AI adoption and customer service
AI customer reply drafts: build a review workflow before automating send
AI can help a service team prepare a reply, but a fluent draft is not a verified answer. The useful first step is usually not automatic sending. It is a controlled draft-and-review workflow that gives the employee the original enquiry, approved evidence, a clear checklist and an easy way to reject or escalate the draft.
By SVS Digital · Published
Start here
Start with one frequent, low-consequence enquiry type. Limit the AI to approved knowledge and the minimum customer data needed for that task. Keep every draft unsent until an authorised employee checks identity, facts, policy, commitments and tone. Route complaints, money, legal or safety issues, uncertain identity and missing evidence to a named specialist. Test representative and difficult cases, record corrections, and allow automatic sending only as a later, separately approved decision supported by evidence.
1. Choose one reply type and define the finish line
Do not begin with “answer customer messages.” Choose a bounded case such as explaining a published return procedure, acknowledging a complete service request or drafting an order-status reply from verified fields. Write the trigger, permitted sources, expected output and action that ends the workflow. A draft ready for an employee is a different finish line from a message sent to a customer.
Microsoft describes customer-service Copilot features as assistance for representatives and says generated content is not intended for use without human review or supervision. Its documentation also notes that knowledge-based answers depend on accurate, current knowledge articles. That makes source ownership part of the workflow: name who updates each policy, price list or procedure and what happens when no current source exists.
- One enquiry type with a clear trigger and owner.
- Approved source documents and their update owner.
- An explicit output state: draft, approved, escalated or sent.
- A stop rule when evidence is missing or contradictory.
2. Give the workflow only the context it needs
Map the fields needed for this reply type. An order-status draft may need the verified order reference, current fulfilment state, approved delivery wording and customer language. It probably does not need the customer's full purchase history, internal notes from unrelated cases or every CRM field. Keep identity checks outside the model where a deterministic system can perform them.
The ICO's AI audit guidance recommends checking that information is adequate, relevant and limited to the purpose, mapping where personal information is used, and reviewing retention and access. Apply that principle to the prompt, retrieved documents, logs and reviewer screen. The correct data design depends on your legal basis, contracts, jurisdiction and systems; this guide is not a compliance assessment.
- Required fields are listed and unnecessary fields excluded.
- Access follows the employee's existing permissions.
- Prompt, draft, correction and log retention are defined.
- Privacy and security owners review the real implementation.
3. Put enquiries into draft, specialist and no-AI lanes
Create a green lane for ordinary cases with a current source and reversible outcome. Use an amber lane for complaints, refunds, unusual exceptions, vulnerable customers, disputed facts or messages that require a supervisor. Reserve a red lane for matters the workflow must not draft or act on, such as credentials, threats, legal claims, safety incidents or decisions that your policy assigns to a qualified person.
Do not ask the model alone to decide whether its own answer is safe. Use deterministic rules where possible, let employees change the lane, and treat customer text and attachments as untrusted input. Microsoft warns that content from emails, social posts or forums can contain prompt injection intended to change output or actions. A customer instruction must never override system rules or authorise a business action.
- Green cases have current evidence and no new commitment.
- Amber cases have a named queue and response target.
- Red cases stop before drafting or system action.
- External text cannot alter permissions, policy or tools.
4. Design a review screen that supports a real decision
Show the original enquiry, customer identity status, facts used, source links, draft and risk lane together. The reviewer should be able to accept, edit, reject or escalate without hunting through several systems. Require a fresh check when an edit introduces a price, deadline, refund, contractual statement or other commitment not supported by the displayed evidence.
Use a task-specific checklist: correct person and case, factual match to source, complete answer, no unsupported promise, appropriate tone, necessary personal data only and correct next step. NIST's Generative AI Profile identifies confabulation, data privacy and human-AI configuration as risks and recommends defined responsibilities for oversight. “A human is involved” is not enough unless that person has evidence, time and authority to stop the message.
- Original message, evidence and draft are visible together.
- Reviewer can reject and escalate as easily as approve.
- Commitments require explicit source or specialist approval.
- Approval records the responsible person and final version.
5. Test the workflow with ordinary and difficult cases
Build a small test set from properly authorised, redacted examples. Include clear requests, missing identifiers, outdated policies, conflicting records, multiple questions, spelling errors, emotional complaints, unsupported refund requests, malicious instructions inside the message and cases that must escalate. Write the expected lane and minimum acceptable reply before running the test.
OpenAI's evaluation guidance recommends task-specific tests that reflect real use, typical cases, edge cases and adversarial inputs, with human feedback used to calibrate scoring. For this workflow, useful checks include correct routing, source-supported facts, prohibited commitments, personal-data exposure and whether the reviewer can detect the problem. Keep critical failures visible instead of hiding them inside one average score.
- Expected lane and acceptable evidence are labelled in advance.
- Typical, incomplete, multilingual and adversarial cases appear.
- A domain owner reviews failures and disputed labels.
- Changes to sources, model or instructions trigger regression tests.
6. Run a draft-only pilot before considering more automation
Begin with a small team and keep sending under normal employee control. Measure the whole job: time to an approved reply, substantial corrections, wrong routing, escalations, missing-source cases and reviewer workload. Read a sample of accepted and rejected drafts. A fast first draft has little value if employees must reconstruct facts or become less attentive because the text sounds convincing.
Set pause conditions for an unsupported commitment, wrong-customer data, missed high-risk case or repeated source failure. Automatic sending is not the default next step; it is a new risk and permission boundary requiring its own use-case review, evidence and rollback path. Some sectors or messages may require specialist legal, privacy, security or regulatory review regardless of pilot performance.
- Draft-only scope, users, duration and owner are recorded.
- Quality and review effort are measured with the same definitions.
- Critical incidents trigger a pause and investigation.
- Expand, revise or stop is a documented business decision.

