Imagine your team starts using AI and doubles the number of drafts, analyses, or code changes it produces. Everyone is busy. The dashboards look impressive.
Then the review queue grows. People approve work they barely understand. Errors travel further before someone catches them. Your strongest employees spend their days checking other people’s output.
Have you improved productivity, or simply moved the bottleneck?
This is the management problem I want to focus on. Once humans and AI work together, producing an answer is only part of the job. Someone still needs to decide whether the answer deserves to become an action.
Start with the work
“We need everyone using AI” is a weak operating instruction. It tells employees which tool to use without explaining which problem to solve.
Start with a workflow instead. Where does work wait? Which steps are repetitive? Which mistakes are expensive? Where does judgment depend on relationships, context, or experience?
Consider a customer support team. AI might summarize a case, retrieve relevant guidance, and draft a response. An employee verifies the facts and decides whether the response fits the customer’s situation. A refund above an agreed amount requires approval.
The manager’s job is to make those boundaries explicit. Access to a tool does not establish authority to act.
Give every workflow a human owner
An AI agent can perform tasks. It cannot carry organizational accountability the way a person can.
Every workflow therefore needs a named human owner who has the authority, time, and knowledge to manage it. That person needs to understand what the agent can access, what it can change, when it must stop, and how to recover from a mistake.
“A human is in the loop” is insufficient if that human is overloaded or cannot inspect the evidence.
Ownership also needs to extend beyond individual outputs. Who maintains the workflow when a policy changes? Who investigates recurring errors? Who can disable the agent? Those responsibilities should be settled before the workflow becomes routine.
Delegate authority in stages
People earn greater responsibility through demonstrated competence. AI workflows should earn greater autonomy through evidence within a defined scope.
| Level | AI authority | Human responsibility |
|---|---|---|
| Assist | Summarize, retrieve, or suggest | Make and execute the decision |
| Draft | Prepare a proposed response or change | Verify and approve before execution |
| Act within limits | Execute defined, reversible actions | Monitor results and handle exceptions |
| Escalate | Stop when limits or conditions are exceeded | Resolve the case and decide whether to resume |
A workflow that reliably tags support tickets has not demonstrated that it can issue refunds. Permission should follow the task, its consequences, and the evidence available.
When the model, instructions, data sources, or task changes materially, reassess the permission. Yesterday’s results may not describe today’s workflow.
Design review as part of the work
If every AI output lands on the manager’s desk, the manager becomes the team’s throughput limit.
Build review into the workflow. Use appropriate automated checks for verifiable requirements. Route specialized decisions to qualified employees. Require approval for consequential actions. For suitable low-risk work, use sampling and exception monitoring once performance has been demonstrated.
Make review easier by requiring evidence: source records, relevant policy, test results, or a clear description of what changed. A fluent explanation from the same model is not independent verification.
Set a limit on how much work can await review. When that queue reaches capacity, slow generation and address the bottleneck. Producing another hundred drafts is unlikely to help the person already struggling to check the first hundred.
Manage trust through participation
Some employees will experiment eagerly. Others will hesitate. Neither reaction tells you whether someone has good judgment.
Ask what is behind the hesitation. Is the tool unreliable? Is the employee unclear about permitted data? Do they fear losing expertise or being evaluated against an unrealistic output target? Those concerns require different responses.
Invite enthusiastic users and skeptical employees to test the same workflow. Compare outcomes. Discuss failures openly. Make reporting an AI mistake an expected part of improving the system.
A useful team agreement establishes permitted uses, required checks, escalation paths, and how saved time will be used. Employees should also know how their performance will be assessed. Tool usage alone says little about their contribution.
Protect the team’s ability to judge
If AI does the first pass every time, managers need to make sure people still develop the expertise required to review it.
Keep deliberate opportunities for independent problem solving, mentoring, and explaining decisions. Ask employees to walk through why a result is correct and where it could fail.
Experienced reviewers need time and recognition for that work. Otherwise, you risk rewarding visible production while hiding the labor that makes it trustworthy.
Measure the whole workflow
More drafts, tickets, or pull requests can signal activity without demonstrating value.
Track the time from request to accepted result, including review and rework. Track escaped errors, customer outcomes, total cost, and the age of the review queue. Examine whether the team can explain and maintain what it delivers.
Treat these measures together. Faster delivery with more consequential mistakes is a different outcome from faster delivery with stable quality.
Start small enough to learn
Choose one bounded workflow. Establish its current performance. Name its owner. Define the agent’s permissions, review method, and stop conditions. Run a limited pilot and compare complete outcomes, including the human effort required.
Expand when the evidence supports expansion. Redesign or stop when it does not.
Managing humans and AI will still require coaching, prioritization, and business judgment. It also requires a deliberate answer to a new question:
When producing work becomes easier, how will we preserve the capacity to understand it, verify it, and take responsibility for it?