Step 1 of 10

Define routing policy and risk boundaries

Design the decision rules that determine which emails can be auto-answered and which must be escalated to a human before any reply is sent.

Why this matters

Routing policy is the skeleton of the entire system. Every other step, classification, drafting, escalation, exists to execute this policy. If you build the automation before defining the rules, you will end up with a system that auto-sends emails you never intended it to handle. Write the policy first, then build around it.

Build instructions

Identify your inbox intent categories

Step 1
Open your email inbox and review the last 50 messages. Group them into categories based on the type of response they require. Common categories: Inquiry (general questions about your service), Follow-up (responses to previous conversations), Support (technical or account problems), Billing (payment, invoice, subscription questions), Legal (contracts, disputes, compliance requests), Unknown (anything that does not clearly fit).
Step 2
Create a new Google Sheet. Name it 'Autoresponder Config'. In Sheet 1, create a table with headers: Intent Category | Route Action | Min Confidence | Hard Stop | Notes.
Step 3
Add one row per category. You will fill in Route Action and Min Confidence in the next sub-step.

Assign a route to each intent category

Step 1
For each intent category, choose one of three route actions: AUTO-SEND (AI drafts and sends without human review), DRAFT (AI drafts but a human must approve before sending), ESCALATE (immediately forward to a human, no AI draft).
Step 2
Recommended starting point: Inquiry = AUTO-SEND (confidence ≥ 0.80). Follow-up = DRAFT (confidence ≥ 0.75). Support = DRAFT (confidence ≥ 0.70). Billing = ESCALATE always. Legal = ESCALATE always. Unknown = ESCALATE always.
Step 3
Mark Billing, Legal, and Unknown as hard stops in the 'Hard Stop' column with 'YES'. These routes apply regardless of confidence, no AI output is ever sent for these categories.

Write rationale for each policy row

Step 1
In the Notes column, write one sentence explaining why you chose that route for that category. Example: 'Support → DRAFT: customers report urgent problems, an auto-send reply risks dismissing legitimate issues that need investigation.'
Step 2
This rationale is not just documentation, it forces you to think about whether your rule actually makes sense. If you cannot write a clear justification, the rule is probably wrong.
Step 3
Share this sheet with any stakeholders who will be affected by automated replies. Get sign-off before building the automation around it.

Common mistakes

Setting all categories to AUTO-SEND with a high confidence threshold. High confidence does not eliminate errors, it just makes errors less frequent. For sensitive categories like billing or legal, no confidence threshold is high enough.
Not defining 'Unknown' as a hard stop. When the classifier encounters an email it has not seen before, it will guess. Unknown intent + auto-send = unpredictable output sent without review.

Pro tips

Start more conservative than you think you need to be. It is easy to relax a DRAFT to AUTO-SEND after you have seen the system work well. It is very hard to recover trust after a bad auto-sent email.

Before you continue

Your policy table has at least five intent categories, each with a route action, minimum confidence (or 'N/A' for hard stops), and a one-sentence rationale. At least two categories are marked as hard stops.

Step result

You have a written, approved routing policy that governs what the automation is allowed to do. Every subsequent step builds on this policy.