What Should AI Do, and What Should People Decide?
Designing Human-in-the-Loop Workflows
Define responsibilities for planning, processing information, operating systems, handling exceptions and approving results.
AutoAgents.AI·October 15, 2026·6 min read
Define what AI can execute and what people must approve. Design clear handoffs, meaningful confirmations and exception handling for enterprise AI agents.
In an enterprise AI agent platform, “AI handles it, a person reviews it” must translate into specific screens and actions. Employees need to see the fields to check, correct mistakes and let the task continue. Restarting the whole task after every correction can erase the time saved.
Assign permissions to specific actions
Consider preparing a purchase request. An employee asks the system to organize quotes, fill out the request and submit it for approval. That involves at least three distinct responsibilities: processing information, making business choices and submitting data to a system.
| Work | What AI can do | What people or established rules retain |
|---|---|---|
| Organize quotes | Extract information, check units and flag missing items | Resolve ambiguous fields and special pricing terms |
| Prepare options | Calculate and compare using agreed rules, with supporting evidence | Choose suppliers and assess exceptions involving lead time or quality |
| Prepare a request | Fill out a draft and check required fields and attachments | Confirm the purchase, amount and budget allocation |
| Submit and route | Submit when authorized and read the resulting status | Complete approvals under company procedures and own the purchase decision |
The allocation must follow company policy. Steps with clear rules and reliable checks can gradually receive more execution authority. AI cannot expand permissions that the business has not granted simply because it judges that doing so would be more convenient.
The same employee may grant different access for different tasks. Permission to read a contract folder does not include permission to edit it or send its contents externally. Permissions should identify both the resource and the permitted action.
Show the real impact of the next action
A useful confirmation shows the action about to happen, its key parameters and its basis. Before submitting a purchase request, for example, show what is being purchased, the amount, the budget allocation, the attachment version and the process that submission will trigger.
A bare “Continue?” button leaves the employee reopening every source to check it. A confirmation screen should bring the relevant differences and risks into view and reduce repeated searching.
Confirmation should also apply to a specific version. If an employee approves a purchase request for RMB 100,000 and its amount, recipient or attachments subsequently change, the system must reassess whether confirmation is needed. When a long-running task resumes, it should also check that earlier authorization remains valid.
Reversible draft preparation can continue within existing authorization. Payments and external commitments need confirmation immediately before execution, enforced by the system. Company rules should define the thresholds; placing them only in a prompt is insufficient.
Separate planned approval from error recovery
Human-in-the-loop design places human participation at defined points in agent execution. This includes both planned approvals and intervention after a failure. These have different purposes and should be measured separately.
Confirming a purchase amount as part of the process is normal collaboration. Rebuilding a request because the system entered incorrect fields is corrective work. Counting every human action as a failure encourages the removal of necessary approvals. Counting only eventual completion can conceal extensive rework.
When a person takes over an exception, the system should show where the task stopped, what has already happened and which business records may be affected. Keep generated files and drafts available. The person can then provide information, correct the result or end the task.
For example, a page may time out after a request is submitted, leaving its status uncertain. Query the business system to see whether the request exists, or ask a person to verify it. Submitting again immediately could create a duplicate. Recovery must depend on the external system's actual state.
How Workflow 2.0 Combines Steps and Human Confirmation
Workflows suit defined sequences, field checks and approval conditions. Agents suit adapting retrieval to the material or choosing a path among permitted tools. Anthropic's Building effective agents makes this distinction between predefined workflows and dynamic agents.
Lingda Agent Platform's Workflow 2.0 combines fixed steps and agentic nodes within one task. For a purchase request, an agentic node can read quotes, select tools for different formats and decide from the results whether it needs to query the product database. Field checks and approval conditions follow the rules defined in the workflow.
If a quote omits shipping costs, the task should pause for the buyer to confirm the pricing terms. Once the information is complete, the agent continues calculating and fills out the draft. Key fields appear for confirmation before formal submission. Permissions control the system action, and its result becomes part of the task record.
When an employee corrects a field, the system needs to save the new value and identify which downstream results are no longer valid. Changing a quote's units usually requires recalculation; correcting a document title may not require retrieving every source again. Preserving task state reduces repeated work and repeated approvals.
In Lingda Desktop, employees can track progress, handle confirmation requests and inspect generated files. A pending request should display the information that needs judgment, such as the purchase amount. After confirmation, the task follows the selected branch; after a data change, it rechecks affected results.
Turn useful feedback into shared methods
A human correction may reveal a lasting rule or a temporary project arrangement. Treat them differently.
Moving one customer's delivery date forward establishes a special arrangement for that task. A business owner confirming a new service standard for an entire customer category provides a basis for updating a shared method.
Organizations can turn recurring, validated corrections into role-specific Skills. A Skill is a reusable task method with clear conditions, steps and checks. If buyers repeatedly correct packaging units, for example, capture the conversion rule and verification method in a Skill. Test it on other quotes before sharing it, and preserve versions so changes remain traceable.
Profy turns specialist methods into digital expert services. In its distillation workshop, creators explain how to assess a problem, which information is essential and when to involve a person. Validated methods become Skills that guide a digital expert as it reads materials, uses tools and delivers an editable document or report. When users request changes, creators can test similar tasks to decide whether that feedback belongs in the general method.
To assess collaboration, record review time, rework and whether intervention leads to completion. Clear responsibilities reduce workload when employees can understand a pending decision and continue the task without repeating earlier actions.
Frequently asked questions
Is a task still automated if a person reviews it?
Yes. Automation can cover part of a complete task. Planned approval is a normal business requirement. Assess how much work AI performs and whether review and correction costs outweigh the benefit.
Do more confirmations make an agent more reliable?
Not necessarily. Repetitive or vague confirmations encourage mechanical clicking. Place confirmation before actions with clear business consequences, show enough evidence for a decision, and enforce execution through permissions and programmatic checks.
Further reading:Connecting AI Agents to Legacy Systems Without APIs · AI Agent Production Readiness: Permissions and Deployment