AI has not understood the project context, so you add documents, revise your instructions, and ask it to answer again, only to get something much like the previous version. By the time the output is usable, you still need to verify the numbers and conclusions. If adopting AI has added these tasks, should managers ask employees to spend less time checking, or improve the information and working practices first? At Intellicon Solutions, our view is to first distinguish between work that prevents incorrect content from being delivered and work that simply repeats the same information or requirements. Only then should you decide where to save time.
According to Business Next’s coverage of the Work AI Index 2026, respondents believed AI saved them about 11 hours a week, but they also spent an average of 6.4 hours supplying information, checking answers, and correcting errors. However, high-performing users who reported improvements in both efficiency and quality spent a larger share of their time on these activities. The study was conducted by an institute affiliated with enterprise AI platform Glean and surveyed digital workers in the United States, the United Kingdom, and Australia. Respondents skewed toward frequent AI users, more senior employees, and people working in technology. The findings were also largely self-reported, so they cannot be applied directly to businesses in Taiwan. What businesses can take from the study is this principle: time spent verifying output should not automatically be treated as waste.
Keep the Number Checks; Fix the Need to Repeat Project Context
The source article uses the example of checking figures in financial statements to illustrate the value of necessary verification. By contrast, repeatedly explaining project context and asking AI to generate much the same answer can add to the workload. We recommend that businesses make the same distinction. Start with a routine report that frequently needs to be redone, and ask the person responsible to record why each addition or revision was needed, rather than simply logging how many minutes it took.
For example, when preparing a procurement quote comparison report, you could group the work into the following categories. This is a suggested approach to reviewing the workflow, not an actual customer case.
- Retain necessary quality checks: The person preparing the report needs to verify the sources of quotations, pricing units, and validity periods, and confirm that the conclusions are supported. These checks are still required during final review, even if the first draft reads smoothly.
- Find out why the same information keeps needing to be added: If employees have to supply supplier terms or the same procurement policy every time, clarify where that information is stored and which version should be used. Do not simply push employees to write longer prompts.
- Understand why work needs to be redone: If the report is missing required fields, make the format requirements explicit. If supplier terms have genuinely changed, update the information rather than treating the resulting rework as wasted effort.
After reviewing the workflow, we recommend having the relevant business manager confirm which checks must remain and who is responsible for review. For a quote comparison report, the person preparing it should be able to explain why a particular supplier is recommended. When the terms cannot be compared directly, the issue should go to a manager for judgment before deciding whether to approve the subsequent purchase. Producing the report does not mean the procurement decision is complete.
If You Keep Adding the Same Information, Check the Documents Before Changing the Prompt
If employees have to explain the project context every time, start by listing the information needed to complete the task. Then check who maintains each item, when it is updated, and who is allowed to use it. For a routine project report, meeting minutes, progress updates, and project closure criteria each serve a different purpose. If the latest progress information is missing, even the most detailed formatting instructions cannot fill that gap.
Information and procedures should also be organized separately. Information provides the basis for the current task, such as a quotation that is still valid. Procedures define the requirements to follow every time the task is performed, such as presenting different pricing units separately or listing items that need confirmation when information is incomplete. The employee responsible and their manager should agree on these requirements before giving them to an AI Agent, so that unresolved expectations are not treated as company policy.
With EgentWrX, for example, businesses can place documents in a department or project knowledge base, allowing an AI Agent to find the information needed for the task. Frequently used formatting requirements and decision rules can be defined as a skill, a standard procedure for the AI Agent to follow, or included in a reusable task card. These features help businesses organize information and procedures and have an AI Agent carry out the work. They do not mean documents will update automatically, nor do they guarantee that every output will be correct.
The platform also has a mechanism for declining to answer when it cannot find sufficient supporting information. During a pilot, we therefore recommend testing not only real work questions to see whether the AI Agent can find enough information, but also questions for which supporting information is missing, to check whether it declines to answer. If it does, first check whether the documents are incomplete, outdated, or inapplicable. Do not simply ask the AI Agent to rephrase and answer again.
A Fast First Draft Is Not Enough: Measure the Total Time to Manager Sign-Off
When evaluating the results of AI adoption, we recommend using a completed deliverable that meets business standards as the benchmark. Record all the time spent supplying information, generating the first draft, checking, revising, and rerunning the task. If an employee produces the first draft quickly but a manager then spends more time reviewing it and sending it back for revisions, that time must count too. Do not report only how quickly the first draft was produced.
Before making changes, ask the relevant business manager to establish an acceptance checklist. It might cover whether all required fields are present, whether important figures can be traced to their sources, and whether conclusions are supported. Specify which content requires managerial confirmation. Separately, the person responsible should record the total working hours spent on each deliverable, why it was returned, and how it was subsequently corrected. Time spent organizing the knowledge base and maintaining rules should also be recorded separately, rather than comparing only the time spent using AI in day-to-day work.
After making changes, compare tasks of similar difficulty using the same acceptance checklist. Switching tools should not mean lowering the standard for deliverables. If necessary checks have been retained and repeated information entry and unnecessary rework have genuinely decreased, assess whether to expand usage. If only the first draft is faster while later revisions take more time, first revisit the information and standard procedures, and incorporate recurring problems into employee training. At the next review meeting, bring a report that was sent back and discuss each reason for its return, along with who is responsible for addressing it.
FAQ
If AI output still needs human verification, can we eliminate that time?
Necessary quality checks cannot be eliminated simply to make the time savings look larger. Start by asking managers to define how important figures, cited sources, and conclusions should be reviewed. Then record tasks such as repeatedly adding context and correcting formatting separately, and prioritize addressing the reasons they keep recurring.
Can we use the figure of 6.4 hours a week spent looking after AI to estimate our company’s adoption costs?
No, it cannot be used directly to estimate working hours for a company in Taiwan. The survey covered digital workers in the United States, the United Kingdom, and Australia, and its findings were largely self-reported. We recommend using your company’s actual work records and including the time spent by both the employees doing the work and the managers reviewing it.
What should we do first if AI keeps producing unusable answers?
First, check whether the information needed for the task is complete and the rules are clear, rather than rushing to ask AI to answer over and over again. Ask the person responsible to keep the incorrect output and the documents used at the time. Determine whether the problem was outdated information, unclear field requirements, or unsupported conclusions, then decide whether to add information, revise the procedure, or refer the issue to a manager.
References
- “AI Babysitting” Comes to the Workplace: Why Are Employees Spending 6.4 Hours a Week Looking After AI—and Feeling More Overworked? — Business Next, compiled and edited by Lin Yu-hsuan, published November 11, 2026. Summarizes the Work AI Index 2026 survey by Glean’s Work AI Institute; the article notes at the end that its first draft was written by AI.



