Human in the loop isn’t just a buzzword. It’s the lynchpin for this next era of deploying autonomous AI agents within the workplace.
Full autonomy still isn’t the goal
An AI that is considered capable of acting autonomously doesn’t mean the same thing as an AI that is trusted to do so. This isn’t a new concept. After all, most employers have (or hopefully have) an employee base that could be trusted to operate without a direct line of authority. And yet, managerial positions and project coordination roles remain a key part of any job market.
For businesses to function healthfully and sustainably, there still needs to be a chain of command where the bittier, more repetitive tasks are overseen by senior team members. Human in the loop is the exact same principle. You can’t have AI monitoring itself, since that would quickly create an echo chamber in which mistakes proliferate just as much as successes.
One bad autonomous action can cost you greatly – perhaps even more than the job role it was intended to replace.
Where oversight still matters
Financial decisions, external communications (particularly with clients), and potentially destructive actions all need to be carefully monitored and considered by a human. AI cannot offer the same nuanced thinking or real-world experience that helps to shape higher level decision-making, and these aren’t areas where any of us can afford to just ‘wait and see’ if it makes a mistake or not.
Even if you were to run year’s worth of simulations of real-world scenarios, would you trust the AI to consistently make the right decisions on your behalf when it actually has the reins of your business in its hands?
During action is another important point for human oversight. You don’t just want to see end results before knowing whether the AI was doing it right or wrong.
Finally, after action: a reconstructable record of what happened and why, for both debugging and compliance
Patterns for giving humans real-time visibility
- Streaming text logs: cheap, but requires the reader to reconstruct visual context mentally
- Screenshot snapshots at intervals: better context, but misses fast-moving or transient states
- Full live session viewing: highest fidelity, whereby a human watches the exact browser state the agent sees, through an embeddable browser view. Best suited to spot checks and incident response rather than constant monitoring at scale.
Make intervention a core part of the design
Remember that visibility is literally only half the story. It’s one thing to see what the AI is doing and know that it’s working toward the right goal, but that doesn’t mean anything at all if there’s no easy opportunity for intervention. Waiting until the task has been completed incorrectly just isn’t feasible.
Escalation triggers that alert a human to a possible point of intervention are vital, as is a takeover hand-off, where the human can take over the AI’s work rather than having to start from scratch. This is how you can ensure that automations work alongside your human employees, rather than slowing the business down or creating new problems alongside the old.
Discussion about this post