AI is already changing what people do at work. Tasks like drafting contract summaries, writing basic code, tagging support tickets, and reconciling accounts can now be partly or fully automated.
That doesn’t always mean a job disappears. More often, the job itself changes. The problem is that most workforce plans track roles and headcount, not the tasks inside them.
This article covers how to identify which tasks AI can take on, spot roles likely to change, plan realistic reskilling paths, and make those decisions before AI is deployed.
What Is Actually Changing in the Work Right Now
McKinsey reports that AI adoption roughly doubled since 2017 and surged again with generative AI in 2025, with the fastest uptake in marketing, product development, and customer service.
The World Economic Forum puts a number on the downstream effect: 44% of workers’ skills are disrupted within five years, and a quarter of jobs change significantly. It also looks at what is on the rise.
Those are big numbers, and they’re easy to nod at without changing anything. So look at what happened inside one function.
GitHub’s research found developers using Copilot completed tasks 55% faster and reported higher satisfaction.
Nobody’s developer headcount dropped because of it. What dropped was the share of the week spent writing scaffolding, boilerplate, and syntax anyone could look up, and what expanded was review, debugging, and deciding what to build. Same title. Different jobs.
Similar story in service teams that handed off call summaries and CRM entries. Also in finance and legal, where contract scanning and reconciliation now run first, and a person reviews second.
This also applies to exposure clusters where work is structured and data-rich. Healthcare, manufacturing, and logistics are shifting too, though mostly around the edges of the role: scheduling, quality inspection, diagnostic support, with a person still physically present and accountable.
None of this is unprecedented. ATMs didn’t end branch banking, spreadsheets didn’t end accounting, e-discovery didn’t end litigation.
Each one reassigned the daily work and changed what a strong performer looked like. What’s different is the clock. Shifts that used to unfold over a decade are landing inside two or three years, which is shorter than the cycle most companies use to hire, train, and promote.
Start With Tasks, Not Job Titles
A title is a container for maybe fifteen recurring tasks. AI doesn’t take the container.
It takes tasks, and it takes a specific kind first: repetitive, rules-based, language-heavy, high-volume analytical.
Which means two people with identical titles in different departments can have completely different exposure, and any analysis that starts from titles will average them together and tell you nothing useful about either.
Decompose the role before you score it
List what the role actually does, then estimate the share of a week each item takes.
The O*NET database works as a scaffold if your internal job architecture is thin, but the time estimates have to come from people doing the job.
Ask three of them separately. Where their answers disagree, you’ve usually found either an undocumented process or a manager who doesn’t know how the work gets done.
Ten to fifteen tasks. More than that and the exercise stalls at role number four.
The five questions that predict exposure
Score the task, never the role:
- Structure: Predictable, rules-based, documented well enough that a new hire could follow it?
- Data: Is there clean digital data behind it, or does it depend on context living in somebody’s head?
- Quality tolerance: Is a reviewed first draft fine, or does the output need to be right the first time?
- Risk: Legal, financial, or safety consequences that require a named human on the decision?
- Human touch: How much rests on empathy, physical dexterity, or trust built over years?
McKinsey Global Institute’s work on automation potential is useful for calibration when internal scores start drifting, which they do.
Teams that just bought a tool score everything as automatable. Teams that fear the tool score nothing.
Run the scenario before you need it
Take a role where 20 to 40% of task time comes back as AI-susceptible, and model it out. Where does the recovered time go? Is there higher-value work sitting there unclaimed, or does the role honestly get smaller?
If it gets smaller: how many people, in which locations, over what period?
Do that for six roles, and now you have a number, a timeline, and a list of names. Which is uncomfortable, and also the only version of this that lets you start training people eighteen months early instead of announcing a restructure with two week’s notice.
Then map adjacency. For each affected role, chart two or three nearby roles someone could realistically reach with six to nine months of learning, given the skills they already have.
Read the Signals in the Labor Data
Some of it is already public. The Bureau of Labor Statistics projects word processors and typists declining 38% between 2022 and 2032, and data entry keyers declining 24%, citing automation and improved office technology.
Those roles are the leading edge rather than the boundary line. What they share is a task list that’s almost entirely structured and language-based.
As traditional career paths become less predictable, some workers are also looking at investing and trading as ways to diversify how they build income and wealth.
Gregor Emmian, Deputy Chief Digital Growth Officer at Rise, works with a platform that gives users access to crypto, stocks, and other markets.
He notes, “As traditional roles evolve, more people are thinking beyond a single paycheck when it comes to their financial future. Freelancing and gig work are part of that, but so is growing interest in trading crypto, stocks, and other markets. Easier access to these tools means workers have more ways to explore additional income and build financial resilience alongside their careers.”
Demand moves in the other direction too. AI product management, model operations, data stewardship, evaluation, human-in-the-loop quality review, and prompt design are embedded inside existing functions.
The scarce person is the one who knows your claims process, or your supply chain, or your underwriting rules, and can also work fluently with the tools. That combination is far easier to grow from your existing bench than to recruit.
Build the Plan Around Skills, Then Fund It
Workforce planning here is a standing process. Not a model somebody rebuilds each October.
Tie tasks to skills, and skills to moves
Once the tasks are mapped, attach the underlying skills to each one.
This changes what a manager can say.
Instead of “your role is evolving,” which everyone correctly hears as a warning, it becomes: four of your twelve tasks are automating over the next year, three of the skills you already use transfer directly into this adjacent role, here’s the training that covers the gap, and here’s the time you’re getting to do it.
While you’re in there, rewrite the job descriptions.
Small task, and it gets skipped constantly. As the work shifts, JDs should show the new responsibility mix and name the specific tools and oversight expectations attached to the role. Skills-first descriptions give current employees a visible target and give candidates an accurate picture of what they’re signing up for.
Pilot small, measure honestly, then write the playbook
Track two numbers on every pilot: time saved and error rate. Both. Always.
A tool that saves an ops team four hours a week and introduces a rework loop nobody logs is not a win, and a team under pressure to demonstrate AI adoption will report the first number enthusiastically and never mention the second.
Ask specifically about the new work the tool created. Reviewing output nobody requested. Correcting confident mistakes that didn’t exist before. Explaining to a customer why the summary was wrong.
When a pilot holds up under that scrutiny, write it up as something another team can copy, including what broke.
Reskilling fails predictably, so name it in advance: the program gets announced, no time gets protected, people are expected to learn on top of a full workload, and participation collapses by week six. Learning time either sits on calendars and gets defended by managers when a deadline hits, or the whole thing is decorative.
External pathways fill gaps faster than building everything internally. Community colleges, bootcamps, and vendors will co-design a course against a specific documented gap. Apprenticeship.gov is a workable starting point for earn-and-learn structures with defined outcomes.
Completing that training should feel like progress, not just another line in an HR system. Recognising new qualifications and professional milestones gives employees something tangible to associate with their development, whether that means sharing a credential internally or displaying it in custom frames.
What HR and Leadership Have to Own
HR sits across job architecture, pay equity, learning pathways, and internal communication, which makes it the only function positioned to connect this work. Left to individual departments, you get four disconnected pilots and no plan.
Leadership’s part is mostly signal, and it’s short. Say plainly why AI matters for the business and for people’s careers. Fund learning with protected hours rather than encouragement.
Reward teams that redesign a process, not only teams that hit an output number, because process redesign loses output metrics every single time unless someone deliberately weights it. Use the tools yourself, oversight steps included.
AT&T’s multi-year reskilling program is worth reading about in detail. Internal career pathways, online learning, tuition support, role redesign, all aimed at moving tens of thousands of employees into in-demand roles rather than hiring externally and letting the existing workforce erode. The interesting output wasn’t the retraining. It was a repeatable system for handling the next shift, and the one after that.
Decide the Ethics Before Deployment, Not After
Automating a set of tasks is a decision about people’s income. It should carry the process weight that it implies.
Before anything ships:
- Run an impact assessment covering reskilling, redeployment, and severance where the role genuinely contracts
- Keep human review on high-stakes decisions, assigned to a named owner instead of a shared inbox.
- Test for bias and define the escalation path before something surfaces, not during
- Bring workers into design and rollout, including unions or employee councils where they exist
Healthcare makes the limits of automation especially clear. Even digitally delivered services such as TRT online, where clinicians evaluate hormone levels, determine whether treatment is appropriate, and prescribe based on the individual patient, still depend on qualified human oversight rather than automated decision-making alone.
Regulation is converging on roughly the same ground.
NIST has published an AI Risk Management Framework covering trustworthy design and use. The EU’s AI Act sets a risk-based structure with stricter obligations on high-risk systems, and several employment applications fall inside that category.
Exposure also isn’t evenly spread. Brookings research indicates automation risk falls harder on workers without a college degree and, in certain occupations and regions, on Black and Hispanic workers.
Aggregate numbers will look manageable while the impact concentrates in specific teams and specific zip codes. Targeted training, wage support during transition, and real internal access to new openings are the mechanisms that address it. Stating a commitment to fairness is not.
What to Do This Week
Pick one role. List its top ten tasks. Tag each as human-led, AI-augmented, or AI-automatable, and note roughly what share of a week it eats.
Then ask the people in that role two questions: where has AI already cleared out busywork, and where has it created work that didn’t exist before. The second answer never appears in vendor reporting.
And as those roles change, the job descriptions need to change with them.
Ongig helps employers create clearer, more inclusive job descriptions that reflect the skills and responsibilities they actually need, giving both current employees and future candidates a more accurate picture of where the work is going.
