How AI Is Changing White-Collar Work: Customer Service & Administrative Work

Editorial Note: This article examines emerging patterns in how AI is reshaping customer service and administrative work. The examples used are composite scenarios based on common workplace patterns, and the three-wave framework represents an analytical perspective rather than a universal timeline.

From Click to Cognition: How AI Is Quietly Rewriting Two White-Collar Jobs (And What to Do About It)

This isn’t another piece debating whether AI will replace humans. That debate dominated much of the early AI conversation, but a more practical question has emerged: how are everyday workflows actually changing?

Let’s cut straight to it: walk into two of the most ordinary white-collar workstations—a customer service desk and an administrative assistant’s cubicle—and look at what the screen actually shows today. Compare that to two years ago. The differences are subtle in some places, glaring in others. And those differences are worth examining, not because they’re dramatic, but because they’re already becoming part of everyday work.

This also isn’t a survival manual. There’s no ten-step framework here for becoming “irreplaceable.” What follows is simply an account of how these workflows are changing, broken down into observable pieces. What anyone does with that information is their own call.

The following examples are composite scenarios based on common patterns found in customer service and administrative workflows.

Part 1: The Pre-AI Reality – Two Scenes That Feel Eerily Familiar

Let’s rewind to two scenes from before 2022. Anyone who’s spent a few years in an office will recognize them immediately.

Scene One: Customer Service Rep – Five Windows, Three Scripts, One Countdown

8:47 a.m. Sarah logs in. She works for a mid-sized retail bank—think Discover or a regional credit union. Her screen is split across five windows: a VoIP phone client, a live chat panel, an email queue, an internal ticketing system, and an SLA dashboard that refreshes every few seconds with current wait times, average handle time, and a running tally that might as well read “how many bathroom breaks you have left today.”

The phone rings first. A cardholder at an airport rental counter just got declined and is already past frustration. Sarah walks the standard de-escalation script while pulling up the account—plenty of credit, clean payment history. She switches to the merchant portal and finds the denial originated from the rental company’s POS terminal. That diagnosis takes about four minutes. During that time, three new chat messages pop up: two are straightforward billing inquiries (copy-paste standard response, resolved), one is a disputed transaction that needs a ticket routed to the back office.

After the call ends, she has ninety seconds to complete the summary: customer name, last four digits of the account, issue category, resolution status, and whether follow-up is needed. Every field is manual. Meanwhile, the next call is already lighting up the queue.

Sarah’s actual shift is eight hours. She spends maybe half of that actually talking to customers. The rest goes to hunting for information, filling forms, switching systems, pasting script templates, and periodically refreshing the knowledge base to see if any policy changed overnight.

Her core competency isn’t banking knowledge alone. It is the ability to manage multiple information streams, coordinate responses, and avoid errors while moving between systems. The job often places heavy demands on attention management, information switching, and short-term coordination rather than complex decision-making.

Scene Two: Administrative Assistant – Excel, Calendars, and the Inbox That Never Empties

2:15 p.m. Mark works in administrative support at a mid-sized SaaS company in San Francisco—about eight hundred employees. He supports three VPs and manages six calendars between them, because every executive has a primary calendar, a shared calendar, and a “private one that nobody else sees.”

His main task this afternoon: coordinate a Q3 planning meeting across four departments. Twelve attendees spread across time zones from Pacific to Central European. He scans the email thread and finds three people have already sent their availability—but with inconsistent formats. One says “Wednesday afternoon,” another says “August 14 at 2 p.m.,” and a third just dropped a placeholder meeting on their calendar without confirming anything.

He builds a time matrix manually, screenshots each person’s calendar view (the company’s permissions are inconsistent—some people’s busy/free details are hidden), identifies two common windows, and fires off a Doodle poll. While waiting for responses, he tackles reimbursements: forty-seven receipts this week from the three VPs’ business travel, client lunches, and a team offsite. Each photo needs to be renamed to “date-purpose-amount” and entered into Concur, line by line, while checking every charge against the company’s T&E policy.

One receipt stands out: a $200 dinner. Policy caps dinner at $75 for domestic travel—unless the meal was with a client. He needs to confirm who was at that dinner. That means digging through old emails.

At the end of the day, Mark’s inbox still shows thirty-four unread messages. He flags twelve as “tomorrow” because they involve policy interpretation or fuzzy questions that need HR clarification—things he can’t decide on his own.

Mark’s real value isn’t Excel proficiency. It’s knowing which decisions he can make and which ones require stopping to ask. That boundary isn’t written in any employee handbook.

What both scenes share: Sarah and Mark’s job descriptions say “communication,” “coordination,” and “problem-solving.” But their actual working hours are consumed by information shuffling, format conversion, system switching, and rule-checking. The common thread across all these activities: clear boundaries, breakable steps, and limited judgment requirements.

And that’s exactly where AI started.

Part 2: The Three Waves – How AI Ate These Jobs, Bite by Bite

If we slice the period from 2022 to today into three phases, each shows a different depth of AI intervention.

Wave One (2022–2023): The Tool Phase

AI arrived as a co-pilot. Customer service systems embedded real-time script suggestions—when a customer said “I want to file a complaint,” the AI popped three response templates and the rep picked one to paste. Email clients added auto-summary plugins that compressed a 300-word customer message into three bullet points. Calendar tools started suggesting meeting times automatically.

The changes were concrete: Sarah’s post-call summary time dropped from ninety seconds to forty, because the system auto-populated key fields from the conversation. Mark’s inbox gained automatic labels—newsletters and internal announcements got shunted to “read later,” and frequent contacts floated to the top.

But the job structure didn’t change. The AI was still a passive tool—it only moved when someone clicked a button. Most practitioners at the time described it as “a better search box.” Useful, but not transformative.

Wave Two (2024–2025): The Agent Phase

This one was qualitative. In some organizations, AI systems moved beyond suggestions and began executing parts of end-to-end workflows.

In customer service, standardized workflows increasingly became candidates for automation. Bill inquiries, password resets, address changes, standard cancellations—these types of requests could be handled without human involvement in some systems. A customer typed a request into the chat window, and the AI could interact with connected backend systems to complete the operation. Human involvement was still required when the system encountered uncertainty, lacked permission, or when the customer explicitly requested a person.

Mark’s world shifted too. Reimbursements increasingly worked like this: he dragged receipt photos into the system, and the AI automatically extracted amounts, dates, merchant categories, and pre-filled the forms—while simultaneously checking each line against policy. That $200 client dinner? The AI flagged it as “requires attendee confirmation” and pulled up the relevant email threads and meeting notes from that date, attached to the review queue. Mark’s job became reviewing and approving. Scheduling turned into: he typed “Q3 planning, 12 attendees, cross-timezone,” and the AI returned three options within ten minutes, complete with each attendee’s historical attendance patterns and timezone preferences.

The core shift: the execution mantle moved. Sarah and Mark went from doing to reviewing and approving. Workload dropped, but decision-making—lighter as it became—moved up a notch.

Wave Three (2025 to present): The Integration Phase

By this stage, AI-first approaches became increasingly common among organizations actively adopting AI workflows. New hires in these environments may be trained primarily on AI-enabled processes rather than only traditional CRM or ERP navigation. Internal process documentation increasingly began categorizing workflows by “AI-automated” versus “human-required.”

Organizational structures changed in some AI-forward companies. Entry-level customer service roles began evolving—instead of only taking calls, some new hires started by auditing AI-handled cases to catch misclassifications. Some companies reduced pure-entry-level hiring, not necessarily because of layoffs, but because the volume of purely manual work was shrinking.

In administrative support, hiring patterns also shifted in some organizations. Pure-execution assistant roles declined in certain environments, while new responsibilities emerged around workflow coordination. These people monitor AI-run processes, identify friction points where human escalation keeps firing, and work with business units to improve processes: either change the policy, or retune the AI.

The signal from this wave: AI is no longer simply an add-on in many AI-forward organizations. It is increasingly becoming part of the workflow design itself. Humans are moving from assembly-line operators toward line monitors and optimizers.

Part 3: The Real Question – If AI Executes, What Do Humans Bring?

Looking across all three waves, a clear pattern emerges: many tasks with clear boundaries, breakable steps, and narrow judgment requirements are becoming candidates for automation. The process has not been sudden, but the direction of change is becoming increasingly visible.

This forces a question that’s worth sitting with—if AI takes more of that work, where does human value actually reside?

Let’s examine this carefully: in well-bounded scenarios, AI systems can already outperform humans on speed and consistency, especially when workflows are clearly defined. This isn’t where “human advantage” lives. The terrain worth examining is the opposite: ambiguous boundaries, incomplete information, and situations where the problem itself needs defining.

In customer service, those scenarios look like this.

First: emotional attribution. A customer says “you guys are terrible.” AI can detect negative sentiment and switch to a soothing script. But it still struggles to determine what the customer is actually upset about—sometimes the customer themselves can’t articulate it. Is it the service, or a bad day, or just needing to feel heard? Sarah once handled a call where a customer ranted for twenty minutes before revealing the issue wasn’t even with the bank—he’d just had a fight with his spouse and needed somewhere to vent. An AI system might classify that as a “service quality” complaint without fully understanding the broader context.

Second: gray-zone cross-functional coordination. A problem touches customer service, billing, and logistics. Each department says “not my problem.” AI can generate tickets and CC everyone on an email, but it still struggles with persuading reluctant humans to cooperate. The person who resolves this knows who to talk to, what that person cares about, and what communication style works.

Third: emotional anchoring. Some customers develop trust with a specific agent during one unusual interaction—maybe the agent went beyond protocol, or asked one extra question at the right moment. That trust carries forward. Customers start asking for “the person who helped me last time.” These moments remain difficult to fully standardize.

Administrative support has its own survival zones.

First: intent interpretation. An executive says, “Can you get me ready for next week’s board meeting?” That sentence mutates depending on context. Does “ready” mean compiling existing data, or building a forecast model? Is this a draft or a final? Who’s in the room? Mark knows which interpretation applies, because he’s been with the company for four years and understands each person’s habits and influence. An AI system, even with access to historical emails, may still struggle to fully understand what this specific person means at this specific moment.

Second: edge decisions outside policy. Company policies don’t cover every situation. A policy-violating expense report sometimes should be rejected, sometimes approved as an exception—the difference lies in the requester’s context, the project’s priority, and the month’s remaining budget. The administrative assistant’s value isn’t policy recitation; it’s judging whether to flex this time. An AI system granted authority here might follow rules strictly. But when the rules themselves are ambiguous, human judgment remains important.

Third: invisible organizational lubrication. Mark knows which VPs have a history and avoids seating them next to each other. He knows who gets irritable after 3 p.m. and schedules important emails for the morning. He knows that while VP A and VP B don’t get along, their assistants are friends—and that channel can carry informal signals. Much of this knowledge is difficult to capture, formalize, and transfer into AI systems.

A discomforting observation: if the bulk of someone’s daily work falls into the “clear-boundary, rule-governed” bucket, that work is increasingly becoming a candidate for automation—it just hasn't reached every workflow yet. That’s not a threat. It’s an early signal.

Part 4: Cognitive Upgrade – Redefining Your Role and Negotiating with AI

At this point, the conversation needs to move from “how to keep a job” to “how to redefine what the job is.”

Drop the question “What can I do that AI can’t?” for a moment. That framing assumes humans and AI are competing for the same slice of the same pie—which may not hold. A more useful question: Given that AI can handle 80%, what should I do with the remaining 20%?

Three Ways to Redefine the Role

First: move from executor to boundary-setter. The core job is no longer completing tasks—it’s specifying which tasks go to AI, which stay with humans, and what conditions trigger a handoff. This requires understanding the business logic deeply enough to know each step’s tolerance for error, and where mistakes are cheap versus where human oversight is non-negotiable. Not a technical skill. A judgment skill.

Second: move from information-processor to judgment-calibrator. AI produces an answer. The human’s job is to review it, question it, and adjust it for the specific context. The value lies in knowing whether this answer fits this situation. That requires broader contextual awareness—company politics, customer history, industry norms—to make effective calls on AI-generated output.

Third: move from service-provider to relationship-operator. Whether with external customers or internal executives, what’s ultimately being paid for is trust. Trust comes from consistency of expectation—people know what to expect from this person, and don’t need to re-establish it every time. AI can support many interactions, but human relationships still depend heavily on context, trust, and personal connection. So this value remains important.

Negotiating Task Allocation with AI: A Practical Approach

This negotiation isn’t with an IT admin. It’s with one’s own weekly time allocation.

One exercise worth trying: take fifteen minutes each week to list every task done in the past seven days, and mark each one by “degree of rule-clarity” and “degree of judgment required.” Tasks high on rule-clarity and low on judgment should be handed to AI—whether or not the company has formally deployed it. Many standardized workflows can already be assisted or automated with existing tools. The purpose isn’t saving time—it’s building the habit of proactively peeling off automatable work.

Then take the time freed up and do one thing: work that nobody asked for, but that makes the whole system run smoother. A customer service rep can use that time to analyze recurring complaint patterns over the past month and send a “what customers aren’t saying” brief to product. An admin can map the company’s informal decision network—who gets pulled into which decisions, and what information needs to reach whom in advance. These have no KPIs. But doing them changes how the role is perceived.

Performance metrics need proactive renegotiation, too. Traditional customer service metrics—volume handled, average response time—lose their discriminating power once AI handles many standardized tasks. New metrics should reflect human-specific contribution: first-contact resolution rate on escalated cases, satisfaction lift after human intervention. Proposing these shifts is, effectively, rewriting the value coordinates of the job.

The Mindset Shift

One cognitive distinction is worth holding onto: competing with AI on speed is unwinnable, and the gap will likely continue to widen in many standardized tasks. Competing on depth of understanding—understanding the business, the people, the particularities of a specific situation—remains a different kind of advantage.

AI can read ten thousand emails. It may not understand what happened between the twelve people in that room yesterday. Humans often do. That kind of contextual knowledge is difficult to capture, formalize, and transfer into AI systems. This is an information-asymmetry advantage, not simply a technical skills advantage.

Another angle: treat AI as a tool that fractures work. A job used to contain twenty different activities, fifteen of which were information-shuffling and format-shifting, five of which required judgment. AI strips out many of the repetitive activities. The remaining judgment-heavy work becomes more visible and potentially more important. That’s actually focusing human work, not diluting it.

Of course, this transition won’t be comfortable for everyone. Role consolidation and hiring freezes are real in some industries. Transitional uncertainty is real. Nobody can guarantee a smooth move for every practitioner. But one thing is clear: the direction of change pushes humans away from some repetitive execution tasks and toward positions that demand more judgment. That position comes with higher requirements—and possibly higher returns, for those willing to redefine what they do.

Conclusion

Go back to the two opening scenes.

Sarah juggling five screens, Mark’s inbox never reaching zero—those images are gradually receding from some parts of the workplace. What’s changing isn’t necessarily the job itself. It’s the part of the job that was “human execution of standardized processes.” That shift will continue, although the pace will vary across industries and organizations.

AI isn’t simply taking anyone’s job. It’s redrawing the boundary of task ownership—clear, automatable things increasingly move toward machines; ambiguous, judgment-heavy things continue to require human involvement. The act of drawing that line matters more than where the line ends up.

Those who land on the other side of the boundary will find a new interface: no clear instructions, no historical data to lean on, just a vague requirement and a room full of people waiting for problems to be defined. That space isn’t comfortable. But it remains one of the areas where human judgment matters most.

References

Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044
Noy, S., & Zhang, W. (2023). Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. Science, 381(6654), 187–192. https://doi.org/10.1126/science.adh2586
Bankins, S., et al. (2024). A multilevel review of artificial intelligence in organizations: Implications for organizational behavior research and practice. Journal of Organizational Behavior. https://doi.org/10.1002/job.2735

Author Note: This article explores how AI may reshape human work by changing the boundary between execution and judgment, rather than simply replacing human roles.