The debate about AI and jobs has been running for years, mostly in the abstract. The practical question (what does a company actually do with its workforce when AI can handle 40% of what its employees do) gets far less attention.
That's the question worth answering. Not "will AI take jobs?" It already has, and will keep doing so. But: how do you actually restructure a workforce when the work changes? Who goes, who stays, what gets harder to manage, and what does a company look like on the other side?
What "AI and Workforce" Actually Means Right Now
Companies are not running experiments anymore. They're making structural decisions. Klarna went from 700 customer service agents to 150. IBM has paused hiring for 7,800 roles it expects AI to cover. Duolingo cut 10% of its contractor workforce because AI now handles content work those contractors did.
These aren't isolated moves. They're a pattern: identify which roles are primarily executing repeatable processes, remove the human cost, keep the outcome. The challenge is that this sounds cleaner on paper than it is in reality.
It's figuring out what to do with the people, the institutional knowledge they carry, and the processes that were built around their presence. Most companies get the tech right and the transition wrong.
Which Roles Go First
There's a reliable pattern to which roles get eliminated first when a company deploys AI seriously. It's not about seniority or department. It's about what percentage of the job is execution versus judgment.
Roles where 70%+ of daily tasks are rule-based, repeatable, and data-dependent are the first to go. Think accounts payable, data entry, Tier 1 support, scheduling, basic reporting, invoice processing, lead qualification. The remaining 30% either gets absorbed by a more senior person or bundled into a monitoring role.
What's less obvious: mid-layer coordination roles are next. The person whose job is to collect status updates, consolidate reports, and push work through a process is doing something AI handles without effort. Project coordinators, operations assistants, and certain HR admin roles are more exposed than most people expect.
What stays: roles that require real judgment under uncertainty, external relationship management, or creative direction. A sales director who closes seven-figure deals stays. The sales ops analyst who builds the reports they use probably doesn't. A senior finance person who structures deals stays. The analyst who pulls the data to build their models is vulnerable.
What Gets Created
This part gets less coverage, but it matters. AI deployment does create roles, just not many, and not for the same people.
The new roles tend to cluster around three areas: AI operations (someone has to monitor what the agents are actually doing), process design (someone has to decide what gets automated and how), and exception handling (the cases AI flags for human review need someone who can actually resolve them).
A company that eliminates 8 administrative roles might create 1.5 new roles in AI operations and process design. That's a net reduction of 6.5 FTEs and a meaningful shift in the skill profile of the team. The person who was processing invoices doesn't become the person managing the AI that processes invoices. Different skills, different background, often a different hire entirely.
This is why "AI will create new jobs" is technically true but practically cold comfort for the people losing existing ones. The new jobs require different skills, and the retraining path is rarely as smooth as the projections suggest.
A Scenario Worth Sitting With
A professional services firm with 85 employees deployed AI across its operations function in Q3 2025. The target was 6 roles: AP clerk, AR clerk, two ops coordinators, a data analyst, and an office administrator. The plan was straightforward: deploy over 8 weeks, keep the humans through a transition period, then let attrition handle the rest rather than execute formal redundancies.
Week 6: the AP clerk noticed the AI was miscategorizing a vendor category that had been manually adjusted 18 months earlier after a client dispute. She flagged it. The error had been running for four weeks and would have caused a material discrepancy in the quarterly close. No one had thought to document why that manual adjustment existed in the first place.
The firm's actual transition took 22 weeks, not 8. Not because the AI didn't work (it did). Because institutional knowledge lives in people's heads, not in process documentation. Every time they thought they were done, someone who was being transitioned out mentioned something they'd been doing quietly that no one had mapped.
They reached the same outcome (from 6 roles to 2) but the transition cost twice what was budgeted and left a CFO who now audits every AI output manually for anything touching a client account. That last part wasn't in the plan.
The Workforce Math Companies Don't Run
Most workforce decisions get made on salary. That's the wrong number. The fully-loaded cost of an employee (including benefits, employer taxes, equipment, real estate, HR overhead, and management time) typically runs 1.5x to 2.2x base salary. A $65K ops coordinator costs the business $100K-$140K per year, every year, reliably. We break this down in detail in our article on the true cost of an employee.
AI agents handling the same scope cost a fraction of that. An autonomous agent operating at the level of a mid-tier operations role runs $6K-$18K per year in ongoing costs once deployed. At Leverwork, our pricing is $15K-$25K for setup and $5K-$10K/month for ongoing management and optimization. That's $75K-$145K in year one to permanently eliminate roles that were costing $100K-$300K per year.
The math is not subtle. What slows companies down is not the economics. It's the organizational discomfort of running the math out loud and acting on it.
What the numbers look like in practice
Managing the Transition Without Making It Worse
The companies that manage AI workforce transitions well share a few common approaches. They're not groundbreaking, but they're consistently ignored.
Map processes before you touch headcount. Every role slated for elimination should have its actual work documented before the person in that role knows they're being replaced. Not the job description. The actual work. What decisions do they make? What exceptions do they handle? What do they know that isn't written down anywhere? You need this before they have any reason to withhold it.
Run parallel operations longer than you think you need to. The instinct is to move fast once the AI is working. Resist it. Run the AI and the human simultaneously for at least 6 weeks before relying on the AI alone. The errors you catch in that window are worth significantly more than the salary savings you'd gain by cutting earlier.
Decide on legal structure before you decide on people. In the UK, redundancies above 20 employees in a 90-day window trigger collective consultation requirements under the Trade Union and Labour Relations (Consolidation) Act. In Germany, works councils have co-determination rights over changes to working conditions. In California, WARN Act obligations kick in at 50+ employees. The legal structure of the transition is not an afterthought. It determines the timeline. For more on the regulatory trends shaping this, our article on the future of work in 2026 covers what's worth knowing.
Separate the decision from the announcement. Once the decision is made, information controls collapse quickly. Treat the implementation plan as need-to-know until you're ready to act. Not for cynical reasons. Because premature disclosure creates uncertainty that degrades the performance of every person in the affected group, including people you intend to keep.
The Roles That Are Harder to Replace Than They Look
Some categories appear automatable but aren't, or at least aren't yet automatable in ways that don't introduce new risk.
Compliance-adjacent roles are a good example. An employee whose job involves regulatory interpretation, not just data handling, carries risk context that's hard to encode. An AI can process the data; it can't absorb responsibility for a judgment call that turns out to be wrong. Some industries require a licensed human to sign off on certain outputs. Automating the production of those outputs doesn't eliminate the human. It changes their job to reviewer and signatory.
Customer relationships at the enterprise level are another. Not customer support (that automates cleanly). The strategic account manager who knows when a client is about to churn three months before it shows in the data, because she notices the tone of their emails has changed, is doing something genuinely hard to replace.
The honest answer is that almost every role has some component that resists automation. The question is whether that component justifies the full cost of the headcount. Usually it doesn't. The exception-handling that requires human judgment in an otherwise automatable role can typically be absorbed by a senior person in a few hours per week, not a dedicated FTE.
What the Workforce Looks Like After
Companies that have completed serious AI workforce transformation tend to end up with smaller, more senior teams. The junior execution layer shrinks significantly. The mid-layer coordination layer shrinks. The senior judgment layer stays roughly the same size and gains capacity because they're no longer managing people doing repetitive work.
The org chart gets flatter. There's less to manage in the people sense, more to manage in the systems sense. The skills that become more important: process design, AI operations, exception handling, and the senior judgment that AI can't replicate. The skills that become less important: data entry, report generation, status tracking, basic analysis.
For a company at $10M-$30M revenue running 50-100 employees, a realistic post-transition state might be 30-60 employees with 15-20 autonomous agent roles handling what the eliminated FTEs did. The agent layer costs roughly $100K-$200K per year in ongoing management. The eliminated headcount was costing $1M-$2M. Our analysis of hiring vs. AI automation runs the full comparison if you want to stress-test the math against your own numbers.
When to Start This Process
The companies waiting for "the right time" are mostly waiting for permission. There's no external signal that makes this cleaner or less disruptive. The disruption is inherent to the change. What changes with delay is how competitive your position is when you get there.
The practical trigger is when your fully-loaded headcount costs are growing faster than your revenue, or when you can identify 3+ roles in your org whose primary work is execution of repeatable processes. At that point, the transition is economically obvious. The only question is how long you want to carry the cost while you decide.
For context on how to think about managing an AI-driven workforce once you've made the shift, that's a different set of challenges. Worth reading before you start the transition, not after.
Find out which roles in your company are ready to go
We'll map your current workforce against what AI can handle today. You'll get a clear picture of which roles are candidates for elimination, what the transition looks like, and what it costs versus saves over 24 months.