Redesign One Workflow Before You Automate It

Faster AI drafts do not guarantee better service. Learn how to redesign a returns workflow with clear ownership, human approval, and controls before automating preparation.

A customer asks to return a product. An AI assistant drafts a polished response. The agent approves it, and the ticket closes. That looks efficient—unless the response misstates the return policy, promises an unauthorized refund, or leaves the customer contacting support again.

Faster writing is not the same as better service. Before adding AI, redesign the decisions, ownership, and handoffs that determine whether a customer’s problem gets resolved.

The following hypothetical workflow shows how to do that for email-based product returns. It is a proposed operating model, not a description of a real company or its results.

Automate preparation first. Keep authority explicit, give exceptions a destination, and measure the customer outcome—not just the speed of the draft.

Start With One Narrow Workflow

For this hypothetical pilot, the scope is email requests to return recently purchased products. AI may summarize requests, retrieve approved policy material, and prepare responses. It may not send messages, authorize refunds, change orders, or close tickets.

This boundary separates preparing work from exercising authority. The NIST Generative AI Profile provides a foundation for risk-based oversight, human review where appropriate, and documented monitoring. It does not prescribe this particular workflow; the organization must define its own controls and responsibilities.

Before: The Manual Process

Customer email → shared inbox → agent review → missing-information request → policy and order lookup → eligibility decision → supervisor review when needed → agent response → documented resolution.

The weak points are not necessarily typing speed. An agent might search conflicting policy documents, discover missing information late, or send an escalation into a queue without a clear owner. Adding AI to those conditions could produce faster drafts without resolving the underlying problems.

Redesign Before Adding AI

  • Improve intake: Ask for an order reference and reason for return upfront. Do not request payment-card details.
  • Establish policy ownership: Maintain one approved, version-controlled return policy with an accountable owner and a process for resolving ambiguity.
  • Preserve case ownership: Assign a named agent who remains responsible through resolution, including during escalations.

These changes should also apply to the manual process. Otherwise, a pilot could mistakenly credit AI for improvements caused by clearer policies or better intake.

Map the AI-Assisted Workflow and Its Approval Points

The proposed workflow keeps the agent accountable while moving preparation work to AI. Each stage needs a defined output, a human responsibility, and a fallback.

Intake and Evidence Gathering

AI prepares: A request summary, suggested category, missing-information checklist, relevant approved policy passages with source identifiers, and permitted order facts.

The agent verifies: That the summary matches the customer’s request, the category is appropriate, and the order record and policy actually apply. A fluent summary is not evidence; the agent must be able to inspect the underlying sources.

Response Preparation

AI drafts: A proposed answer explaining eligibility, required information, and the next step. Unresolved questions appear separately rather than being filled in with plausible assumptions.

The agent approves: The facts, applicable policy, amounts, dates, commitments, and tone. If eligibility is unclear, the agent routes the decision instead of approving a confident guess.

Decision and Business Action

AI prepares: A suggested disposition and internal note.

An authorized person decides and acts: The supervisor approves any exception outside the agent’s authority and records the rationale. The appropriately authorized employee performs any permitted transaction through the normal business system. AI does not acquire transaction authority because it suggested the outcome.

Communication and Resolution

AI drafts: The final customer reply, case summary, and proposed outcome tag.

The agent approves and completes: The agent reviews the exact final message, sends it, confirms resolution, and selects the appropriate ticket status.

Approval must cover the wording, not merely the underlying decision. Intercom’s guidance on AI features in the inbox warns that a generated response can recast an approved goodwill exception as policy eligibility.

For example, in this hypothetical workflow, a supervisor might approve a return as an exception. If the draft says the customer qualifies under the standard policy, it changes the meaning of that decision. The agent must correct and review the final text before sending.

Give Every Exception a Destination

“Escalate when necessary” is not an operating procedure. Define where each exception goes and who remains responsible.

  • Missing order information: The assigned agent requests only the details needed to proceed.
  • Missing or contradictory policy: The policy owner clarifies the rule; the agent handles the reply manually.
  • Refund outside policy or agent authority: A support supervisor decides and documents the rationale.
  • Suspected account compromise or sensitive-data exposure: The security or privacy team handles it through the organization’s incident process.
  • Customer disputes the answer or requests a person: A human-owned support queue takes over without another automated loop.
  • Tool failure or unavailable reviewer: The case follows the manual process. An unavailable reviewer never becomes permission to send an unapproved draft.

Every escalation should include a handoff packet: the customer’s request, verified facts, policy references, attempted steps, unresolved decision, and named owner. The customer should not have to reconstruct the case for each recipient.

Enforce Boundaries Technically

Restrict AI access to necessary data and approved sources. Treat customer messages and attachments as untrusted input, not instructions that can override policy or approval requirements. Remove sending, refund, order-editing, and ticket-closing permissions from the AI’s tool access.

Retain appropriate records of drafts, approvals, policy versions, and actions under an agreed retention policy. NIST identifies data protection, monitoring, incident response, and change management as relevant governance controls. Human review should reinforce those controls, not substitute for them.

Measure Resolution, Not Just Drafting Speed

A quicker draft can still create more checking, rework, and repeat contacts. Zendesk’s guidance on customer-support metrics distinguishes active handling time from elapsed resolution time and highlights reopened tickets alongside speed. Use those distinctions, but define pilot success independently of a vendor dashboard.

  • Human effort: Total active minutes per case, including review, corrections, escalation, and rework.
  • Customer wait: Median and 90th-percentile time to the first substantive reply and final resolution. An automatic acknowledgment does not count as substantive.
  • Durable resolution: The percentage of cases reopened or followed by same-issue repeat contact within seven days. This is a proposed measurement window, not an industry benchmark.
  • Quality: Blinded human review of factual accuracy, policy compliance, completeness, and tone, using the same rubric across both processes.
  • Customer experience: Satisfaction scores, survey response rates, and complaint themes. Silence is not proof of resolution.
  • Safety: Unauthorized commitments, privacy incidents, and missed escalations.
  • Economics: Labor, AI, integration, and quality-review costs per durably resolved case.

Suggestion acceptance can help diagnose usability, but it is not proof of correctness. Evaluate outcomes after comparable follow-up periods, and do not quietly exclude unresolved cases.

Test the Redesign Before Expanding It

First, measure the redesigned manual process. Then randomly assign eligible cases to that process or the same process with AI assistance. Keep policy versions, staffing coverage, and case eligibility comparable.

Set improvement targets and acceptable quality tolerances before launch. Report case counts and uncertainty rather than presenting an inconclusive result as success. Audit a random sample alongside every complaint and safety exception.

Expand only if total effort or customer wait improves without unacceptable deterioration in quality, repeat contacts, or satisfaction. If results are unclear, extend the pilot. After an unauthorized action or sensitive-data exposure, pause the affected workflow and investigate before resuming.

Prelaunch Checklist

  • Is the scope narrow, with explicit prohibited actions?
  • Does every case have a named owner?
  • Can reviewers inspect authoritative policy and order evidence?
  • Must a person approve the exact final message?
  • Does every exception have a destination and manual fallback?
  • Are baseline measures, success criteria, and pause conditions documented?

Make the Workflow Better Before Making It Faster

The useful starting question is not “Where can AI write for us?” It is “Which decisions and handoffs must work reliably for this customer problem to be resolved?”

Choose one workflow, assign its owner, and map its approval points before enabling AI. Establish the redesigned manual baseline, then test AI as a preparation layer. Expand only when the evidence shows better service—not merely faster text.

Browse all insights · Contact Bart McDonough