Changelog

What we shipped, when, and why.

A running record of how support-arc is becoming itself. Public, terse, dated. No marketing fluff — if it isn't real, it isn't here yet.

0.9.0 April 22, 2026 cohort beta

The first six teams are live.

After three months of design partner work, we opened to a six-team beta cohort. Onboarding tightened from two weeks to seven days. Most of what we shipped this release is the result of pushing real conversations through the system at production volume for the first time.

  • New

    Cohort onboarding flow. A guided seven-day setup: connect inbox → import knowledge → review the first 100 AI drafts → set confidence thresholds → go live. Each step is a checkpoint we walk through with the team.

  • New

    Reasoning trace. Every AI reply now exposes its sources, the policies it consulted, and a one-line summary of why it answered the way it did. Click any reply in the team view to expand the trace.

  • Improved

    Journey timeline density. We were showing too many micro-events. The timeline now collapses sequences of similar actions and surfaces only the moments a human would actually want to see.

  • Fixed

    Email threading across forwarded chains. Forwarded support emails were creating new journeys instead of attaching to the existing one. They now reattach correctly.

0.8.0 March 11, 2026 confidence

Confidence becomes a first-class control.

We rebuilt the AI's output to expose confidence per reply, and turned it into the operational primitive teams use to decide what auto-sends and what gets reviewed. Single biggest behavior change since we started.

  • New

    Per-topic confidence thresholds. Set the auto-send threshold differently for refunds, password resets, billing disputes, etc. Tune over time as the AI learns your edges.

  • New

    "Why I'm unsure" annotations. When the AI scores a reply below threshold, it explains why in plain language: missing context, conflicting policies, low source coverage, etc.

  • Improved

    Calibration over time. Confidence scores now self-correct against actual outcomes. If the AI was 90% sure on something it got wrong, it learns that pattern is a 70%, not a 90%.

  • Removed

    Single global confidence threshold. One number for all reply types was always too blunt. The setting no longer exists — per-topic is the only mode.

0.7.0 February 4, 2026 journeys

The customer is the unit of work, not the message.

We removed the ticket queue. The home view now shows journeys — one per customer, ordered by what needs human attention. This is the change we'd been building toward since the start.

  • New

    Journey view. Every customer gets a single thread that includes all conversations, product events, jobs the AI is working on, and prior outcomes — across email and in-app.

  • New

    Jobs panel. Each journey breaks down into discrete jobs the AI is tracking ("waiting on customer reply," "verifying refund eligibility," "drafting follow-up"). Clear status, no statuses-as-tags.

  • Removed

    The ticket queue. No more list of one-line subjects. The team home view now shows journeys requiring judgment, in the order they need it.

  • Removed

    Status field. Open / pending / on hold / solved is a queue artifact. A journey is either in progress or it isn't, and the AI tracks that without anyone setting a flag.

0.6.0 December 17, 2025 tools api

The AI gets hands.

An AI that can only draft replies is half a system. This release was about giving the AI the ability to actually do the things it was answering questions about — refund a charge, resend a verification email, update a subscription.

  • New

    Tools API. Connect any internal endpoint as a tool with a JSON schema and a permission scope. The AI can call it, with audit logs and per-tool confidence thresholds.

  • New

    Approval steps. For sensitive tools (refunds over a threshold, account merges), the AI assembles the action and queues it for one-click human approval instead of executing.

  • Improved

    Knowledge ingestion. Markdown, Notion, Google Docs, and public help-center crawls now ingest cleanly with semantic chunking. Re-ingest is incremental.

0.5.0 October 3, 2025 dual write

Run alongside your existing stack, then leave it.

We heard the same thing from every prospect: I want to try this without ripping out my current support tool. Dual-write mode lets support-arc shadow your existing inbox and write back into it, until you're ready to switch.

  • New

    Dual-write to Zendesk and Intercom. support-arc receives messages, drafts replies, and writes them back into your existing tool as agent drafts. No customer-facing change while you evaluate.

  • New

    Shadow mode. The AI handles every conversation without sending. You compare its drafts against what your team actually shipped, side by side, for as long as you want.

  • Improved

    Switchover. When a team is ready to go live, primary routing flips with a single setting. Historical context is retained.

0.1.0 August 14, 2025 origin

We started.

First commit. Three founders, a thesis, and an inbox we were both running and replying to ourselves to feel the work.

  • New

    An idea. If the AI is the primary actor in support, the entire shape of the software should be redrawn around the customer's journey instead of the team's queue. Everything since has been an attempt to be specific about what that means.