LoginSign Up
← Back to all sparks
S

Subbly

E-COMM
Velocity2.5

Subbly's AI builder becomes interruptible, months after retention intelligence moved in-house

ai-builderagent-steeringchurn-predictionsubscriptionscredits
◆Current state
Development is uneven across the product. The AI Builder is where the recent motion is: model intelligence was upgraded in June, and August makes the agent steerable mid-run — a new prompt can be queued while it works and gets inserted between steps rather than waiting for the current one to finish. The subscription platform itself has been quieter since spring, when Subbly replaced a third-party churn model with an in-house one it measures as 68% more accurate, added a cancellation type setting for immediate versus end-of-period, and shipped biometric two-factor with anomaly detection.
◆Where it's heading
The builder work is converging on keeping a person in the loop while the agent runs rather than making the agent more autonomous — the stated motivation is course-correcting a run that is going wrong and avoiding wasted credits. That framing, credits as the cost of a bad run, is worth noting: it treats agent time as a metered resource the merchant is spending, which shapes what gets built next. The retention side has not seen comparable investment since the churn model landed in May.
◆Prediction
Expect further mid-run controls on the builder along the same line — editing or cancelling a queued step, or rolling back a bad one — since the credit-waste problem this release names is only partly solved by inserting prompts.

◆Recent moves

  1. 27d ago

    Queue new prompts while the AI builder is still working

    A new prompt can now be sent while the agent is mid-run, with the builder inserting it between steps rather than making the user wait for the current one to finish. It makes the existing builder interruptible rather than adding capability, and Subbly frames it explicitly around saving credits on runs that go wrong.

  2. 3mo ago

    Upgraded Normal intelligence in the AI Builder

    An intelligence upgrade to the AI Builder's Normal mode. A model refresh rather than a change in what the builder does, and the start of the run of builder work that the mid-run steering release continues two months later.

  3. 4mo ago

    Improved Churn Predictions: New model is 68% more accurate

    ⚡ SPARK

    Still the most consequential release in this window and the last major move on the retention side of the product. Everything shipped since has been builder work, which makes the churn model look like a finished project rather than the opening of a line.

  4. 4mo ago

    Cancellation Type Setting: Immediate or End of Period

    Merchants can choose whether a cancellation takes effect immediately or at period end. A small billing control, but it sits directly next to the churn work — the difference between the two settings is how much time a win-back offer has to land.

    View source ↗
  5. 5mo ago

    Account security: biometric 2FA, anomaly detection, sessions

    Biometric two-factor, anomaly detection, and session management for merchant accounts. Platform hardening rather than merchant-facing capability, and the oldest substantive release in this window.

  6. 5mo ago

    🔐 Improved account security: biometrics + anomaly detection + more

    A second capture of the same account-security release, carrying an identical timestamp and content. A feed duplicate rather than a separate shipment.

    View source ↗