Artificial Intelligence
You’ve Been Prompted
Imagine you’re hiring an AI agent. How do you onboard and manage the new hire nobody interviewed?
Julie Colwell
Principal Strategist
Workday
Imagine this job posting.
Operations Associate
Company-wide · Remote (everywhere) · Full-time (24/7)
Reports to: TBD
About the role
We’re looking for a proactive, detail-obsessed teammate to join every department at once. In this role, you’ll have visibility into all open requisitions, invoices, and support tickets across the company. You’ll spot what’s slipping before anyone raises it, draft the fix, and bring it to a human for review and approval.
What you’ll do
- Monitor every open requisition, invoice, and support ticket in the organization
- Identify issues before they’re escalated
- Draft solutions and route them to the right person for approval
- Follow up as many times as it takes
What you bring
- Deep familiarity with every policy we’ve ever written
- The ability to follow established guidelines from day one, with no ramp-up period
- Seamless integration with every other system in the company
- Availability at 3 a.m. and every other hour
- Persistence that doesn’t fade
What we offer
- Salary: none
- Desk: none
- You’ll appear on the budget as a line item
Report
AI Agents Are Here—But Don’t Call Them Boss
Same Orientation, New Kind of Hire
Typical new hires get a manager, the right access, a role, reviews, and a team. Agents need the same foundation to do their best work.
Companies are moving fast, and the ones pulling ahead are building governance early. Deloitte surveyed 3,235 leaders and found:
74% of companies plan to deploy autonomous agents within two years.
21% have a mature governance model: clear limits on what agents decide alone, real-time monitoring, and audit trails.
What success takes: Strong runtime enforcement and systems that work together are what avoid runaway costs, unclear value, and increasing risk. Model capability isn’t the key performance indicator.
Who manages the agent: Gartner research from this month found only 24% of leadership teams agree on who should be accountable for them. That gap is an opportunity for people leaders, who already know how to onboard, supervise, and support a new team member. In fact, during his keynote at Rising, Workday co-founder and CEO Aneel Bhusri said agents need to be “managed like humans.”
Workday co-founder and CEO Aneel Bhusri said agents need to be “managed like humans.”
No Ramp-Up Necessary
For as long as we’ve had workplace software, people have told it what to do. The cursor blinked until you typed. The email sat in the inbox until you hit send. People started every workflow.
For as long as we’ve had workplace software, people have told it what to do.
Autonomous agents change that order. They start within an existing process, and by design, they proactively find what needs to be done, and ask a human manager for the final decision.
This shift is happening quickly. Gartner projects that 40% of enterprise applications will feature task-specific agents by the end of 2026, up from less than 5% in 2025.
The Human Loop
With the right context and governance in place, autonomous agents will be trained to act with human approval. But keeping a human in the loop raises two questions: Which loop? And what role does that person play?
Match the person to the decision. Gerrit Kazmaier, Workday’s president of product and technology, explained where people belong in three tiers:
In the loop: anything that materially affects income or employment.
On the loop: routine approvals.
Out of the loop: high-volume checks where machines are more accurate.
Let the loop change over time. People should be most involved the first time an agent runs a task, with autonomy growing as trust is earned. Wharton’s Ethan Mollick says agents should ask people for ideas and expertise, not just review and approval.
No Path to Management
Toyota figured this out early, as Workday’s Laila Almounaier, vice president of product and engineering, pointed out at Rising this month. In its early days, Toyota’s manufacturing facility had a power loom that stopped itself when a thread broke and called a person over to help. This safeguard let one worker run 20 looms. The machine handled the routine, while the person handled the broken thread.
Without the right design, agents that ask can drift into agents that direct.
That balance matters. Without the right design, agents that ask can drift into agents that direct. Research shows what happens when that balance tips:
A 2026 European Commission study linked algorithmic management, where software directs workers, to lower autonomy, fewer breaks, and higher stress.
Researchers at Curtin University note these systems were built around employer efficiency, with little regard for their impact on the people using them.
The hopeful part: Wellbeing is better when workers and employers agree on autonomy, and negative effects diminish when workers feel represented and valued.
In other words, people handle direction best when they have a say in it. Agents shouldn’t be people managers.
Must Show Your Work
Toyota’s model worked because the loom only called a person when it needed one. Agents work best the same way, because every request requires someone’s attention.
Transparency helps. In one small study, proactive agents improved efficiency but disrupted workflows when there were excessive notifications. Showing what the agent was doing, and why, reduced the disruption.
Smart design pays off. Companies that redesign work around agents are more than twice as likely to beat their cost goals, according to Gartner.
So, how can worker dignity, autonomy, and value be preserved while taking advantage of the benefits agents have to offer?
First 90 Days
Start small, then scale. Deloitte found that companies succeeding with agentic AI start with lower-risk use cases, build governance, and expand from there.
One use case at a time. FIS, a fintech company with 55,000 employees, rebuilt HR processes instead of layering agents on top, rolled out one use case at a time, and brought in IT, data, and security from day one. Steve Begley, FIS senior vice president of people operations, said the first step is examining your processes and data.
Ask the tough questions. Whatever your starting point, ask:
- What can this agent do without asking?
- Who is the named owner when it’s wrong?
- How does it explain why it’s asking?
- How does someone say “not now”?
- How will we know oversight is real, rather than assumed?
Reports to: You
Once the agent goes live, the questions start: Should this invoice go out? Does this ticket need escalating? Can this requisition move forward? Each one lands with a person, and someone has to own the answers, including the ones that turn out wrong.
Why that person matters: Joel Hellermark, Workday’s chief AI officer, said that when AI can produce analysis on demand, human judgment becomes the scarcest and most valuable part of the work. The agent can gather the facts and draft the fix, but a person decides what’s right for the business and the people affected.
Now the agent writes the prompts, and the human job is answering them: knowing what to approve, what to question, and when to say “not yet.”
For years, getting value from workplace software meant learning to prompt it. Now the agent writes the prompts, and the human job is knowing which to approve, which to question, and when to say “not yet.”
Report
AI Agents Are Here—But Don’t Call Them Boss
A gap between effort and impact is the real ROI problem. Discover the solution.