A category-and-confidence rule decides where each support email goes in this Claude triage pipeline

One call per message does both jobs. Claude returns the classification and the drafted reply in the same structured response ^[4][6]. The shape is a JSON Schema passed directly in the request, and the post says the API guarantees the response matches it ^[12]. Truncated output, prose wrapped around the object, and a field that is a string on one call and an array on the next all go away ^[13].

The schema is where the decisions sit. `category` is one of four fixed strings: sales inquiry, support-technical issue, billing question, complaint ^[14]. `priority` is High, Medium or Low ^[15]. `intent` is a free snake_case label the model invents, such as `sso_redirect_loop_outage`, and routing does not read it ^[16]. `confidence` is a number from 0 to 100, and `confidence_tier` is computed from that number in code instead of being asked for in the response ^[17]. Moving a threshold is therefore an edit to a comparison.

Routing keys on category and confidence together ^[5]. The assignee comes from a lookup table in code, and the model never names a person ^[18]. The author's reason is specific: a model asked for an assignee can return someone who does not work there, or drift between "Sam" and "Samuel" until the routing table misses ^[19]. The rule as the post states it is to "Keep anything with real-world consequences (assigning work to a specific person, sending money, deleting data) in deterministic code" ^[20]. A human clicks Approve or edits the draft in Gmail, and nothing sends automatically ^[9].

Write access arrived through a convenience. The build started on `gmail.readonly` plus `gmail.send`, then the mark-as-read step needed write, so the author moved to `gmail.modify`, a superset of readonly ^[21]. List readonly beside modify and Google silently drops the redundant one from the actual grant; `google-auth` then raises `Scope has changed` at the next token refresh, which is well after the edit that caused it ^[22]. The author found it by reading the `google-auth-oauthlib` source after the second refresh failure ^[23].

The body extractor is the other documented failure. Its first version walked the MIME tree and, at each level, fell back to HTML when plain text was not found at that level ^[24]. Gmail messages frequently arrive as `multipart/alternative`, and the structure can nest a multipart inside another multipart ^[3].

Three external dependencies total, SQLite for state, and no web framework; the CRM view is a plain `http.server.BaseHTTPRequestHandler` serving an HTML table ^[10][11]. Re-running the pipeline cannot double-process a message, because the result is written to SQLite and the message is marked read ^[8].

The post opens by calling "AI will save you time" a claim everyone makes and almost nobody backs with numbers, and promises a breakdown of running cost against what it replaces ^[25]. The published text stops at the third bug, before any figures ^[26]. Two of the three announced bugs are written up ^[27]. For that cost comparison to transfer to another shop, you would need tokens per message, messages per day, and the hourly rate of whoever answers them now.