Features · Investigate problems
A likely cause in the alert, with the evidence behind it.
When a critical incident opens, Operanse gathers what was observed and an AI model reviews it: the likely cause, how confident it is, and what it could not check. Every statement cites its evidence, and inference is never presented as fact.
On InsightThe evidence is gathered for every incident on Monitor; the AI review is on Insight.
Payments failing: 7 of 9 first attempts since 09:31 [E1]
Calls to the gateway’s API answered normally [E3]
The gateway plugin update at 09:12 [E5] likely changed how payments are confirmed on this store.
Whether the gateway reports these payments as declined on its side.
Evidence cited (3)
- [E1]payments.failed, Tue 09:31–10:02, against the usual for this hour
- [E3]outbound calls to the gateway host: 0 failures in the window
- [E5]change: payment gateway plugin 8.1 → 8.2, administrator, wp-admin
An alert says what is wrong, not why
“Payments failing” arrives at ten in the morning. The next hour goes on opening the dashboard, the changes list and the gateway’s logs, trying to line them up, before anyone can say whether it is the gateway, the update or the host.
- The first hour of an incident spent collecting, not deciding.
- A developer who needs the evidence before they can start.
- An AI answer elsewhere that sounds sure and cites nothing.
What you can see
The evidence, gathered without AI
For every incident, on Monitor and Insight, a rule-based planner collects the relevant observations and puts them in order: the readings, the store’s usual for that hour, the changes just before, the services it called. No model is involved in this step.
- E1ReadingPayments failed: 7 of 9 first attempts since 09:31
- E2BaselineUsual for Tuesday 09:00–10:00: 0–1 failures
- E3ReadingCalls to the gateway API: answering normally
- E5ChangePayment gateway plugin 8.1 → 8.2 at 09:12
Inference, labelled and cited
On Insight, an AI model reads those labelled records and writes the review. Every statement must cite evidence that exists; a validator rejects a review that does not.
The model’s conclusions are stored apart from the observed facts and never replace them.
- observed
- read from the store and the service; cites a record
- inferred
- written by the model; cites the records it rests on
- confidence
- low, moderate or high, stated on the cause
- not verified
- what the store’s data cannot show
What it could not check
A review says what is outside its reach: the gateway’s own dashboard, a mailbox, a setting it is not allowed to read. That list is where a person starts, rather than taking a confident guess as settled.
- Whether the gateway reports the webhook deliveries as failed.
- Whether the update changed the webhook signing secret.
The review arrives in the alert
The first message is held briefly so the diagnosis is in it, not in a later one.
CRITICAL open since Tue 14:21 (store time)
What changed
9 orders carrying a Stripe transaction id have been pending for more than 15 minutes, totalling $1,180. Completed payments in the last hour: 2, against this store’s usual 11–17 for a Tuesday afternoon.
Comparison
This store’s own normal for Tuesday 14:00–15:00, learned from 6 weeks of history: usually 14 completed payments and no paid-but-pending orders. Both readings are outside this store’s normal range.
What this is worth
Priced from this store’s hourly takings: a shortfall of about $1,240 over 3 hours, carried as an estimate. Renewals due in the window are included; retries are not counted twice.
AI review (evidence-cited)
The orders were paid at Stripe and the store did not complete them. The site’s own calls to api.stripe.com answered normally through the hour [E3], so the outbound side is healthy; what stopped is the inbound completion. A recorded change at 13:58 updated WooCommerce Stripe Gateway 9.4.1 → 9.5.0 [E5], 23 minutes before the first stuck order [E1]. Likely cause: the gateway’s webhook handler is no longer completing payment on this store since the update. Confidence: moderate.
What Operanse did not verify
Whether Stripe’s side reports the webhook deliveries as failed, and whether the update changed the webhook signing secret. Both are visible in the Stripe dashboard, not from the store.
If the review has not finished within 15 minutes, the alert is sent without it and the review appears on the incident page when it completes.
A review is a starting point to check. It may suggest what to look at next; Operanse is read-only and changes nothing on the store.
Coverage and limits
What it covers
- An AI review with critical alerts. The likely cause, every statement citing its evidence, labelled as inference and kept apart from what was observed.Insight
- The evidence, gathered. For every incident, the relevant observations collected and put in order, with no AI involved.Monitor
Limits
- Reviews are for critical incidents. Warnings are not reviewed.
- 6 reviews per site per day are included. Beyond that, an incident gets the gathered evidence and no review, unless you allow extra reviews.
- Two reviews for one site are spaced at least 30 minutes apart, so a run of related alerts is not reviewed once each.
- A likely cause is an inference, with a confidence. It is not proof, and it is labelled as AI inference wherever it appears.
- Reviews pause while a payment for the plan is overdue; the evidence is still gathered.
How an incident is investigated · Why an AI model cannot change anything
Start using it
- Install the plugin from wordpress.org. It is read-only: it cannot change the site.
- Pair the site with the token from your dashboard. Setup, step by step
- Reviews start with the first critical incident on Insight or the trial. Nothing to configure.
Insight On Insight. The evidence is gathered for every incident on Monitor; the AI review is on Insight. Compare plans
Questions
Does every alert get a review?
No. Critical incidents do, up to 6 per site per day, at most one per site every 30 minutes. Warnings are not reviewed. Every incident on Monitor and Insight gets the gathered evidence.
Can reviews cost extra?
Only if you allow it. Extra reviews are $1.50 each, up to a number per site per day that you set. It starts at zero.
What does the model see?
Labelled evidence records from your store’s monitoring, already redacted. Store content such as error messages or setting names is passed as evidence, never as instructions.
How is this different from asking my own AI?
A review runs automatically on a critical incident. Asking your own AI is a conversation you start, on Monitor and Insight, with your own AI client doing the reasoning.