Redesign an AI-assisted workflow by following one request all the way to an accepted result, then changing the steps, information and responsibilities that determine whether it gets there. Compare removing work, standardizing it and using ordinary automation before deciding where AI belongs. Test the receiving person's work as carefully as the generated output.
For a process owner, the important question is what will change after the draft appears. A faster summary can leave the same missing information, approval queue and manual re-entry behind it. This article shows how to examine those dependencies and make a bounded change. The tools and workflows guide covers the broader choice of systems and support.
For a transfer between colleagues using different tools, define the sources, review status and acceptance needed at the handoff.
Trace one request to its accepted result
Suppose a wholesale distributor receives an email asking to change the quantity and delivery date of an order that has not shipped. A customer-service coordinator checks the request, a fulfillment planner determines whether the change is feasible, and the coordinator updates the order and confirms the agreed change to the customer.
The endpoint is more demanding than “draft a reply.” The approved change must be recorded where fulfillment staff will act on it, and the customer must receive an accurate confirmation. An order that has already shipped follows a different process and stays outside this trial.
Follow a recent case with the people who sent and received it. Ask them to show the actual records rather than describe the ideal procedure. Include a case that went wrong or needed clarification.
Record:
- The trigger and endpoint. What starts the request, and what evidence shows it is complete?
- The information. Which original message, order record and operating rules does each person need?
- The handoffs. Who receives the work, what they check and how they accept or return it.
- The delays and repeats. Where the case waits, gets retyped or returns for missing details.
- The authority. Who may change the process, approve the order change and stop the trial?
OpenAI Academy's recurring-workflow worksheet offers a useful starting structure for scope, current work and responsibility. Its readiness recommendation explicitly does not authorize building or deployment. A completed map still needs the people who do and own the work to confirm it.
If you do not yet have a specific request to trace, start with discovering AI use cases from everyday work. “Improve customer service” is too broad to test as one workflow.
Compare changes before choosing AI
In the distributor example, suppose the coordinator copies the email into a spreadsheet, sends a separate summary to the planner and later copies the planner's answer into the order system. An AI-written summary may help, but it leaves the duplicate records and the planner's missing information untouched.
Consider each kind of change on its own merits:
| Change | Application to the order request | What must be checked |
|---|---|---|
Remove unnecessary work | Retire a spreadsheet entry that only duplicates the approved order record | Confirm who uses it and preserve any required history or control |
Standardize the input | Give coordinators one required set of order and change fields | Keep a route for customers whose emails lack those details |
Use established automation | Validate required fields and route a complete case to the right queue | Test routing rules, permissions and failed updates |
Add AI assistance | Extract the requested changes from varied email wording into a draft | Link back to the message and require a person to resolve ambiguity |
Retain human judgment | Have the planner accept, reject or revise the proposed fulfillment change | Make the decision and its conditions visible to the coordinator |
These choices can coexist. You do not need a model to check whether an order number field is empty. You may benefit from one when customers describe the same request in many different ways.
The UK Government's AI Playbook makes the same basic technology-selection point: established technologies may solve a problem better than AI. It is government guidance, not a performance guarantee for a private company's redesign.
Start with the smallest change that addresses the observed problem. If the planner's queue lacks cover when someone is absent, better extraction will not provide that cover. The process owner needs to resolve the staffing or routing question.
Design what the receiving person can accept
A handoff needs more than a notification. The receiving person needs enough information to act, a clear decision to make and somewhere to send an incomplete case.
For the order-change trial, the coordinator's packet should contain the original customer request, the verified order identifier, the specific requested changes and any unresolved facts. The planner should not have to reconstruct the request from a polished paragraph with no source attached.
Information preparation may matter before any new automation. In a September 2026 account of seller enablement, Mark Stephen Ware described curating internal source materials for an onboarding initiative so that the collection was current, authoritative and relevant to the seller's task. Sellers and managers still had to interpret the customer context and validate the response. His account is a practitioner's self-report, not an independently measured productivity result.
For the distributor, apply that question to the order record and fulfillment rules. Which version is authoritative? Can both colleagues access it? What happens when it conflicts with the email? Resolve those issues before asking AI to combine the information.
A proposed route for one order change
Keep one case record from intake to confirmation.
Prepare the request
The coordinator verifies the order and checks AI-extracted changes against the original email.
Decide feasibility
The planner accepts, rejects or revises the change, recording any conditions in the same case.
Record the decision
The authorized coordinator updates the order and confirms that the saved record matches the decision.
Confirm and close
The coordinator sends an accurate customer confirmation and closes the case with its evidence attached.
Return incomplete cases visibly
Missing or conflicting information returns to the coordinator. A hold remains visible until resolved; it does not silently close the request.
For each changed handoff, agree the acceptance check, owner and exception route. “Human review” is too vague. In this example, the coordinator checks extraction against the message; the planner checks feasibility against the current order and fulfillment information. Those are different judgments.
Test the changed route beyond the draft
Run approved cases through the complete proposed process. A demonstration that stops before the order update cannot establish that the handoff works.
Use representative cases, including important failure conditions:
- A complete ordinary request. Check whether the planner can decide without requesting information already available upstream.
- A missing or conflicting detail. Verify that the coordinator receives a visible unresolved case and can recover the original source.
- A request outside scope. Confirm that an already-shipped order leaves the trial route and reaches the existing process.
- A repeated request or failed save. Check that retrying does not create duplicate changes or falsely report success.
- An unavailable reviewer or tool. Demonstrate who takes over and how an open case remains accounted for.
Compare the old and proposed routes using the same definition of completion. Measure elapsed time, all involved people's effort, accepted outcomes, returns for correction and unresolved cases. Use the net time savings guide when defining effort, including review and rework.
Treat gains as a hypothesis. In Dillon and colleagues' study of an individual Copilot rollout, changes in email work did not come with comparable changes in meeting activity. The authors discuss the coordination needed to change shared work. The experiment did not test this redesign procedure, and application records did not directly establish output quality.
A full-process trial also needs capacity. In a Reddit discussion about introducing agents, one user recommended running human and agent paths together before switching. Another Reddit user challenged that advice, saying the operators they spoke with lacked time to do the work twice and could leave agent outputs unused. That is an unverified account, but it exposes a useful planning question: who will do the comparison work, and when will it end?
Budget that work explicitly. A controlled set of approved past cases may help test the route before a bounded live trial. The appropriate coverage depends on the workflow's risks and variation; a fixed number of weeks or a universal accuracy threshold does not settle readiness.
Retire duplicate work with an explicit cutover
An improved route can still add workload if everyone keeps maintaining the old one indefinitely. Before expanding the trial, agree which old steps will end, which controls remain and what would trigger a return to the fallback.
| Decision | Distributor example |
|---|---|
Authoritative record | The approved order in the order system, with the change case linked to it |
Work to retire | The separate status spreadsheet, only after its users and required records are accounted for |
Control to preserve | Planner acceptance before an authorized order update |
Fallback | The existing manual route, with open trial cases reconciled before switching |
Owner and review | The process owner checks results with service and fulfillment leads and authorizes the next scope |
For the detailed acceptance and retirement decision, use the guide to ending old work after an AI workflow is accepted.
Distinguish retaining a fallback from requiring everyone to operate both routes on every case. A fallback must be usable when needed; routine duplication should have a defined purpose and endpoint.
Do not remove an approval merely because it is slow. Establish what it protects, who has authority to change it and how that protection would be maintained. Some delays are avoidable coordination work; others reflect a decision the business still needs.
The first change may be modest: one case record, clearer required information and fewer requests sent back. Keep or expand it only if the receiving team's work and the final outcome improve. The guide to the gap between personal productivity and company results explains why a local time saving alone cannot establish wider value.
Questions about AI workflow redesign
How is workflow redesign different from task automation?
Task automation changes how a particular activity is performed. Workflow redesign changes how the complete result is produced, including the sequence, information, handoffs and responsibilities. An AI draft can be part of either. Trace one recent case to its accepted endpoint, then use the comparison of change options to decide whether the main problem needs removal, standardization, established automation or AI assistance.
Should AI workflow redesign remove human approvals?
Only an authorized owner can decide whether a particular approval should change. First establish what the approval protects and whether the proposed route preserves that control. AI drafting does not itself justify allowing AI to approve or commit the result. In the distributor example, planner acceptance remains necessary even when extraction and routing improve. Record the decision, reviewer and conditions before testing a changed approval route.
What should happen when the AI output is incomplete?
An incomplete output should leave the routine path and reach a named person with the original information and unresolved issue attached. Keep the case visibly open until it is resolved or explicitly closed under the existing process. Avoid guessing missing facts or treating a notification as a completed handoff. The proposed order-change route shows how the coordinator receives such exceptions.
How do you know when to stop running the old process?
Stop routine duplication when the authorized process owner accepts evidence that the new route meets the agreed outcome, control and support requirements. Reconcile open cases, preserve required records and establish a usable fallback before switching. If the evidence is weak or comparison work has no owner, narrow or revise the trial. Use the cutover decisions to make that choice explicit rather than leaving both routes running indefinitely.
