Suppose the sales department deploys an AI Agent to read customer requirements, query a database, and generate a draft quotation. A quoting task that once took 40 minutes can now be completed in eight. Over a year, the hours saved are substantial, and the project owner has concrete before-and-after figures to demonstrate the results.
Yet customers may not notice any difference.
What customers care about is how long they wait between sending an inquiry and receiving a quotation. That time can be divided into two broad categories: the time someone spends actively working on the request, and the time the documents sit somewhere waiting for the next person to take over. Deploying AI at a single step usually reduces the first. If the second accounts for most of the total, making one step faster will do little to accelerate the overall workflow.
Actual processing time may be shorter than you think
The lean manufacturing classic Learning to See includes a representative value-stream example. At a stamping plant, the actual processing time required to manufacture a part totaled just 188 seconds, yet the part took 23.6 days to move from incoming raw material to finished-goods shipment. Processing accounted for only about 0.01% of the total lead time. The rest was spent waiting or sitting in queues.
Knowledge work follows a similar pattern. In 2023, Asana surveyed 9,615 knowledge workers across six countries. Respondents reported spending an average of 58% of their working time on “work about work,” such as coordinating tasks, following up on progress, and searching for information.
Communication and coordination are unavoidable. The problem is that the step a company chooses to accelerate with AI may not be the most time-consuming part of the overall workflow. Focusing only on the minutes saved in a single task can easily lead companies to overestimate AI’s impact on the process as a whole.
Map the quoting workflow to see where time gets stuck
A typical manufacturing quotation passes through seven steps:
- Sales intake: Receive the customer’s drawings, quantities, delivery requirements, and material specifications, then check whether all attachments are included.
- Engineering review: Interpret the drawings, tolerances, materials, and process constraints, then list any items requiring clarification.
- Procurement inquiry: Request quotes from suppliers for materials and outsourced processes, then consolidate their responses.
- Production scheduling estimate: Check equipment capacity, tooling status, work in progress, and the production schedule.
- Cost consolidation: Combine material, processing, outsourcing, packaging, and transportation costs.
- Management approval: Review margins, pricing exceptions, and payment terms.
- Sales delivery: Apply the quotation template, confirm the version, and send it to the customer.
The workflow looks straightforward, but once you follow an actual quotation through it, you find that every handoff can create delays.
Attachments must be forwarded manually, and when engineering sees them often depends on whether someone follows up. After engineering completes its review and places the results in a shared folder, procurement and production planning must each wait to be notified. Supplier responses may sit in an inbox before anyone consolidates them. Production planning cannot begin until both engineering and procurement have provided their results. If even one piece of information is missing, the entire quotation stalls until someone follows up. The pricing approval then sits in a manager’s inbox until they have time to review it. Once approval comes back, sales still has to convert the information manually into the official quotation format.
Here, handoff cost refers specifically to the time during which information is already back in the company’s hands but no one has begun working on it. External delays, such as waiting for a customer to provide missing documents or a supplier to respond, are excluded because the company cannot shorten them directly.
A faster individual step may simply reach the queue sooner
Consider the opening example. Even if an AI Agent can complete the engineering review or cost consolidation quickly, the existing delays will remain as long as attachments are still forwarded manually and approval requests continue to sit in inboxes.
Engineering may finish earlier, but if the next step has not started, the work has merely entered the queue sooner. From the perspective of one department, processing time has clearly fallen. From the customer’s perspective, the wait after submitting an inquiry may remain unchanged.
In practice, the cases that make customers feel lead times are unreliable or excessively long are often the small number that require escalation for approval or several rounds of additional documentation. These cases are not necessarily the most difficult to process, but they may repeatedly stall at several handoff points. To identify the problem, companies must first record the waiting time.
How to measure handoff costs consistently
The method is straightforward. At each stage of the workflow, record four timestamps: receipt, start of processing, completion, and handoff. The difference between receipt and the start of processing is the waiting time at that stage.
With these records, companies assessing an AI deployment will no longer mistake the time saved at one step for an improvement across the entire workflow. Managers can see which stage has the longest wait and which cases repeatedly get stuck. They can then determine which repetitive tasks an AI Agent should complete “in advance.”
For example, as soon as a case arrives, an AI Agent can check the attachments and conduct a preliminary review. When the next stage takes over, the person responsible does not have to begin by organizing the information and can work directly with content that has already been prepared. The savings then extend beyond one person’s processing time to include downstream waiting and repeated clarification.
The three deployment models require different success criteria
In a 2025 study, McKinsey used customer service case resolution to compare three approaches to AI deployment. Providing AI assistance at a single step improved resolution time and productivity by approximately 5% to 10%. Connecting AI Agent to an existing workflow saved 20% to 40% of the time and reduced case backlogs by 30% to 50%. Redesigning the workflow around tasks that AI Agent could handle autonomously shortened resolution time by 60% to 90%.
These figures come from a single use case and cannot be applied directly to every company. They do, however, clearly illustrate the differences among the three approaches. Each requires a different level of investment and a different degree of process change, so they should not be evaluated against the same criteria.
If a company deploys AI only to assist with one task, a reasonable goal is to reduce the processing time for that task. If it expects improvements at the workflow level, it must address how work is handed off between upstream and downstream steps, and may even need to redesign the entire process.
AI Agent must be able to hand work off for the workflow to become faster
For multiple AI Agent instances to pass work to one another, the output from one AI Agent must serve directly as the input for the next. They must also access the same body of enterprise knowledge and share permission and audit mechanisms, allowing managers to monitor the cost, quality, and exceptions at every stage.
A platform can provide the foundation for connecting the workflow, but end-to-end management still requires someone who is accountable for the results of the entire process. This process owner must continuously track two things: the total time from receiving the customer’s request to sending the quotation, and the stage where the case is currently stuck.
Returning to the quotation example, the most important metric for evaluating the deployment is the customer’s total wait from inquiry to quotation, not the processing time at any single stage.
This article is excerpted from Intellicon Solutions’ Agent-Ready: AI Applications for Taiwan’s Manufacturing Industry. The full white paper explains how companies can progress through three stages, from individual tasks to cross-departmental workflows, and includes a 12-question self-assessment.