviewer works through them in whatever order makes sense to them, flags issues in a shared document, schedules a discussion with relevant stakeholders, and eventually produces a consolidated set of comments and recommendations. The handoffs are informal. The status is tracked in someone’s head or in a shared spreadsheet.
A process designed with AI as a native capability: contracts arrive and are immediately triaged by an AI system that categorizes them by type, value, and risk profile, and routes them to the appropriate human reviewer with a pre-populated analysis already complete. The human reviewer focuses specifically on the flagged anomalies and judgment-dependent clauses, applies their expertise to the things that genuinely need it, and the AI handles the compilation, formatting, and routing of the final output. Status is tracked automatically. Metrics are generated without anyone having to compile them.
The second process isn’t just faster. It’s structurally different. It generates better data. It has fewer handoff failures. It scales in a way the first one doesn’t. And it asks something different of the human involved: less stamina for repetitive review, more expertise in the moments that matter.
Given all of this, the question that most individuals and organizations are quietly anxious about is the practical one: what do you actually need to be good at, in a world where AI handles the execution layer of knowledge work?
This question has generated a lot of confident, frequently contradictory answers from people who arguably should know better. “Creativity will matter most.” “Emotional intelligence is the key.” “Learn to code.” “Don’t bother learning to code, AI will write the code.” “Become a prompt engineer.” The signal-to-noise ratio is not great.
Let me offer a more grounded framing, based on what the evidence from early AI-augmented workplaces actually suggests.
The tasks that remain stubbornly human are the ones where there is no objectively correct answer, where context changes what “right” means, and where the cost of getting it wrong is high enough that you want someone with genuine stakes in the outcome making the call.
Should we settle this contract dispute or litigate? Is this marketing campaign on-brand in a way that protects our long-term reputation, even if it performs well in the short term? Is this customer’s complaint a one-off or a signal of a systemic problem? Does this candidate have the cultural adaptability to thrive in our specific environment?
AI systems can inform all of these decisions. They can surface relevant data, flag patterns, present analogous cases from the past, and in some instances offer probabilistic assessments. But the decision itself, in all its accountability-carrying, contextually embedded, sometimes-gut-feel reality, remains human.
The practical implication: the skill of making good decisions under genuine uncertainty, including the metacognitive ability to know when you have enough information versus when you need more, is going to be worth more, not less, as AI handles more of the information-processing work that feeds into decisions.
There’s a newer skill that doesn’t have a great name yet, but is already separating high performers from average performers in AI-augmented environments. It’s the ability to effectively direct AI systems: to decompose goals into tasks, evaluate outputs critically, catch errors before they propagate, and iterate on instructions until you’re getting what you actually need.
This is not the same as prompt engineering in the narrow technical sense. It’s closer to the skill of being a good manager, knowing how to communicate what you want clearly, knowing how to evaluate whether you got it, knowing when to give more specific instructions versus more latitude, and knowing which tasks to delegate and which to keep.
People who already have these skills, who have managed teams, who have written good specifications, who have learned to brief creatives or contractors effectively, tend to adapt quickly to working with AI systems. People who have spent their careers executing work rather than directing it have a steeper learning curve.
This has interesting implications for career development advice. Learning to be a good director of work, not just a good executor of it, is going to be a more durable investment than learning any specific tool or technical skill, which will be obsolete before you’ve gotten good at it.
There’s been a persistent argument in career advice circles that the future belongs to generalists, to T-shaped professionals and multi-hyphenate creators who can move fluidly across domains. And there’s something to this, particularly in creative and entrepreneurial contexts.
But the evidence from AI-augmented workplaces is actually pushing the other direction in certain ways. When AI handles the execution layer of work, what humans add is most clearly visible when that human has deep, hard-to-replicate expertise in a specific domain. The specialist who can recognize when an AI’s analysis of a legal clause is subtly wrong, because they’ve spent fifteen years doing this, is worth more in an AI-augmented environment, not less. Their expertise is now more purely expressed rather than being diluted by the manual work that surrounded it.
Shallow generalism, knowing a little about a lot of things, is precisely the knowledge profile that large language models are very good at. If your value proposition is knowing things that are widely known, AI is a genuine threat to that. If your value proposition is knowing something specific very deeply, in a way that takes years of practice to acquire, AI is more often an amplifier.
This doesn’t mean specialization at the expense of any cross-domain fluency. The most effective people in AI-augmented environments tend to have genuine depth in at least one domain combined with enough breadth to understand how their domain connects to adjacent ones, and enough comfort with AI tools to direct them effectively. That’s not a revolutionary insight, but it’s grounded in what’s actually working.
Individual skills matter, but the bigger lever is organizational. How companies structure themselves, how they define roles, how they think about what humans are for in a workflow, these decisions will determine who captures the value that AI workflow automation creates and who simply has their cost structure disrupted without gaining the corresponding benefits.
One role that’s quietly becoming critical in forward-looking organizations doesn’t have a standardized job title yet. It shows up variously as “AI Operations Lead,” “Workflow Automation Specialist,” “Prompt Engineer” (used loosely), or just gets folded into the responsibilities of whoever was already in charge of process improvement.
What this role actually involves is something like: understanding both what AI systems can do and what the organization needs to accomplish, then designing the human-AI collaboration architectures that connect those two things. It requires technical literacy without requiring deep engineering skills, domain knowledge without being siloed in a single function, and a systems-thinking orientation that can hold both the granular detail of a specific workflow and the broader organizational context it operates in.
This is, in some ways, a new version of the business analyst role that was important in the early enterprise software era. When SAP and Oracle started displacing custom-built enterprise systems in the nineties, organizations needed people who could translate between what the software could do and what the business needed. The same dynamic is playing out now, at greater speed and with more organizational surface area.
The automation of knowledge work also has specific and underappreciated implications for management. A significant portion of what middle managers do in large organizations is information routing: gathering status updates from their teams, synthesizing them for reporting up the chain, communicating decisions back down, coordinating handoffs between people and teams, tracking progress against plans.
If AI systems handle a substantial portion of information routing, status tracking, and coordination work, the traditional justification for many management layers weakens considerably. This is not an argument for eliminating management. It’s an argument for being honest about what management should be for.
The managers who will thrive are the ones whose value was always in the harder things: building team capability, making judgment calls when data is ambiguous, navigating organizational politics, understanding what motivates specific individuals, representing their team’s interests in resource allocation decisions. These are the things that can’t be routed through an AI system.
The managers whose primary value was aggregating and routing information are in a more precarious position, and the honest thing is to name it rather than dance around it.
There’s a macro-level story running underneath all of this that business strategy conversations are slowly starting to address but haven’t quite reckoned with fully: the automation of manual workflows is going to create enormous competitive divergence between organizations, and it’s going to do it faster than most strategic planning cycles are designed to handle.
The economics are simple and a little brutal. If Company A successfully automates sixty percent of its workflow overhead and Company B doesn’t, Company A can either produce the same output at dramatically lower cost, or produce dramatically more output at the same cost, or redirect the liberated capacity toward higher-value activities. Any of these creates a durable advantage. All three simultaneously creates a gap that’s very hard for Company B to close from behind.
The implication is that this is not a technology that organizations can choose to evaluate at a leisurely pace, wait and see, let others work out the kinks. The kinks are being worked out right now. The organizations that are building operational competency with AI-augmented workflows today are developing institutional knowledge and organizational capability that will compound. The ones waiting for the technology to fully mature before engaging with it may find that when they’re finally ready to start, the competitive gap has become structural.
This is particularly true in professional services, where the output is intellectual and the cost structure is heavily weighted toward human labor. Law firms, consulting firms, accounting practices, marketing agencies: the economics of these businesses are being repriced in real time by firms that have figured out how to do more with less manual overhead. The clients are already noticing.
Step back far enough and there’s a larger question lurking behind all the specific conversations about which workflows get automated and which skills become more valuable. What kind of work do we actually want humans to do?
This sounds philosophical, but it has practical weight. The choices organizations make right now about how to deploy AI in their workflows are, collectively, choices about what human labor is for. If we automate the execution layer and keep humans in the loop for judgment and creativity and relationship-building, we’re making a statement about what we think humans are good for and what we find valuable about human involvement in work.
That’s worth being intentional about, rather than letting it be determined by default, by whatever is cheapest or most immediately efficient.
There’s an optimistic version of where this goes: a broad shift toward work that is more intellectually engaging, more genuinely creative, more oriented toward the kinds of problems that require human empathy and judgment and imagination. Work that people find more meaningful because the meaningful parts are no longer buried under an avalanche of administrative overhead.
There’s a more complicated version where the efficiency gains mostly accrue to capital rather than labor, where the people displaced from workflow-heavy roles don’t smoothly transition into the higher-value work that was supposedly waiting for them, and where the gap between those who can direct AI systems effectively and those who cannot becomes a new axis of economic inequality.
Both of these versions are plausible. Which one we get is not predetermined by the technology. It’s determined by decisions that organizations, policymakers, educators, and individuals are making right now, most of them without fully appreciating the stakes.
In offices throughout the 1970s, there were people whose entire job was to carry clipboards. Literally: to walk between departments, pick up forms, carry them to wherever they needed to go next, and track their status. When computers arrived, that specific job description was eventually eliminated. But the people who had been carrying clipboards didn’t disappear. Some retired, some moved into other roles, and the information they had been manually routing started moving through networks instead.
Nobody mourns the clipboard carrier job. But the transition had a human cost that took years to work through, and it happened unevenly, with the people least equipped to adapt often bearing the highest burden.
We are at an inflection point that’s structurally similar but orders of magnitude larger in scope. The clipboard is now a dashboard that someone manually updates every Monday morning. It’s the weekly report that takes four hours to compile. It’s the inbox that requires two hours of daily triage. It’s the form that gets filled in three systems because they don’t integrate. Enormous amounts of human effort, applied to moving information between containers.
That effort is going away. The question isn’t really whether. The question is what we build in its place, who gets to benefit from the time and capacity that’s freed up, and whether we’re honest enough about the disruption to actually help the people it affects.
The technology, for once, is the easy part. It’s the human and organizational questions that are genuinely hard. And unlike the technology, those questions don’t get answered by better models or faster chips. They get answered by choices. By people deciding what they value, what they’ll invest in, and what kind of work they want to exist in the world.
The manual digital workflow had a good run. It outlasted its usefulness by about fifteen years. What replaces it is still, usefully, up to us.
