I think HR should tell employees, in writing, how far each AI agent is allowed to act before the agent goes live. Vendors now ship the controls. BCG’s new survey data shows that four in ten employees want more openness about AI, and that few CEOs worry about it.
The vendors have handed HR a dial
This week Workday announced autopilot agents for hiring, payroll and accounting. In its release, the company says customers “decide which actions an agent can take on its own, which need review, and when a person makes the final call,” that every action is logged, and that customers can set how often agents run. We described the details in our report on the announcement.
That is a useful design, and Workday is one example of a pattern. Each setting on that dial is a decision about people. A recruiter’s outreach either goes out after approval or it does not. A payroll anomaly either gets drafted into an employee message for the payroll team to review or it does not. Somebody inside each customer will choose, task by task. In most companies that choice will land in a configuration screen that employees never see.
Employees already say they are in the dark
BCG published the second edition of its CEO Insomnia Index on October 14. In the release, BCG reports that four in ten employees say their CEO has not been transparent enough about how AI will affect their roles and career opportunities. Only 15% of CEOs worry that their approach to AI transformation could erode employee trust. Nearly 90% of CEOs are confident employees trust their leadership, and 46% of employees agree.
The same release attaches a cost to the gap. Half of the employees who say their CEO’s handling of AI has increased their trust are more willing to go above and beyond than they were three months ago. Among employees whose trust has eroded, fewer than 10% say the same. BCG’s survey covered 645 CEOs of companies with at least $100 million in revenue, and this edition added employee perspectives.
Matthias Tauber, a BCG managing director and senior partner, offered a diagnosis in the release: “CEOs assume that since employees hear their words, that’s enough to build trust, but employees also judge CEOs by their actions.” An agent that sends, flags or drafts without a visible rule is an action employees will judge.
What I would publish
I would write one page per agent, and it would answer four things in plain language. What the agent does without asking. What it does only after a named role approves. Who reads the log, and how often. How an employee or candidate can ask for a person to look at an outcome. HR already owns the vocabulary for this, because it writes the policies the agents are told to follow. Workday’s release says its agents inherit each company’s policies, permissions and approval rules, so the policy documents are the right place to start.
The page should describe the rule and leave the configuration out. It can say that outreach to candidates needs recruiter approval, a limit Workday describes for its hiring agent, without publishing a screenshot of the admin panel.
The strongest objection
The best case against publishing is practical. Settings change as agents mature, and a page that is out of date is worse than no page. Legal teams will also ask what a published limit commits the company to, and security teams will not want a description of controls circulating. I take all three seriously.
Each has an answer. Dated pages with a named owner solve the first, and a review whenever a limit changes keeps them accurate. The second is solved by describing outcomes, such as who approves what, which carries no more legal exposure than the policy it restates. The third is solved by publishing the rule and withholding the configuration. A company that cannot describe in a paragraph what its agent may do alone probably has not decided yet, and finding that out before launch is the point of writing the page.
Where to start
The earlier piece on inventorying hiring AI before January described the list of systems. Add one column to it: the highest-impact action each system can take without a person. Fill that column in before pilots begin, then hand the finished pages to the CEO. BCG’s numbers suggest the CEO may not yet know employees are asking.
Source: Boston Consulting Group, “Forty Percent of Employees Say Their CEO Isn’t Transparent Enough on AI”
