how to suggest support replies from a knowledge base without sending unchecked AI mail

why agents rewrite the same shipping and returns answers all day

Agents burn half their day rewriting the same paragraphs about delivery windows, return labels, and password resets. Macros help until the shop runs three shipping partners and a seasonal returns exception that no single macro covers without sounding robotic. New hires shadow seniors for weeks just to learn which article is current, while outdated Notion pages and PDF manuals sit beside the helpdesk with no link into the ticket. Customers notice the inconsistency: one agent promises fourteen-day returns, another quotes thirty because they found a different page. That inconsistency becomes a chargeback argument when finance later asks what the shop officially promised. Shops on WooCommerce add another layer of friction. The correct answer often depends on order status, fulfilment warehouse, and whether the line is a subscription renewal. Opening the shop admin, the carrier portal, and the knowledge base for every "where is my parcel" ticket is slow even for experts. Suggested replies that ignore those facts are worse than silence because they invent arrival dates. Teams that tried generic chatbot auto-sends under NDA usually rolled them back after a wrong refund promise; the durable pattern is assistive drafts with citations, not unsupervised sends. Managers then trust the queue again because every draft shows which article justified the wording. The knowledge problem is also a compliance problem. GDPR deletion answers, warranty wording, and B2B pricing exceptions must come from approved sources. If the draft cannot point to an article id, agents should not see a confident paragraph. Automation here is really retrieval plus constrained generation plus an edit gate - not a replacement for ownership of the knowledge base. Content editors need a publish path that marks articles safe for customer use per locale, including German AVV-sensitive wording and Spanish AEPD-facing privacy answers, otherwise the model will keep surfacing stale internal notes that were never meant for buyers.

when a ticket category or macro should load knowledge base draft replies

A ticket lands in a category that has a mapped knowledge collection, an agent clicks Suggest reply inside the helpdesk, or a priority below urgent waits more than a configured number of minutes without a public response. Re-runs happen when the category changes after triage or when WooCommerce order status updates while the ticket is open, because shipping answers go stale when the parcel is scanned. Batch jobs do not auto-send; they only pre-stage drafts on high-volume categories during peak hours for agents to accept or edit. Tickets marked legal, vulnerability report, or media inquiry never enter suggestion flows. Agents working night coverage can request a draft before they write a first reply so they are not guessing warehouse rules from memory. Suggestions also refresh when a linked order moves from processing to completed mid-thread, because a returns answer written against an undelivered parcel becomes wrong the moment the carrier marks delivered.

how to retrieve kb passages draft a reply and require agent approval

1. Read ticket category, language, customer segment, and linked order id; fetch order status, line items and shipping method from WooCommerce when the id resolves.
2. Query the knowledge base or vector index with a hybrid of category filters and the latest customer message, keeping only articles marked approved for customer use in that locale.
3. Build a draft reply that cites article titles or ids inline, inserts factual order fields as placeholders, and stays inside a tone guide that forbids new policy promises.
4. Write the draft to an internal helpdesk note or private draft field, never as a public comment, and notify the assignee that a suggestion is ready.
5. Require an agent accept, edit, or discard action before any public send; store which article ids were used and whether the agent edited the draft for later quality review.
6. If retrieval returns nothing relevant, post a short internal checklist of questions to ask instead of inventing an answer from the model alone.

how n8n connects helpdesk vector search and wooCommerce to draft replies

n8n receives the helpdesk webhook or button webhook, loads ticket text over REST, and enriches with WooCommerce REST when an order id is present. Approved articles sync nightly from Notion, Confluence, or a WordPress knowledge CPT into an embeddings store, with locale and segment tags. The generation step calls a model with a system prompt that lists allowed article ids and forbids policies not present in the retrieved passages. Make or Zapier can cover lighter stacks, but n8n or custom PHP usually wins when you need word-count limits, redaction, and private draft write-back on Help Scout or Zendesk. Drafts and citation metadata stay in the helpdesk; only truncated prompts leave your network. Action Scheduler can refresh article embeddings after editors publish, so agents are not suggesting last season's returns policy. Prompt logs should drop email addresses and payment fragments before storage, and retention should be short enough that support operations remain the documented purpose rather than secondary analytics on customer tone.

what happens when kb retrieval is empty or the model invents a policy

Empty retrieval yields an internal checklist template, not a model essay. If the model cites an article id that was not retrieved, the workflow discards the draft and alerts content ops that a hallucination guard fired. WooCommerce timeouts still allow KB-only drafts with a warning that order facts are missing. Rate limits on the model provider fall back to the top three macro snippets for that category. Agent discard reasons are mandatory when the draft is wrong so editors know which article to fix. Public-send webhooks without an accept token are rejected. When embeddings lag behind a same-day policy change, editors can force a reindex flag on the article and the next suggestion run prefers lexical search until vectors catch up, avoiding drafts that quote yesterday's holiday cutoff.

where generative AI belongs in suggested replies versus hard macros

Generative AI drafts wording and selects among retrieved passages; it does not invent shipping promises, refund amounts, or legal positions. Temperature stays low, the prompt includes only retrieved chunks plus structured order facts, and the output is a suggestion for a human. Classification models may choose which collection to search, but policy truth lives in the knowledge base. If you cannot name an owner for article freshness, do not enable AI drafts - fix macros and ownership first.

when canned macros beat AI suggested replies for support teams

Your knowledge base is a pile of conflicting PDFs with no approver. Volume is under a few dozen tickets a week and macros already cover ninety percent of asks. You want fully automatic public replies on day one without an edit gate - that is a chatbot project with different risk, not this use case. Highly regulated medical or financial advice lines need specialist tooling beyond a general support draft flow.

which metrics show suggested replies reduce handle time safely

• Median agent handle time on tickets where a suggestion was shown versus matched controls
• Share of suggestions accepted with only minor edits versus discarded
• Count of hallucination-guard discards per week
• Percentage of drafts that cite at least one approved article id
• Knowledge article edit latency after agents flag a wrong draft