A sales team can produce quotations faster and still take just as long to send them. The drafting step improves, but the approval queue remains. Meanwhile, more drafts arrive for the same few managers to review.
That is a plausible outcome of optimizing the most visible task instead of the process that delivers the result. AI makes the local improvement easy to demonstrate. It can also make the original constraint more expensive.
Lean provides a useful discipline here. The Lean Enterprise Institute describes it as creating needed value with fewer resources and less waste. Applied to AI investment, that means understanding which work matters, what prevents it from flowing and which steps should disappear before deciding what to automate.
The Workflow
Follow the work, including the work people hide
A procedure describes how work is supposed to happen. A real case shows what the organization does when information is missing, an exception appears or the person with approval authority is unavailable.
Follow an order, a customer request or a financial reconciliation from start to finish. Include the spreadsheet maintained outside the main system, the email asking for clarification and the correction made by someone who knows which field is usually wrong.
Those interventions may be essential to keeping the business running. Their existence does not make the employees inefficient. It often shows where the formal process relies on knowledge that has never been made explicit.
Growth exposes this dependence. A workaround that one experienced colleague could handle becomes a queue when volume increases. Automating the paperwork around that queue will not resolve the missing decision rule.
The people doing the work need to help define the problem. Otherwise, a technically neat redesign can remove an apparent inefficiency that was quietly preventing a more serious error.
The Waste
Remove unnecessary work before assigning it to AI
Every existing step should be able to justify its cost. Some approvals protect the business. Others survive because a problem occurred years ago and nobody revisited the response.
If a report has no reader, producing it faster adds little. If two teams repeatedly reconcile the same records because neither trusts the shared system, an AI reconciliation tool may be useful, but it should not prevent examination of why the records diverge.
This is where process improvement and automation can pull in different directions. An automation project needs a task to deliver. A process review may conclude that the task should be removed. Management needs to make that conclusion acceptable; otherwise, the project is rewarded for preserving the work it was supposed to improve.
THE THROUGH-LINE
I would ask the process owner to describe the desired workflow without naming any AI product. That makes the operational requirement visible before a vendor demonstration shapes the answer.
The AI Role
Give AI a specific operating role
Once the workflow is understood, AI can earn a precise role in it. It might extract information from varied documents, detect patterns across transactions, forecast demand or prepare a recommendation from scattered evidence.
These capabilities matter when the volume or variability of the work makes manual handling costly. They also need reliable inputs and a defined response to uncertainty. A demand forecast cannot correct stale inventory records on its own, and an exception alert cannot decide who has authority to change a purchase order.
For an illustrative distributor, AI could interpret incoming requests and identify possible substitute components. The approved catalog, available stock and commercial rules would still determine what can actually be offered. The process owner would decide which substitutions need expert approval.
The value would be measured in acceptable quotations reaching customers sooner, with controlled error and rework. The number of documents the model processed would explain activity, but it would not establish the result.
The Response
A faster signal is useful only if someone can respond
AI can help surface recurring delays or anomalies across more cases than a team can examine manually. That is useful observational capacity. It creates no improvement until someone interprets the signal and changes the work.
Define who receives an exception, how quickly they need to respond and what they are authorized to do. Also define when the correct response is to leave the process alone. A system that escalates every unusual case can consume more expert capacity than it releases.
Feedback needs to return to the process. If employees keep correcting the same output, the organization should examine the data, task design or decision rule. Repeated manual correction should not become an invisible subsidy that makes an automation appear reliable.
Measure the cost of responding to the system alongside the cost of running it.
The Measure
Measure from the customer’s request to the finished result
A useful scorecard follows the entire flow: elapsed time, cost per completed case, error and rework, unresolved exceptions and the quality the customer receives. Value-stream mapping is helpful precisely because it examines the connected flow of information and work rather than an isolated station.
Suppose AI halves the time spent preparing a document. If reviewers now spend the saved time checking its contents, the benefit may be improved consistency rather than lower labor cost. That can be valuable. It needs to be named honestly.
Recovered time also requires a decision about its use. It may support more volume, better service or reduced overtime. It does not automatically become a cash saving. The business case should say which result management intends to obtain.
The Call
Choose a process with room to improve and an owner who can change it
A sensible starting point has a measurable problem, usable data and a process owner with enough authority to change how the work is done. Begin with a bounded intervention and compare the resulting workflow with the original one, including failures and exceptions.
There may be no reason to use AI at the first step. Removing a redundant approval or connecting two existing systems can be the better investment. If AI is useful later, a cleaner process makes its contribution easier to test.
Before approving automation, ask the process owner to identify the work that will stop, the work that will change and the result that should improve. Those commitments give the AI team something worth building. Without them, faster execution can simply produce a larger queue.