An employee cuts the time needed to draft a customer reply from 15 minutes to 5. Their manager then spends another 12 minutes checking what the AI has promised the customer. Has the business actually saved time?
These are hypothetical figures, but they expose a useful question to ask before buying more tools. Drafting may become faster while review, approval and accountability stay in the old process. If every piece of work still waits for one manager, individual speed does not necessarily become team speed.
The task is to design the work people do together. Who may use AI to prepare information? Who may decide without further approval? What requires escalation? How will the business know whether results have improved once checking and corrections are included?
A manager beginning to use AI does not need to answer these questions for every department at once. They should be able to answer them for the first workflow they intend to trial.
Individual speed does not establish team results
Generative AI at Work, published in The Quarterly Journal of Economics in 2025, studied an AI assistant introduced to 5,172 customer-support agents. Issues resolved per hour increased by 15% on average, with larger gains among newer and initially less-skilled workers. This was a result in the setting studied, rather than a rate every business can expect after adopting AI. Read the published research
For a team leader, the variation behind that average matters. A new employee may gain access to useful response strategies they have not encountered before. An experienced employee may need a different kind of assistance. Requiring everyone to use AI equally and raising everyone's target by the same percentage may therefore misread what is happening.
A separate six-month experiment covered 7,137 knowledge workers in 66 firms using AI within email, meeting and writing applications. In the experiment's second half, treated workers who used the tool spent about two fewer hours on email each week. The researchers did not detect changes in the quantity or composition of tasks from providing the tool to individuals. Shifting Work Patterns with Generative AI
That study is a working paper, and some authors worked for Microsoft, the tool's producer. A result that was not detected does not prove that organizations cannot change. The narrower interpretation is that individual time savings alone do not establish that an entire workflow has changed.
An employee may write a document faster while still waiting for the same pricing information, the same manager and updates scattered across messaging groups. The delay may simply move elsewhere. Managers need to follow the work from the initial request to resolution, including the stages where AI contributes little.
Define responsibility before adding approval points
"You can use AI, but someone must check it" is too vague for daily operations. Reviewers need to know whether they are checking factual accuracy, the suitability of a response or a commitment the business will have to honor.
An employee may verify names and document numbers correctly without having authority to approve a discount. A sales manager may approve a price without being responsible for deciding whether customer information may be sent to a particular AI tool. A system administrator may configure access without owning the business decision.
At a minimum, identify the owner of the workflow, the person authorized to make the relevant decision and the person responsible for conditions governing data use. One person may hold several roles in a small business. Naming the roles still helps people understand which rule to revisit when something goes wrong.
NIST's AI Risk Management Framework 1.0 calls for clear roles, responsibilities and communication concerning AI risk, including differentiated responsibilities for human-AI arrangements and oversight. It is a voluntary framework rather than Thai law and does not prescribe one approval matrix for every company. NIST AI RMF Core, Govern 2 and Govern 3
The practical step is to replace the catch-all phrase "someone checks" with a description of what each person checks, what they may approve and when they escalate.
Start with the consequences of a decision
Consider a customer-service team at an equipment distributor. The following is a proposed workflow design, rather than a mandatory standard or a confirmed process used by Enersys or any customer.
| Work | What an employee may do | When to escalate | What must be checked against an actual source |
|---|---|---|---|
| Draft an answer from an approved company manual | Use AI to draft and send after completing the agreed checks | The manual does not cover the issue, or the answer introduces information with no identifiable source | The current manual and relevant product details |
| Communicate shipping status | Use AI to phrase information confirmed in the system | The customer requests a delivery commitment that the responsible person has not confirmed | Order status and information from the person responsible for delivery |
| Propose a discount or refund | Prepare facts and options within policy | The request exceeds the delegated amount or conditions | Approved pricing, sales terms and approval authority |
| Handle a complaint involving potential loss | Gather information and draft an acknowledgment | There is a significant claim, dispute or conflict between facts | Original documents and the responsible person's assessment |
Approval needs usually follow the consequences for the business, rather than the length of a message. A single sentence drafted by AI may lead a customer to believe the company has given a guarantee.
Conversely, if an employee works from an agreed manual, can check the source and stays within delegated authority, having a manager read every word may add a queue without improving the checks that matter. Managers should relate review points to consequences and decide how to sample work employees may send themselves.
High-impact work or decisions that are difficult to reverse may require expert review every time. A workflow with limited consequences might begin with every item reviewed during a trial, then change in response to the errors found. There is no universal sampling rate suitable for every task.
Escalations should also arrive with enough information to make a decision. What does the customer want? Which facts have been checked? Which condition is blocking the work? What does the employee recommend, and when is an answer needed? If the manager must reconstruct the entire case each time, approvals will remain slow even with AI helping to draft.
