For the last several years, one question has dominated the conversation about artificial intelligence and work: which jobs will AI replace?
It is an understandable question. It may also be the wrong one.
A recent working paper by Benjamin Verschuere and Angus Cameron, Hiding in the Mean: The Two Margins of AI's Employment Effect, starts from a deceptively simple problem. Occupations are made of tasks, and those tasks do not all relate to AI in the same way. Some activities are candidates for substitution. Others become more valuable when paired with AI. Many remain largely unaffected.
The authors classify every task by what it requires of a human, then measure how much of each kind sits inside an occupation. Weighted by employment, they find US work is roughly 32 percent substitution, 15 percent complement, and 54 percent inert. Collapse all of that into a single occupation-level exposure score and the opposing forces disappear into the average.
That distinction matters. Their post-2021 analysis associates greater exposure to substitutable tasks with weaker employment growth, and greater exposure to complementary tasks with stronger growth. The authors are careful about how far to push this. They grade the complement result robust in sign and describe the substitution effect as consistent with AI rather than established. That caution is worth preserving when the finding gets quoted.
But the broader idea is powerful even before the macroeconomic debate is settled:
AI does not encounter a job title. It encounters the tasks inside the job.
For companies, that changes the question from which jobs disappear? to something more practical: how should the work inside a job be redesigned?
A job title is a useful shortcut, and an increasingly lossy one
Companies still organize work through titles. Executive Assistant, Marketing Manager, SDR, Bookkeeper, Project Manager.
We need those labels. They help with compensation, career paths, accountability and communication. But they are weak specifications of what someone actually does.
Consider a company that says it needs an administrative assistant. In practice that might mean managing an inbox, coordinating a calendar, following up with vendors, maintaining a CRM, documenting SOPs, preparing weekly reports, and knowing which exceptions need to reach the founder.
Those activities sit inside one job description, but they have very different relationships with AI. Recurring reminders can often be automated. An AI system can help triage an inbox or draft a process document. But deciding that a frustrated customer's email needs immediate executive attention is not simply another inbox task. It involves context, consequence and judgment.
The title tells us where the person sits in the organization. It does not tell us how the work should be divided between software, AI and a human operator. That is increasingly the distinction that matters.
We started seeing the same problem from both sides of the talent market
At Avila we did not begin with a grand theory of workforce architecture. We began with an ordinary operating problem: how to describe client demand and talent more precisely.
On the demand side, clients often arrive using familiar role language. Discovery conversations quickly get more specific. A request for a marketing VA may actually involve email nurture, community management, list growth and content repurposing. A request for an administrative profile may turn into CRM implementation, process documentation, scheduling and recurring follow-up. A sales request can decompose into prospect research, outbound calling, qualification, objection handling, CRM updates and appointment booking.
The role title is the beginning of the conversation, not the specification.
Then we looked at the supply side and found the mirror image. Candidates carrying similar broad labels often have materially different capabilities once we inspect interviews, work history, tools, portfolios and placement evidence. In one case, a profile that could easily have read as creative or video-oriented resolved much more clearly into SEO and content operations once we reviewed newer evidence. In finance, a broad "financial assistant" category separated into very different clusters: senior bookkeeping and month-end close on one side, financial and risk analysis on another.
That matters for matching. Those people are not interchangeable merely because a taxonomy puts them in the same family.
It also changes how we read a talent shortage. Sometimes a company really does need to recruit. Sometimes the apparent shortage is an observability problem: the capability already exists in the talent pool, while the title, profile or verification system fails to reveal it.
In our own demand and supply review, several apparent gaps became less severe once we stopped searching by broad role label and started looking for evidence of the underlying tasks. Not every gap closed. Outbound sales and appointment-setting capability still looked comparatively thin when we required recent evidence of call volume, objection handling and outcomes. But "thin and under-verified" is a very different diagnosis from "we have no candidates," and it points to a different action.
From job architecture to Task Architecture
This led us to a framework we have started calling Task Architecture.
The idea is to stop treating a role as the smallest useful unit of workforce design. Instead, begin with the outcome a business needs and decompose the work required to produce it:
client outcome → task bundle → canonical task → work mode → evidence → owner → human checkpoint
For each task, we ask which of three modes best describes how the work should operate today.
Automate. A task is a strong automation candidate when it is bounded, repeatable, and can run safely end to end once its inputs, rules and controls are defined. Recurring reminders, deterministic data movement, standard report compilation, routine system updates.
Augment. AI materially increases speed or quality, but a human still reviews, contextualizes, decides or communicates. Research synthesis, first-draft content, variance analysis, meeting preparation, workflow mapping, inbox triage.
Human-own. Some work depends on judgment, trust, accountability, relationships, negotiation, taste, or consequential decisions. AI may support the person doing that work. It should not quietly become the owner of the outcome. Handling a sensitive client escalation, deciding a financial exception is material, negotiating a stakeholder tradeoff, protecting brand reputation, making a consequential hiring decision.
These are not permanent labels. A task that needs human review today may become safely automatable as tools and controls improve. The same task can belong in different modes at different companies, because risk, data quality and business context differ. That flexibility is a feature of the model, not a flaw.
The goal is not to preserve every human task
There is a tempting response to AI disruption: identify what machines cannot do and build a defensive wall around it. That misses the opportunity.
The valuable worker in an AI-enabled organization is not simply the person performing the least automatable tasks. Increasingly it is the operator who can do several things at once:
- own consequential judgment;
- use AI to increase throughput and quality;
- recognize repeatable work that should be automated;
- operate and improve the workflows around that automation;
- recognize when AI output is unreliable and needs escalation.
Complementarity is not only a property a researcher assigns to a task. It can become an operating capability in a person.
A strong executive assistant may spend less time manually coordinating routine scheduling and more time preparing context, anticipating conflicts and protecting executive attention. A marketer may produce fewer first drafts from scratch and spend more on positioning, audience judgment and performance interpretation. A financial operator may automate deterministic processing while becoming more valuable at anomalies, controls, and explaining what the numbers actually mean.
The job does not vanish. Its internal architecture changes.
This changes recruiting too
Suppose a business says it needs an SDR. The traditional workflow begins by searching for people with SDR experience. A task-oriented workflow starts one level deeper. What outcome is the company trying to produce, and what work actually produces it?
The bundle might include building prospect lists, researching accounts, writing or adapting outreach, making outbound calls, qualifying interest, handling objections, updating the CRM, booking meetings, and following up with no-shows.
Now the matching problem is clearer. Some of that work can be automated. Some can be accelerated by AI. Some, particularly live persuasion, qualification and nuanced objection handling, still needs evidence that a human can own it well.
This also gives recruiting a more useful definition of "verified." Instead of asking only whether someone is a verified SDR, we can ask which parts of the sales task bundle the person has actually demonstrated, what evidence supports those capabilities, and where human review is still required before we represent that person to a client.
That is a much higher-resolution talent market.
Workforce software will need the same shift
Most workforce systems were built around people, jobs and org structures. AI introduces another kind of actor: software that can perform, assist with, or coordinate pieces of work.
To use that responsibly, companies need a representation of work that both humans and machines can read. A future workforce system might treat a task as something closer to:
task → work mode → risk → evidence required → owner → automation candidate → human checkpoint
An agent could execute bounded Automate tasks, assist a human on Augment tasks, and recognize when a Human-own task requires escalation rather than autonomous action.
That is a different vision from "AI employee replaces human employee." It is closer to an operating system in which work is deliberately allocated among humans, AI and traditional software according to capability, context and consequence.
The job title is not going away
None of this makes job titles useless. People need professional identities. Organizations need reporting structures. Labor markets need shorthand. Careers need recognizable pathways.
But the job title may no longer be precise enough to serve as the fundamental unit for designing AI-enabled work.
The paper that prompted this makes the academic version of the point visible: opposing AI effects can vanish when tasks are aggregated into occupational averages. Our experience operating a talent marketplace suggests a practical parallel. Both client demand and worker capability get clearer when you decompose broad labels into the actual work underneath them.
That is why the future-of-work conversation should move past asking whether AI replaces humans. The more useful question is how we architect the relationship between them.
The organizations that adapt best to AI may not be the ones that replace the most people. They may be the ones that get best at deciding which work belongs to software, which work belongs to an AI-enabled human, and which decisions should never leave human hands.