Once a company decides to implement AI Agent, departments quickly produce a long list of requests: Sales wants to organize inquiry emails, Procurement wants to compare supplier quotations, Customer Service wants to generate draft responses, and managers want AI Agent to help handle cross-departmental approvals. With so many needs to address, the team must first decide which task is most worth tackling first.
We recommend that the first use case not be the most high-profile, the most creativity-intensive, or the one involving the largest number of departments. Start with a routine task. Confirm that the required data is accessible, the processing steps can be clearly explained, and someone can validate the completed work. Then, based on the potential impact of errors, introduce human approval where appropriate. This approach allows the team to first learn how to structure work into an executable, verifiable process before deciding whether to expand the scope.
What Do Scientists Most Often Ask AI to Help With?
Anthropic analyzed 207,049 Claude and Claude Code scientific research sessions conducted between August 2 and September 30, 2026. The results showed that data analysis accounted for 33.7% of Claude sessions and 48.1% of Claude Code sessions. Writing and communication also represented 32.0% of Claude sessions. Meanwhile, researchers spent 28.6% of their time gathering information, but this type of work accounted for only around 11% of sessions across the two interfaces. The use of AI for ideation and hypothesis generation was also relatively limited.
This research suggests that users are more likely to ask AI to handle work for which it is well suited. However, this does not mean every enterprise should start with data analysis. After identifying the top 20% of tasks theoretically best suited to AI, the researchers found that Claude usage for these tasks was two to four times the level estimated based on how researchers allocated their time. Enterprises can apply the same principle: the importance of a task and its suitability for early implementation are two separate considerations. Strategic planning is certainly important, but if its objectives shift as discussions evolve and no one can clearly define what constitutes an acceptable answer, it is not a suitable first AI Agent project.
The research also used models to estimate that broad adoption of current AI capabilities could potentially double research output in some fields. However, enterprises should not treat this estimate as a guaranteed project benefit. The research sample covered only consumer plans during a specific period, while the accompanying survey of 528 scientists primarily involved active AI adopters in the life and health sciences. Implementation teams can therefore draw on the principle of first assessing whether a task is suitable, but they should not copy scientific research productivity estimates directly into the company’s return-on-investment projections. Instead, they should begin by asking process owners: Which tasks recur every day or every week, and for which tasks can employees provide both acceptable and unacceptable examples that the team can compare in practice?
Use Four Criteria to Select the First Workflow
Begin by reviewing the routine work performed by a single department over the past month rather than collecting requirements from across the entire company at the outset. The team can score each task using the following four criteria. The purpose of scoring is to give managers, business users, and IT a common set of questions for discussion; there is no need to rush into designing a sophisticated weighting system.
- Frequency of repetition: Does the same set of steps recur regularly? For example, Sales may organize inquiry emails every day, while Procurement may compile quotations on a recurring basis. A task must occur often enough for the team to continuously refine its instructions and validation methods.
- Completeness of inputs: Before AI Agent begins working, does it have access to all necessary emails, forms, documents, and fields? If employees still need to ask around and fill in missing information, the team should first establish clear intake rules rather than allowing AI Agent to guess what is missing.
- Verifiability of outputs: Can the person responsible clearly specify which fields the completed work must include, which sources it must cite, and which errors are unacceptable? For example, in procurement quotation comparisons, each product name, specification, unit price, and term can be checked for consistency. For customer service drafts, reviewers can verify whether the response cites company policies and addresses every question raised by the customer.
- Risk of errors: If the result is incorrect, will it merely require additional internal revision, or could it lead to external commitments regarding price, delivery schedules, or contractual terms? High risk does not mean a task can never be implemented, but the process must pause for confirmation by an authorized person before content is sent, contracts are signed, or amounts are approved.
For the first workflow, prioritize a task that occurs frequently, has complete input data, is easy to validate, and allows errors to be intercepted before anything is sent externally. For example, AI Agent could first organize the details of sales inquiries, identify missing fields, and generate a quotation draft, after which a salesperson confirms the price and delivery schedule. However, if historical data is scattered across multiple locations and quotation rules still depend on individual judgment in the moment, the data and rules must be organized first. After completing the assessment, the team should select only one task, clearly document its owner, required inputs, acceptable outputs, and the potential impact of errors, and then decide whether to begin implementation.
Clearly Define the Process, Decision Rules, and Human Approval
After selecting a task, the team must break down every step that employees currently perform based on experience. For a procurement quotation comparison, the process might include collecting all supplier documents, verifying product names and specifications, listing differences in prices and terms, flagging missing information, and finally submitting the results to Procurement for confirmation. Each step must clearly state where the data comes from, who receives the output, and what qualifies as completion. Otherwise, even if AI Agent produces a neatly formatted summary, managers will still be unable to confirm whether important terms were overlooked during processing.
Human approval must be placed at clearly defined points in the process. It is not enough to write, “A manager will review when necessary.” For example, the process can be required to pause when an amount exceeds the department’s threshold, payment or delivery terms are unclear, information from different sources conflicts, or a document is about to be sent to a customer or supplier. EgentWrX can connect recurring tasks into a workflow and require human approval before each handoff. After AI Agent completes the preceding work, the process waits for a person to authorize its release before passing the result to the next stage. The platform provides the workflow mechanisms, but the enterprise must still define which situations require approval based on its responsibilities and risk policies.
Some tasks become stalled not because of the processing steps, but because experienced employees have not clearly articulated their decision rules. In these cases, ask the process owner to choose one common type of decision and document when it applies, the required data, the decision rules, any exceptions, and examples of acceptable results. EgentWrX can structure standardized practices as skills for AI Agent to apply when relevant situations arise. If an employee-created skill is to be rolled out across the company, it must first undergo administrative review. Begin with a pilot in a single team, record which situations require adjustments to the rules, and then have the business manager or governance unit decide whether the skill should be shared more broadly.
Before implementation begins, the process owner and IT should walk through the complete process together, confirming the data sources, validation methods, responsible parties, and conditions under which the process must pause at each step. Whenever a task involves financial amounts, external commitments, ambiguous terms, or significant judgment, retain human approval. After the team has accumulated real execution records, it can decide which task to implement next.
FAQ
Does the first AI Agent project have to focus on data analysis?
Not necessarily. Data analysis was simply one of the more common use cases observed in the research. Enterprises should still begin by confirming whether a task occurs frequently, whether its input data is complete, whether its results can be validated, and whether errors can be intercepted before anything is sent. Document organization, quotation comparison, and draft responses can also be prioritized for testing if they meet these criteria.
Is cross-departmental approval suitable as the first use case?
It is not advisable to begin with a complete cross-departmental approval process, because these workflows typically involve more data sources, responsibilities, and exceptions. Instead, begin with one segment, such as organizing application materials and flagging missing information. After confirming that the necessary data is accessible and the results can be validated, evaluate whether to connect the subsequent stages of the process.
Can AI Agent handle tasks involving financial amounts?
AI Agent can first organize information, compare terms, and generate drafts. However, before prices, payment terms, or external commitments are approved, the process should pause for confirmation by an authorized person. The team should also define financial thresholds, exceptions, and return procedures in advance.
How should we begin documenting decisions when experienced employees say they rely on judgment?
Choose one type of decision that has occurred frequently in the recent past. Ask experienced employees to explain which information they review, which rules they follow, and which situations require exceptions, and have them provide examples of both acceptable and unacceptable results. The team should first test and refine the rules in a single scenario rather than rolling them out to every department from the outset.
References
- How scientists use Claude — Anthropic Research analyzes how Claude and Claude Code are used in scientific research.
