During the first week of an AI implementation, many companies begin by sending a needs-assessment survey to each department, asking employees to describe “what they want AI to help with.” Two weeks later, half the forms come back blank, while the other half contain requests such as “help organize data,” “automatically generate reports,” or “reply to customer emails.”
These responses do reflect genuine needs. The problem is that they are too broad and insufficiently specific. With surveys like these in hand, the project lead has little basis for deciding where to begin—or how to evaluate the results once the work is complete.
The real obstacle is often the way the question is framed.
One Seemingly Simple Question Actually Conceals Three Difficult Challenges
“What do you want AI to help with?” sounds like an open-ended question. In practice, however, it asks respondents to do three things at once: understand what AI is capable of, break down their own work into tasks that can be delegated, and assess the potential value and risks.
For employees who have never really used AI, the first step alone is difficult to answer, making the next two steps even harder. It is like asking someone unfamiliar with machinery specifications to define the design requirements for a new production line. Even if they understand their own work extremely well, they may not know how to articulate their needs in an actionable, measurable form.
Reframing the question changes the situation. Ask, “What do you normally do at work?” and nearly everyone can begin by listing activities that occur every day, week, or month. The discussion also shifts away from imagined new features and back to the work the company is already paying people to perform—the work consuming labor hours and affecting delivery schedules. Whether AI can perform that work, and which parts it should handle, can be assessed in the next step.
Use Four Fields to Define Responsibilities Clearly
The inventory can be conducted by department, with each employee listing their responsibilities. The first round does not require a complicated form. Start with four fields:
- Responsibility: How is this task performed today? Where does the initial data come from? Who receives the completed output?
- Frequency: Is it performed daily, weekly, monthly, or only when a particular event occurs? Where possible, enter the number of times it is performed.
- Time per task: Record the actual average time an experienced employee needs to complete it under normal conditions.
- Number of users: How many people in the department perform the task or use its output?
Of these four fields, the responsibility description is the one most likely to miss the mark.
For example, “procurement management” and “quality management” are merely functional labels. They do not show where the work begins or what is ultimately delivered. A more useful description would be: “Download open purchase orders, filter for overdue items, and email suppliers to expedite delivery.” A single sentence identifies the data source, the actions performed, and the resulting output.
During the inventory stage, preserve the complete sequence of steps rather than trying to divide it too early. During the interview or evaluation stage, the team can then determine which parts AI should take over and which still require human involvement.
Use the Five TURBO Criteria to Identify the First Tasks
Once the responsibility inventory is complete, apply the five TURBO criteria: Time-consuming, User-wide, Repeat, Buffer for human-AI collaboration, and Operable.
In the initial stage, prioritize work with clear rules, consistent inputs and outputs, and errors that are easy to detect. These tasks are easier to scope and make it easier to establish acceptance criteria.
Among the clients we have advised, the first tasks to achieve stable execution have mostly involved document recognition, data entry, and reconciliation. The reason is straightforward: the inputs and outputs are clearly defined, and errors are easy to identify. Once an AI Agent completes a task, users can quickly determine whether the result is correct, while the project team knows exactly where adjustments are needed.
A Good Inventory Naturally Produces Three Types of Decisions
Consider the following sample responsibility inventory for a procurement department. The frequencies and time estimates are illustrative only; actual figures should reflect each company’s circumstances:
- Consolidate supplier quotations and standardize currencies, payment terms, and delivery schedules to produce a bid comparison table—approximately 12 times per month, requiring 45 to 90 minutes each time. This task is suitable for an AI Agent because the data fields are consistent. AI can extract the data and highlight discrepancies, while negotiations remain the responsibility of employees.
- Draft expediting emails in each supplier’s language—approximately 15 times per week, requiring 8 to 15 minutes each time. This work is suitable for implementation as an AI task, but a procurement employee must confirm the promised delivery date and recipient before sending.
- Assess the manufacturing capabilities and partnership risks of new suppliers—approximately twice per month, requiring two to four hours each time. AI can help organize the information, but supplier selection decisions should not be delegated to AI.
- Negotiate with suppliers and coordinate responses to material shortages—approximately four times per week. This type of work relies heavily on negotiation, interpersonal relationships, and situational judgment, so it should not be included in the first group of tasks.
A quality assurance department’s responsibility inventory will reveal similar layers. Comparing suppliers’ material certificates against procurement specifications is suitable for AI, which can classify results as “match,” “discrepancy,” or “requires confirmation.” When deciding whether to accept nonconforming products under concession, however, AI can organize historical cases for management’s reference, while decision-making authority remains with the authorized manager.
A good responsibility inventory should naturally include three types of conclusions: “suitable,” “suitable only for assistance,” and “not suitable.” If every item in the inventory appears suitable for AI, the team should reassess whether it has defined the task scope too broadly.
It is also important to remember that a responsibility inventory is not the same as a development backlog. For example, the first version of a “quotation consolidation” task might handle only field extraction and the creation of a discrepancy table. Once the output is stable, historical quotation searches can be added, while negotiation decisions should always remain with employees. A clearly defined task scope ensures that the team knows what each version is expected to accomplish and makes the results measurable.
The Four Most Common Forms of Resistance During an Inventory
A responsibility inventory touches everyone’s work, so some resistance during implementation is normal. The following are the four situations we encounter most often.
The first is concern that reported working hours will later be used in performance evaluations.
This concern cannot be dismissed with a simple assurance that it “will not happen.” Before the inventory begins, the company must clearly explain its purpose and access controls: the inventory is used only to estimate task value and establish a pre-improvement baseline, not to rank individuals. The original inventory should be visible only to department managers and the implementation team, while executive meetings should review department-level summaries only.
If the time required for certain tasks is genuinely difficult to estimate, begin by recording their frequency and steps without collecting individual time estimates. Keeping the inventory moving is more practical than insisting that every field be complete from the outset.
The second is the belief that one’s work is not worth documenting.
Experienced employees often internalize so many decisions as common sense that they write only “routine operations.” Continuing to ask, “Is there any other work?” rarely produces more detail. During interviews, ask instead: “Which files did you deliver last week?” “What must be completed before the end of every month?” or “Which steps need to be handed over when you take leave?” You can also open the most recently processed file and work with the employee to identify its inputs and outputs.
When you encounter a step that depends on experience-based judgment, continue asking which principles the experienced employee uses to make the decision. Do not avoid the effort, and do not impose your own logic to make their process appear more coherent. At many companies we have advised, seemingly simple tasks—such as form recognition, special account allocation, and shift rules—conceal long lists of exceptions. This know-how can be preserved only by uncovering it one layer at a time through persistent questioning.
The third is concern among managers that inventory results will be used to scrutinize staffing levels, leading them to submit only low-value, inconsequential tasks.
The first inventory round should establish a clear premise: its purpose is to improve how working hours are used by having AI take on structured work, not to rank departments. If managers remain concerned, begin with a pilot in one benchmark department. Once other managers see how the figures are used to improve processes rather than challenge staffing levels, subsequent inventory efforts usually proceed much more smoothly.
The fourth is concern among senior employees that documenting their work will make them replaceable.
Companies can position senior employees as mentors and knowledge stewards responsible for defining rules, validating results, and handling exceptions. The message can be direct: “In the future, people may not need to perform every task manually, but the company will need more experienced employees to direct and review the work produced by AI Agent.”
Knowledge contributions must also be recognized in performance evaluations. Organizing rules and transferring experience cannot become additional, unrewarded work for senior employees. If a company asks experienced employees to share their know-how without providing corresponding recognition, even the best inventory methodology will be difficult to implement.
What You Should Have at the End of the First Inventory Round
After completing the first inventory round, each department should have three deliverables: a responsibility inventory with clearly defined inputs and outputs, a screened list of candidate tasks, and a pre-improvement baseline for each task.
With these three items in place, the team will know how to define the scope when building each task. When evaluating results, it will have a baseline for comparison. And when deciding which tasks are worth deploying, it will have concrete evidence to support those decisions.
The advantage of starting with responsibilities is that employees do not need to speculate about which new capabilities AI might deliver. Instead, they begin by clearly describing the work they are already doing. As the inputs, processes, and outputs become clearer, the first tasks suitable for AI Agent will naturally emerge.
This article is excerpted from Intellicon Solutions’ “Agent-Ready: AI Applications in Taiwan’s Manufacturing Industry” white paper. The white paper includes complete sample responsibility inventories for procurement and quality assurance departments, along with a 12-question self-assessment that enables companies to evaluate their current level of Agent-Ready maturity in just ten minutes.