Let’s talk
GUIDE

How do you automate a support inbox without sending a wrong reply?

Let the automation sort each message and draft a reply, and keep a person as the one who presses send. Messages about money, refunds, legal or health matters, and anything the model is unsure about, skip the draft and go straight to a person. If the model returns something malformed or empty, treat it as a failure and route the message to a person; never send it.

Wrong replies usually come from a draft that invented an order number, a date or a policy. The drafting instructions forbid those, and a check in code runs after the model, because the model is not trusted to obey. Sending without approval can come later, one category at a time, once the team trusts the drafts in that category.

Updated Sep 28, 2026

How it works

  1. Sort into a few categories

    Read last month's inbox and name five to eight categories, each with an urgency and an owner. Billing, refunds and complaints are their own categories.

  2. Draft only from approved help content

    The model receives the help articles and policies that apply, and is told never to invent an order number, a date, a refund amount or a policy.

  3. Check the model's answer in code

    Ask for a strict format: category, urgency, a one-sentence summary, the draft, and whether a person is needed. Anything malformed goes to a person instead of becoming a blank or broken reply.

  4. Send sensitive cases straight to a person

    Billing, refunds, legal, health, complaints and anything the model flagged skip the draft. A person writes those replies.

  5. A person approves every send

    The reviewer edits, approves or rejects. Approval is the only thing that sends, and each message can be approved once.

  6. Widen slowly

    Track per category how many drafts are approved unchanged, edited or rejected. Only a category with steady unchanged approvals is a candidate for sending on its own.

What it costs to run

Running cost has two parts: the helpdesk where the team works, usually billed per user, and the model that sorts and drafts, billed per token by the model provider, or per resolved conversation when the helpdesk's own assistant answers customers. These are list prices from the vendors' pages; volume and message length set the total.

ItemList priceSource
Help Scout Standard plan (helpdesk)US$25 per user per month; the page lists a discount for annual billing. A free plan covers 5 users and 1 inboxHelp Scout, Help Scout Pricing: Affordable Customer Support SoftwareRead on Sep 28, 2026
Help Scout AI AnswersUS$0.75 per resolution (a conversation resolved without a person)Help Scout, Help Scout Pricing: Affordable Customer Support SoftwareRead on Sep 28, 2026
Intercom Fin AI Agent on an existing helpdeskUS$0.99 per outcome, no seat cost, with a monthly minimum (Intercom gives 50 outcomes as an example)Intercom, Intercom Pricing | Plans for every team sizeRead on Sep 28, 2026
Claude Haiku 4.5 through the Anthropic APIUS$1 per million input tokens, US$5 per million output tokensAnthropic, Claude API Docs, PricingRead on Sep 28, 2026
Anthropic's own worked example: 10,000 support conversations of about 3,700 tokens each on Claude Haiku 4.5About US$37 in total (Anthropic's estimate)Anthropic, Claude API Docs, PricingRead on Sep 28, 2026

Third-party list prices, read on the date shown. They are not Betterlane prices.

What we learned building it

From our own workflows, September 2026: a support-inbox demo run on invented messages, and the reply triage that runs Betterlane's own lead handling on n8n Cloud, tested on our own seed addresses. Neither is a client system.

Distrust the model's answer
Our support-inbox demo asks the model for strict JSON. A malformed answer becomes a failure routed to a person, never a blank reply, and our test file checks that scenario along with the others.
Money never gets an automatic draft
Anything about billing takes the lane where a person writes the reply. In our own reply triage, code also sends a draft to that lane if it contains money, a number or a link to another site.
Approval is the only thing that sends
An approval for a message that is not in the queue is refused with a reason. In our own reply triage, a second approval of the same reply was refused.
Test the approval step too
The first version of our reply triage read the approval form wrongly, because the answers were keyed by the field's label. That run was declined and the fix re-tested before use.

When it is not worth it

  • If the inbox gets a few messages a day, saved replies in the helpdesk do the job.
  • If most messages need decisions the automation cannot see (refund approvals, account changes, custom quotes), the drafts will mostly be rewritten.
  • If nobody has written down the policies, write them first: the drafts can only be as right as the help content behind them.

Questions

Can it send replies on its own?

It can, but start with approval on every reply. Move one category at a time to automatic sending, and only once the team has seen enough drafts there approved without changes.

Do we have to change helpdesk?

Usually not. The workflow can sit beside the current helpdesk through its API or email forwarding, so the team keeps working where it does now.

What always stays with a person?

Money, refunds, legal and health matters, complaints, and anything the model marks as uncertain.

How do we know it is working?

Count, per category, the drafts approved unchanged, edited and rejected. A rising share of rejections in one category means the help content for it needs fixing.

Want sorted tickets and drafts your team approves before anything is sent?

See the service: Support inbox automationLet’s talk

All guides