Meta AI Executive Says Agents Could Reshape Entry-Level Work

Clara Shih’s view of AI and work changed when she watched agents compress a product development process at Meta. Work that had involved researchers, designers, product managers and several kinds of engineers could, she said, move from an idea to a tested prototype with one or two people. The experience led her to reconsider who companies need to hire—and to leave her role building business AI products to focus on workers facing the transition.

Shih now remains a senior advisor to Meta while leading the New Work Foundation, a nonprofit, and Dear CC, a consumer initiative for early-career workers. Their free offerings include a podcast where hiring managers discuss what they seek, a tool called Field Report that estimates AI exposure across majors and occupations, and Game Plan, which connects rejected applicants with peers and a mentor.

Her account complicates the familiar assurance that automation will simply take routine tasks off workers’ hands. Shih once expected customer support staff to use AI for repetitive work and spend more time on complex problems and customer relationships. Looking back, she says that outcome has happened to some extent, but “primarily not been true.” That admission carries weight because she helped build products intended to automate customer service.

The moment agents changed her expectations

Shih described the turning point as a practical demonstration, not a forecast. A conventional product process moved through user research, surveys and interviews, mockups, requirements documents, and handoffs to front-end, back-end and machine learning engineers. At Meta, she saw agents and coding tools help reduce those steps to a smaller group working together.

One or two people could develop an idea, generate a prototype through what’s commonly called vibe coding, test it with real users and simulated users, then bring a leaner team into the work of building it for production. Shih said the agents matched and sometimes surpassed strong employees on real tasks. The speed of that change made her imagine the same pattern spreading beyond product development.

She saw possible applications in marketing, distribution and privacy review. Those workflows still involved human experts checking the final output, she said, but tasks that had taken ten steps and ten days could be compressed into minutes. Her concern was less about whether this made a company more productive than about what the shift would mean for people whose jobs were changing alongside it.

That concern became personal when Shih began removing entry-level job postings from her team. She said AI made it possible to move faster under pressure, and that she no longer felt she needed some of those roles. Her decision illustrates a difficult feature of workplace automation: an employer may adopt tools to accelerate delivery without setting out to replace a specific worker, yet the staffing consequences can follow directly from the new process.

Shih’s career gave her a view across several stages of software development and deployment. After early roles at Google and Salesforce, she founded Hearsay Systems and led it for a decade. She later returned to Salesforce to run Service Cloud, became CEO of Salesforce AI in 2023 and led the launch of its agent platform, Agentforce. In 2024, she moved to Meta to build and lead its business AI group, whose agents answer customer messages on WhatsApp, Messenger and Instagram.

Her experience is one account of how agents can alter work inside a company; it doesn’t establish how quickly the same changes will spread across industries. Agents still have limits, and a separate report cited in the original interview described underperforming systems as one factor in the failure of a planned large-scale staffing reduction at Meta. The contrast matters: a workflow that looks dramatically faster in a controlled product task may still stumble when deployed across a company.

Three ways automation can change a job

Shih uses a three-part framework based on research by MIT economist David Autor and his lab. The outcome depends on how much of a job’s work is automated and which tasks are affected. A job title alone, she argues, doesn’t tell workers whether AI will assist them, substitute for their work or change who can enter the field.

The first outcome is direct substitution. When AI takes over a significant share of the tasks that define a role, fewer people may be needed to do that work. Shih points to translation and customer service as areas where this could happen. She doesn’t predict that every position will vanish; rather, enough work may be displaced that a field becomes a less attractive place to enter or remain.

The second outcome is the optimistic case: AI handles routine bottlenecks and lets people focus on harder work. Shih cites code snippets and test case generation as tasks that can be automated while leaving software engineers more time for other responsibilities. If lower costs or faster work increase demand for the service, the field may expand—a dynamic often described as the Jevons paradox.

She expects this pattern to benefit some software engineers, radiologists and workers in other roles. The key condition is that automation removes tasks that constrain a professional’s work without removing the expertise that gives the role value. If that happens, employees can use their training on decisions and responsibilities that tools don’t handle as well.

The third outcome is less comfortable. AI can make a job easier to enter by automating some of the expert knowledge previously required, increasing the supply of workers even as the market grows. Shih compares this with taxi driving after GPS directions and rideshare platforms became widely available. More people could take up driving, and demand for rides grew, but she argues that in places such as New York and London, driver earnings have fallen below a living wage.

That example separates job creation from job quality. A larger number of available roles doesn’t guarantee that workers earn enough or have stable work. Even when automation brings in new customers or increases demand, the gains can be uneven if the number of people able to perform the work rises faster than demand.

For employers and workers, the three cases point to different questions. Does a tool remove a small bottleneck, or a large share of the work? Does it leave a worker more time for tasks that require judgment, or lower the expertise required to enter the field? And if the work expands, do workers share in the resulting value? Shih’s framework doesn’t answer those questions for every occupation, but it gives teams a more useful way to assess exposure than simply asking whether AI can perform a task.

Customer service offers a test of the optimistic case

Shih had helped build AI systems aimed at customer service, including Agentforce at Salesforce and Meta’s Business AI agents. These products can answer customer messages and are intended to handle support work for companies. The stated promise was that workers would use AI to take care of routine requests, leaving more time for complex problem solving and relationship building.

Shih says she once believed that optimistic account herself. In her description, the idea was present while these products were being built, but the team was primarily excited about making them work. The hoped-for shift toward higher-value tasks has happened “a little,” she said, but it “primarily” hasn’t. Her comment challenges the assumption that a tool’s ability to automate a task automatically translates into better work for the people who performed it.

The gap can arise because companies may adopt automation to handle more interactions with fewer staff, rather than to preserve the same workforce for newly expanded responsibilities. The interview doesn’t quantify how often that happens or establish a single cause. It does show why evaluations of workplace AI need to examine staffing and job design, not just whether a system can answer a customer.

Shih also points to market pressure. At a company trying to deliver quickly, a tool that can compress a workflow has an immediate appeal. Once teams see that result, they look for other bottlenecks to remove. The same logic can carry into multiple departments, multiplying the effect beyond the team where a tool first proved useful.

Why she expects roles built around handoffs to face pressure

Shih estimates that about one in five roles inside companies involves preparing an artifact for another employee to review. That might mean drafting a brief or slide deck, or preparing an order form for a salesperson to deliver to a customer. She expects these positions to face particular pressure as AI becomes better at producing the inputs that other employees use.

In her view, the person ultimately accountable for an outcome—and often in direct contact with a customer or candidate—may find it quicker to use AI than to coordinate across several people producing and passing along materials. She offers salespeople, recruiters and senior legal staff as examples of employees who deal with outside parties. The vulnerable point is the chain of handoffs: if a tool can prepare a useful draft quickly, the company may need fewer people to create intermediate work.

That prediction doesn’t mean every role involving documents disappears. It suggests that organizations may reassess how many people they need for work that feeds into someone else’s decision. The distinction matters for early-career workers, who often begin by supporting more experienced colleagues and preparing the material they use.

Shih sees a related challenge for junior employees. Effective use of AI requires enough knowledge to provide context, judge what a system returns and improve the result. Without that grounding, an inexperienced employee can simply pass a request to an AI tool and forward its answer without adding much value. Employers may then decide that the role is unnecessary, even though those jobs have traditionally helped people gain experience.

Dear CC is intended to help young workers build that familiarity in a field. Shih describes the needed capabilities as practical knowledge of how to prompt and configure agent systems, paired with the ability to evaluate and refine their output. The goal is not just to get a tool to generate something; it’s to know whether the result is accurate and useful.

There is a tension here. Employers may want candidates who already have enough expertise to supervise AI, while entry-level positions are one of the ways people acquire that expertise. Shih’s proposal puts weight on projects and practice as a bridge: workers should learn to use AI in a specific area and gain experience reviewing its work. Whether that approach can replace the training opportunities provided by early jobs remains an open question.

Software skills may spread into other professions

Shih is less pessimistic about software engineering than some forecasts suggest. She argues that computer science teaches ways of thinking—breaking problems into algorithms, building reusable modules and evaluating systems—that are well suited to working with agents. As companies adopt agents, she expects demand for people who can set them up and judge their outputs to grow.

Her more provocative prediction is that software engineers may move into work now associated with legal, marketing and accounting teams. That’s a personal forecast, not a measured projection. Its logic is that employees with the technical background to assemble agent workflows could take on tasks in other fields, especially when they can combine software skills with enough subject knowledge to check the work.

The same argument also points to a limitation: configuring an agent isn’t a substitute for knowing what a good result looks like. Shih’s advice to students emphasizes deep expertise alongside technical skills. For professionals, the practical question is how to combine domain knowledge with the ability to direct and assess AI systems, rather than treating tool fluency as a complete qualification.

In this view, demand for software engineers could rise even as AI automates some coding tasks. That outcome depends on the tools enabling more software to be built and on engineers being able to take on broader work. It also depends on whether companies value technical skills enough to move those workers into professions with distinct standards and responsibilities.

What Dear CC offers job seekers

Shih says young graduates are the first group Dear CC is trying to support. Many followed the path they expected would lead to a suitable job, sometimes taking on college debt, and now struggle to find work in the field they studied. Her first aim is to make the forces changing work easier to understand, so people can make decisions with a clearer picture of what’s happening.

But she doesn’t think an AI mentor can provide everything a person needs. When the natural response to a shortage of human support is to build another agent, Shih argues, it can miss what job seekers are asking for: conversations with people, trust and a sense of connection. Dear CC’s mentoring model pairs applicants with peers and a mentor, aiming to make a difficult search less isolating and to add accountability as people work through practical steps.

Those steps include revising a résumé and LinkedIn profile, deciding where AI can help improve an application, avoiding generic material that reads like spam, and learning how to contact a hiring manager. Shih also describes preparing candidates to reach a person when employers use AI screening. The emphasis is on applying AI selectively while still building the human connections that matter in a job search.

The organization’s survey found that after six months without work, respondents began questioning their self-worth and identity. Shih says this can hit especially hard for people who have been high achievers and suddenly find themselves doing gig work while facing pressure from parents who may not understand changes in the labor market.

Peer groups can offer support as well as practical advice, she argues. Applicants who are navigating similar challenges can encourage one another and hold each other to a plan. The model takes a different view of what technology can provide: AI may help with applications and career research, while human relationships help with motivation and belonging.

Shih founded the New Work Foundation with a broader aim than supporting one group of job seekers. She says no organization has a reliable forecast of exactly how the transition will unfold, which is why she favors experimentation and humility. Her work is free to users, and its tools focus on explaining AI exposure, offering hiring advice and connecting people to mentors.

Work, income and a sense of purpose

Shih argues that support for people affected by AI will need to come from employers, schools, individuals and public institutions. Her foundation can’t manage the transition on its own. She says employers may need to change how they train and hire, while schools and individuals will also have a part to play.

She disagrees with Andrew Yang, a founding advisor to the organization, about whether universal basic income would address the growing gap by itself. Shih says people first need a living wage, but also autonomy, mastery and purpose. For much of history, she argues, paid work has supplied those things; if work changes substantially, support has to consider more than income.

She points to a volunteer project in Arlington Heights, Illinois, as one example of how communities might connect people around practical needs. The group, Connections to Care, matches volunteers—including recent retirees—with older residents who need rides to appointments or help getting groceries. Similar tasks can be purchased through delivery and ride services, Shih observes, but the volunteer connection also creates relationships and a shared sense of purpose.

Her example suggests one possible response if AI changes the supply of paid work: organize people to meet needs that markets don’t address well, while creating meaningful connections. Shih mentions local schools, nursing homes and neighbors as places where unmet needs may exist. She presents community service as an idea to explore, not a ready-made answer for workers whose jobs change.

The proposal raises a practical challenge of its own. Community roles may provide belonging and useful activity, but the interview doesn’t show how they could replace earnings or career progression. Shih’s argument is that a living wage and a sense of purpose both matter; any response focused on only one could leave an important need unmet.

The disruption may reach workers as they gain experience

Shih and interviewer Casey Newton describe current signs of strain among some early-career workers and people in occupations with high exposure to AI. The interview doesn’t present a society-wide jobs crisis. Instead, it focuses on groups where the effects may appear first and on the possibility that those effects will spread as agent capabilities improve.

Shih expects the pressure to extend beyond graduates one or two years out of school to people with three or four years of experience. That moving target makes it harder to prepare for a single job market. Skills that help someone get hired now may not be enough to keep their work valuable as systems change.

Her recommendation is to cultivate a habit of continuous learning and an entrepreneurial approach to work. She advises young people to prepare for volatility both financially and psychologically, rather than assuming they will join an organization and progress steadily for a decade. The point isn’t that stable careers will disappear, but that workers should be ready to adapt as roles change.

Newton describes the available evidence as mixed: aggregate employment figures still look broadly stable, while some studies point to early pressure in exposed and entry-level roles. Shih agrees with that framing and says the trend data she has seen is moving in a concerning direction. The interview doesn’t attach a numerical forecast to what comes next.

That uncertainty is central to Shih’s message. She doesn’t claim to know how many jobs will be affected or what a durable response should look like. Instead, she connects what she saw inside Meta with the choices employers are already making and with the experiences of job seekers who need help now.

Advice for students entering exposed fields

For a college student considering law school, Shih recommends three steps. First, study something that holds your interest instead of picking a major only because a parent prefers it or because it pays well today. She argues that deep expertise is valuable, and sustained interest can help someone build it.

Second, learn serious AI skills. She distinguishes that from introductory familiarity with chatbots, emphasizing agent systems, context, tool calls and evaluation. Her suggestion is to build and release something that solves a real problem. A student might create an agent to help someone contest a parking ticket, for example; a useful test is whether someone outside the builder’s family chooses to use it.

Third, invest in people skills and relationships. Shih expects the next few years to be uneven, and says a support network matters alongside financial preparation. She compares the potential disruption with manufacturing shifts in the 1980s and 1990s: changes in employment can be difficult for affected workers even when broad economic measures improve.

The three recommendations reinforce one another. Subject expertise gives a worker a basis for judging an agent’s output. Building a real tool offers hands-on practice. Relationships can provide advice, opportunities and support when an individual’s plans need to change.

Shih’s story is a reminder that AI’s effect on work may show up first in decisions made inside teams: fewer handoffs, smaller project groups and fewer openings for people still building experience. The tools may help a company move faster, but workers’ outcomes depend on how employers redesign roles and whether new routes into skilled work emerge. For people entering the labor market, the most concrete preparation she offers is to develop deep knowledge, learn to evaluate agents and build relationships that can help them adapt.