Imagine going to a job board, making a hire for a chief of staff, and onboarding that new hire within a few minutes.
That’s the fantasy that artificial intelligence companies are making a reality with virtual teammates. Optimizely recently released virtual teammates to provide marketers with role-specific AI coworkers, ranging from a chief of staff to a marketing analyst.
The question is no longer if virtual teammates will be joining the workforce, or even when. They’re here. The question has shifted to how these virtual teammates get onboarded and set up for success, just like a human employee on their first day.
To find out, I talked with Optimizely and asked security and HR leaders what they would put in place first. Here are the best practices for onboarding a virtual teammate.
Provide a clear job description
Job descriptions have a reputation problem. In a Monster survey of 1,060 U.S. workers this summer, only 10% said job descriptions always reflect the reality of the role.
A human hire can work around a bad job description by asking questions and watching how the team operates. A virtual teammate can’t. The input of a job description defines the outputs a virtual teammate is able to provide. Give it a vague role and the work comes back vague.
Optimizely gives its virtual teammates preassigned roles and descriptions, so when someone hires a chief of staff, the virtual teammate’s performance lines up with what the role expects.
The big difference for organizations designing for a virtual workforce is what needs to be defined. For example, most onboarding documents or job descriptions don’t say, “Check with your manager before wiring money to a vendor.” An employee would typically know to ask. A virtual teammate knows only if the job description lays it out explicitly.
“Explicit autonomy boundaries and escalation triggers, also known as human-in-the-loop rules, are a critical addition to an AI teammate’s job description,” says Megan Barbier, chief human resources officer at Xactly, a sales performance management platform. An AI teammate, she says, “requires precise, explicit parameters defining what it can execute independently versus when it must stop and ask a human for approval.”
Barbier offers two examples of triggers. Hand the work back to a human when the teammate’s confidence score falls below 85%. Do the same when it picks up negative sentiment in a client thread. Skip that step, she says, and “you risk deploying an unguided tool instead of defining a job role.”
Add connectors to provide memory
The difference between an AI agent and a virtual teammate is long-running memory. An agent operates within a window of context. A virtual teammate needs to recall an event or a critical detail from three weeks ago and let that history inform what it does next.
That’s why the first step Optimizely walks its customers through is “connectors,” which plug a virtual teammate into email, campaign data, and the other systems where the work happens. Kevin Li, senior vice president of product at Optimizely, says this is an important step, that a human oversees, to make sure the virtual teammate gets access to the right information.
“Employees on day one are given a computer, credentials, and access to the systems and information they’ll need to do their job,” says Li. “Things are no different with virtual teammates. Connecting the AI to Slack, email, and other sources of institutional knowledge is the equivalent of handing a new hire their logins.”
Unlike human employees, who can be overloaded with information during onboarding, a virtual teammate prefers to drink from a firehose. Virtual teammates are exceptional at ingesting information and using it to determine what matters to the job at hand.
But an abundance of information doesn’t mean an abundance of permissions. The teammate gets its own login and its own permissions, rather than inheriting the access of whoever set it up. Everything else about guardrails builds from there.
Set least-privilege access
One difference between a virtual teammate and a human one is speed. A person pauses when a request feels off. A virtual teammate carries out instructions, including malicious ones, at machine speed.
That’s why guardrails are an absolute necessity for virtual teammates. Optimizely’s default is a human in the loop. The teammate proposes its work, and a person approves it before it acts. Over time more autonomy can be granted, just like you’d give a good employee more freedom to operate.
The next guardrail covers what it can touch in the first place. According to Mehdi Houdaigui, Deloitte’s US Cyber AI leader, “least-privilege access” is the foundational guardrail to ensure teammates get only the permissions a job requires.
“An email agent that drafts messages should be able to read approved folders and create drafts, but not send, delete, forward or export contacts,” says Houdaigui. “Similarly, a campaign agent should not be able to publish or change spending. Those limits must be enforced by the underlying systems, not simply written into the agent’s instructions.”
Instructions and descriptions are good guides. But a job description tells a virtual teammate what to do. It doesn’t stop the teammate from doing something else. If the email connector allows sending, a manipulated teammate can still send, whatever the description says. The guardrail has to live in the permissions.
Log the work
Permissions cover what types of work can be done. Logging shows what work actually got done, and that’s what lets you manage and improve a virtual teammate.
“You cannot govern what you cannot see,” says Leslie Nielsen, chief information security officer of Mimecast, a cloud-based cybersecurity platform, who puts full logging and monitoring ahead of every other guardrail. Without it, he says, “you end up with shadow AI, where agents quietly accumulate access no one ever approved.”
Tightly scoped teammates can have too much in the way of permissions, and that access review requires continuous oversight. Just like scheduling a 90-day review for a new hire, calendaring an access review is what’s required from a hirer of a new AI coworker.
The last mile questions
Virtual teammates invite questions with tough answers. Where do virtual teammates sit on the org chart? Who does the AI chief of staff report to? When AI makes a call a human wouldn’t have made, whose name is on the decision?
Virtual teammates are proving that they can work. But one thing is clear: People own the outcome, no matter what actions the teammates take.
That’s why the descriptions and guardrails matter to the org chart. Every action traces back to a person who approved it, or who decided it could run on its own. The work is still your job.