In a quotation workflow, AI Agent may have already organized the request data, compared it with historical orders, and drafted a quote, yet still be unable to handle a few exceptions. When managers see this, they may conclude that the project is "not ready to use." Employees may also feel that if they still have to check every item at the end, it would be faster to do everything themselves from the start. But the problem is not necessarily that the model lacks capability. The company may still be using "Can it completely replace a person?" as its acceptance criterion.
Palantir CTO Shyam Sankar made a similar point in an interview with Stratechery: companies adopting AI should treat the workflow as the basic unit, rather than aim to replace an individual knowledge worker. If a project requires AI to replace people entirely, the company may still declare the entire project a failure even when AI completes 98% of the work, simply because the remaining tasks are unfinished. In his view, companies should redesign workflows around AI's current capabilities. Just as factories did not simply replace one machine when they began using electricity, businesses need to rethink how the entire operation is organized.
If AI completes 98% of the work, why does the workflow still break down?
Sales quotations are a common example. AI Agent may be able to organize customer requirements, search historical prices, compare costs, and draft a quote. However, sales, finance, or management may still need to determine whether a discount is appropriate for the current competitive situation, whether special payment terms are acceptable, or whether the delivery schedule involves cross-departmental commitments. If the project goal is for "AI Agent to complete quotations automatically," the company may declare the entire project a failure because the last few tasks still require human judgment. By contrast, if the goal is to reduce the time spent organizing information, prevent omissions, and have managers handle only exceptions, the team can begin to see the results of human-AI collaboration.
Using supply chain analysts as an example, Sankar noted that improving an individual's productivity does not necessarily improve return on investment. When evaluating a supply chain project, a company therefore cannot look only at how much faster an analyst works. It must also confirm whether procurement, inventory, scheduling, approvals, and other parts of the workflow have actually improved. When reviewing an existing project, start by asking: "Who receives the information prepared by AI Agent? Can that person use it immediately, or do they have to reorganize it?" If the next employee still has to verify everything from scratch, the company should first improve the handoff and clarify responsibilities before discussing whether to replace the model.
Divide the workflow into execution, judgment, and approval tasks
When redistributing responsibilities, start with a workflow that has a clearly defined beginning, end, and owner. For example, a procurement price comparison process could begin when a request is received and end when a manager approves the purchasing recommendation. Then divide each step into three categories: tasks that AI Agent can perform according to explicit rules, tasks that employees must assess based on the circumstances, and tasks involving responsibility or commitments that require human approval. The purpose of this classification is to give every handoff clearly defined inputs, outputs, and owners. There is no need to distribute the work evenly among the three categories.
AI Agent can first handle tasks such as collecting information, standardizing formats, checking rules, and preparing drafts. If the data conflicts, the conditions are unclear, or the current rules cannot address the situation, an employee should make the judgment. When an amount exceeds a department's limit, content is about to be sent to a customer, or procurement terms may create a long-term commitment, the workflow should pause until someone with approval authority reviews it. The company must also define which source materials, differences, and supporting reasons the approver must examine. Managers should not be asked to approve a recommendation simply because it looks complete on the surface.
EgentWrX can connect tasks into workflows and require human approval before a handoff. Once AI Agent finishes the preceding work, the workflow pauses until a person reviews and releases it. Companies can begin by testing recently completed real-world cases to identify which steps most often result in rejection, requests for additional information, or rewrites. They can then adjust how work is divided between AI Agent and employees. During the initial rollout, companies can first identify higher-risk steps, such as those involving ambiguous conditions, amounts above established limits, or external commitments, and decide who should review them. There is no need to move every task into the system at once.
Turn experienced employees' knowledge into reviewable skills
After breaking down the workflow, companies can divide tasks that "require human judgment" into two categories. The first covers situations where the company already follows a consistent approach but has never documented it. The second covers work where circumstances genuinely differ from case to case and professional judgment is required. For example, when an experienced procurement employee sees that a supplier has changed its payment terms, they may routinely check the supplier's past performance, alternative sources, and contractual restrictions. New employees, however, often see only the final approval or rejection and do not know which information the experienced employee reviewed along the way.
Companies can interview the employees who actually perform the work and document the three to five most commonly used decision criteria, exceptions, and output formats. They can then validate each item against real cases. Repeatable practices can be created as skills (AI Skills) in EgentWrX, allowing AI Agent to organize information according to the same rules every time. If the company wants to deploy employee-created skills across the organization, it should first submit them to the back-end review queue. They can be rolled out only after passing review. Decisions that affect contracts, pricing, or customer commitments should still require human approval to prevent an experienced employee's rules of thumb from being treated as automatic decisions.
Evaluate the entire workflow and actual business outcomes
After responsibilities have been redistributed, the acceptance criteria must also change. Teams can record how many handoffs each task requires from start to finish, where delays occur, whether work is returned because information is missing or because a judgment is disputed, and whether the final result meets the original business objective. This information reflects the company's actual costs more accurately than whether AI produces fluent answers. It also helps the team determine whether the next round of improvements should focus on data, rules, or the allocation of responsibilities.
Sankar has observed that customers seeking to increase revenue and improve market competitiveness are more likely to succeed than customers focused on reducing headcount and costs. However, the source does not provide the number of customers, their industry distribution, or a quantitative comparison, so this observation should not be treated as a universal conclusion. Companies can instead use it as a point to examine: in addition to calculating how much labor AI could save, they should assess whether AI can help sales teams handle work they previously lacked time to address, respond to customers more quickly, or reduce the risk of overlooking critical cases. For the selected workflow, first document the expected outcomes, risks, and owners, then validate them with real cases. Expand the rollout to other units only after confirming that the workflow achieves the intended results.
FAQ
Which tasks should be assigned to AI Agent first?
Start with tasks governed by clear rules, supported by accessible input data, and producing outputs that are easy to verify, such as organizing inquiry information, comparing existing terms, and drafting documents. If a task involves pricing commitments, contractual responsibilities, or exceptions that existing rules cannot handle, an employee should evaluate it before someone with the appropriate authority gives final approval.
At which points should human approval be required?
Human approval can first be introduced when an amount exceeds a department's limit, conditions are unclear, data conflicts, or content is about to be sent outside the company. Approvers must be able to review the source data, before-and-after differences, and reasons supporting the recommendation so they can make a substantive judgment. They should not see only an "Approve" button.
If AI Agent can complete 98% of the work, should the company keep switching to more powerful models?
Not necessarily. First determine whether the remaining errors result from insufficient model capability, missing data, unclear rules, or a flawed handoff process. If the errors are concentrated in a small number of higher-risk judgments, redesigning the workflow and retaining human approval may serve the project's goals better than requiring the model to handle everything.
How can a company determine whether a skill is ready for organization-wide deployment?
Validate the rules, exceptions, and output format against real cases first. Confirm that different employees following the same approach can consistently produce verifiable results, then submit the skill to the back-end review queue. It can be deployed across the company only after passing review. If the skill affects contracts, pricing, or external commitments, the workflow should still retain human approval.
References
- An Interview with Palantir CTO Shyam Sankar About the AI Religious War — Ben Thompson's interview with Shyam Sankar, published October 29, 2026.
