The Roles Employers Plan for AI Agent Rollouts

Employer plans point to specialized work around agents, while the evidence still measures intent more often than completed hiring. New responsibilities cluster around deployment, evaluation, security, adoption, and value measurement.

Seventy-eight percent of leaders in Microsoft’s 2025 Work Trend Index said their organizations were considering hiring for roles focused on AI. The list included AI trainers, data specialists, security specialists, AI agent specialists, ROI analysts, and business process consultants. That figure measures plans, not completed hires.

Microsoft published the report on April 23, 2025. Edelman Data x Intelligence conducted its 20-minute online survey among 31,000 full-time or self-employed knowledge workers, including 9,037 leaders, in 31 markets from February 6 through March 24, 2025. Microsoft sells AI products, and the methodology does not identify a separate external funder, so readers should treat the findings as vendor research.

The survey still provides a useful view of the responsibilities that organizations expect to add. It also shows why a job seeker should look past a new title and ask what decisions, evidence, and operating work the employer expects that person to own.

Hiring plans show where responsibility is moving

Microsoft asked leaders which roles their organizations were considering during the next 12 to 18 months. The report published percentages for ten titles, including the six below.

The descriptions in the final column are BIJ2 AI’s interpretation of the operating work that each title would need to cover. Microsoft reported the titles and percentages, but it did not publish a standard job description for each role.

The report also asked managers a related question. Thirty-two percent of managers said they planned to hire AI agent specialists to design, develop, and optimize agents, while 28% were considering AI workforce managers to lead teams of people and agents.

Microsoft defines managers as knowledge workers who manage a team or group, but the cited page does not report the manager subgroup size. These results come from a separate group and question, so they should not be merged with the leader percentages in the table.

These results show that employers were assigning names to new responsibilities. They do not establish how many openings appeared, how long the roles lasted, or whether employers filled them.

Agent specialists own more than the build

An AI agent specialist needs to translate a business task into an operating system with defined inputs, tools, permissions, checks, and handoffs. The model call represents only one part of that system.

Microsoft describes this role as designing, developing, and optimizing agents. An employer can make that description more useful by assigning responsibility for failure cases, release criteria, monitoring, and rollback decisions. Without those duties, the specialist may ship a capable demonstration without owning its behavior in daily work.

This role can suit software engineers, automation specialists, product managers, and operations experts. A candidate for this role needs to understand both the technical limits of an agent and the work that surrounds it.

Trainers connect agents with actual practice

Microsoft reported that 32% of leaders were considering AI trainers, and 35% of managers were considering trainers who would guide employee adoption. The title can create confusion because some companies use trainer for model development, while others mean workforce education or agent improvement.

An employer should state which meaning applies. A workforce trainer may teach employees how to give context, inspect an answer, and escalate a weak result. A trainer responsible for agent behavior may turn approved examples, reviewer feedback, and policy rules into test cases or system updates.

Domain experts can enter this work without becoming model researchers. A claims specialist, accountant, nurse, lawyer, or engineer can identify realistic cases and explain why an output fails. That judgment gives a technical team the material it needs to improve the system.

Evaluation work may sit inside several titles

Microsoft’s top ten list does not include a separate AI evaluator. Yet agent specialists, trainers, security specialists, and process consultants all need evaluation evidence to perform their jobs well.

The International Labour Organization’s 2026 research brief explains why. Its authors reviewed experiments, firm data, labor platform studies, and representative surveys from Australia, Denmark, Germany, Korea, Kuwait, the United Kingdom, and the United States. They found that productivity gains varied by task and setting, while users’ ability to detect unreliable output affected results on complex work.

The ILO authors also found limited evidence of large-scale job displacement at the time of review. Worker-reported time savings had not produced clear increases in measured output, earnings, or employment across the representative evidence they examined. The ILO published the web record on June 1, 2026, while the PDF carries a May 2026 date.

These findings make evaluation part of operating work rather than a final inspection. Someone must choose representative cases, define acceptable results, record failure patterns, and check whether performance changes after a model or workflow update.

Organizations may place that work under quality assurance, risk, operations, or a domain team instead of using the evaluator title. People seeking AI evaluation work should search for the responsibility as well as the label.

Security specialists define the agent’s reach

Microsoft found that 31% of leaders were considering an AI security specialist. Agents raise familiar security questions in a new operating context because they can use tools and take actions instead of only returning text.

A useful security role defines which data an agent may read, which systems it may change, and which actions require approval. The role also needs an incident path when an agent exposes information, follows a malicious instruction, or performs an unauthorized action.

Employers may divide this work among identity, application security, privacy, and governance teams. The title matters less than a clear record of who can limit access and stop the system.

ROI analysts test whether local gains reach the business

Microsoft reported that 29% of leaders were considering AI ROI analysts. The title reflects a measurement gap that many agent projects encounter after a promising pilot.

The ILO review found that people could report time savings without producing a corresponding change in measured output or earnings. An ROI analyst therefore needs a baseline, a defined outcome, and the full cost of operation. Review time, repair work, and the cost of failures belong in that calculation.

An analyst should also separate a measured result from a forecast. A pilot can estimate how a workflow changed for a defined group, while a business case projects what may happen after adoption expands. The second calculation depends on assumptions about use, quality, and demand.

This work could suit economists, accountants, financial analysts, and operations researchers who can design a comparison and trace a result into business records. Familiarity with an AI tool helps, but measurement discipline carries more value than a collection of prompts.

Process consultants redraw the handoffs

Microsoft’s list included AI business process consultant, with 28% of leaders considering the role. The title points to work that starts before a team selects a model.

A process consultant maps the current task, the information it uses, and the decisions that move it forward. The consultant then identifies where an agent can act, where a person must review, and what should happen when the system cannot continue.

The ILO review concluded that generative AI was beginning to reorganize bundles of related tasks instead of simply removing whole occupations. That finding supports a cautious interpretation of the new titles. Organizations may create a specialist role for the transition, then distribute parts of the work into operations, product, risk, or training teams.

Broad forecasts cannot validate exact titles

The World Economic Forum’s Future of Jobs Report 2025 offers a wider employer view. Its dataset contains 1,043 unique company responses representing more than 14.1 million workers across 55 economies and 22 industry clusters. The Forum collected survey responses from May through September 2024 and published the report on January 7, 2025, as a forecast for 2025 through 2030.

Two-thirds of participating employers said they planned to hire people with specific AI skills, and half expected to reorient their businesses in response to AI. The report also placed AI and machine learning specialists among the fastest-growing roles by percentage.

Those statements capture employer expectations rather than observed future employment. The Forum identifies survey partners and data partners, including ADP, Coursera, Indeed, and LinkedIn, but the public report page does not name a separate funder. Readers should also avoid applying a global forecast to a particular country or occupation without local evidence.

Taken together, the Microsoft and Forum surveys show planned demand for AI skills and clearer ownership around agent systems. Neither survey proves that every proposed title will become a large occupation.

Portfolios can show operating judgment

A new title can disappear even when the work remains. Job seekers can prepare for that possibility by showing artifacts that apply across labels.

An agent specialist can present a workflow map that marks tools, permissions, review points, and fallback steps. An evaluator can show a test set with a scoring guide and a record of failures. An ROI analyst can document a baseline, cost boundary, and decision rule for expansion.

Domain experts can add examples that expose the limits of a generic benchmark. A legal reviewer can build cases around conflicting authority. A finance reviewer can test how the agent handles missing records or a reconciliation that does not balance.

These artifacts demonstrate judgment that an employer can inspect. They also reveal whether the candidate understands the consequences of a weak result in the target domain.

The labor market is still naming the work

Current evidence supports a measured conclusion. Employers expect agent rollouts to create work in design, training, evaluation, security, process redesign, and value measurement, but broad surveys still track intentions more often than filled positions.

That gap matters for anyone entering AI work. A candidate should treat a new title as an invitation to inspect the underlying responsibilities, reporting line, and decision rights. Across the proposed roles, employers need people who can turn an agent into a controlled process and show whether that process performs.

Sources

1. Microsoft, 2025 Work Trend Index Annual Report, “2025: The Year the Frontier Firm Is Born”, April 23, 2025. Microsoft reports a 31,000-person online survey across 31 markets, conducted by Edelman Data x Intelligence from February 6 through March 24, 2025, including 9,037 leaders, but the cited page does not report the manager subgroup size. Microsoft publishes AI products, and the methodology does not identify a separate external funder.
2. International Labour Organization, “The Impact of GenAI on Jobs, Productivity and Work Organization: A Review of the Empirical Evidence”. The PDF is dated May 2026, and the ILO web record is dated June 1, 2026. The brief reviews experiments, firm data, labor platform studies, and worker and firm surveys, and it does not present a single survey sample or a separate funder.
3. World Economic Forum, Future of Jobs Report 2025, January 7, 2025. The survey dataset contains 1,043 unique company responses representing more than 14.1 million workers across 55 economies and 22 industry clusters, collected from May through September 2024. The report identifies survey and data partners but does not name a separate funder on the cited public pages.