Operations Exceptions, Predictive Packs and AI Controls

Operations › Exceptions brings problems from eleven sources into one inbox: smart alerts, security incidents, workflow exceptions, connector dead letters, QuickBooks sync failures, mobile offline conflicts, procurement match exceptions, support tickets past their SLA, ledger integrity checks, POS cash variances and trash purge failures. Each module still decides what is wrong; the inbox gives every answer the same owner, SLA clock, timeline and closure. An exception resolves itself when its source stops reporting it, and if the problem comes back later it opens as a new row rather than reopening an old one.

The SLA clock starts when an exception is first seen, and its length follows severity: 4 hours for critical, 24 for high, 72 for medium, 168 for low and 336 for information. At breach, the owner is notified, or the module’s people if there is no owner. Every further 24 hours the escalation steps up a level, to a maximum of three, and adds the people who manage settings. A daily digest goes to each person with the high and critical exceptions for the modules they can already open, so a storekeeper is not sent ledger problems. Anyone can also subscribe for themselves from the inbox’s SLA & Digests button, choosing the modules, the lowest severity, email, in-app or push, and daily or weekly; a subscriber is then left out of the default role digest.

In practice: a QuickBooks sync failure and a POS cash variance appear on the same morning. The finance manager takes the sync failure, the branch manager takes the variance, and both clocks start. The variance is closed within the hour with its reason; the sync failure is not, and at 24 hours the finance manager’s name appears in the escalation with the settings managers added, which is what gets the expired connector password renewed.

Open an exception to own it, acknowledge it, comment, attach evidence files and triage it with a verdict and a root cause, such as data entry error, process not followed, configuration, integration, system defect, fraud or misuse, or expected activity. Each source has a playbook of steps to follow. If the computed impact score misjudges an exception, override it with a reason; the reason stays on the row and the timeline, and clearing the override returns the computed score. If a rule is too noisy, tune it with a stated reason, for example a longer SLA for that rule in your workspace. Closing needs a reason: fixed, false positive, accepted risk, duplicate, handed off to support or no action needed. Everything lands on the timeline, so a closed exception carries its own history.

Hand off to support creates a support ticket from the exception. Only an allow-listed summary goes with it: the module, severity, title, a summary, when it opened and a reference. Customer names, amounts and anything else in the evidence stay behind, because a support ticket is read by more people than the exception was. The handoff preview shows exactly what the ticket will carry before you send it, on the web and in the mobile app. A closed exception stays closed while its condition persists, so closing a false positive does not reopen it at the next sync.

The executive view (Operations › Executive) sits over the inbox. It shows a heatmap of module by severity, the size and age of the backlog, whether SLA breaches are rising, and what is exposed in money, stock and relationships. Money is shown per currency and never converted, so a business with shilling and dollar exposure sees both. In practice: a director checks the executive view on Monday, sees procurement high-severity items rising for three weeks, and asks the procurement lead about match exceptions rather than about the whole department.

AWRA AI › Predictive shows seven packs, each a scored list of things likely to go wrong: stockout risk, supplier delay, late customer payment, cash pressure, till shrinkage, approvals likely to breach their usual turnaround, and asset custody risk. None of them is a trained model. Each is a stated rule or statistic over your own workspace’s records, refreshed every morning, and no language model is involved, so a workspace that has switched third-party AI off, or used up its AI allowance, still gets every pack.

Every row carries its explanation, shown under Why this on the web and in the app: the drivers with their values, the tables and time window it read, how many records, how confident the score is and why, the method in plain words, and the recommended action. Forecasts and anomaly findings carry the same kind of explanation. If a row is wrong, the explanation shows which input to fix, rather than leaving you to argue with a number.

You can mark a row helpful or not, and AWRA measures the outcome 14 days later, for example whether the predicted stockout happened, so the feedback is checked against what followed rather than only recorded. In practice: a pharmacy sees a stockout risk on an antibiotic. Why this shows 31 days of sales averaging 18 a day, 140 on hand and a supplier lead time of nine days; the buyer agrees, raises a request, and two weeks later the outcome shows the shelf never ran empty.

From an insight you can raise a procurement request, a stock transfer request, a follow-up task, an in-app alert to a colleague, a report pinned to the dashboard, or a support note. Each one is written through the module’s own process with its usual starting state: a procurement request starts pending and a transfer moves no stock until it is approved and dispatched. An action raised from an insight is reviewed exactly as one raised by hand.

Actions that commit money or stock and are worth at least your workspace’s threshold wait for a second person who holds Approve AI actions; the person who raised it cannot approve it. The threshold defaults to 50,000 in the workspace currency and is set in Settings › AI & data. The permission belongs to the Pro plan and above, and is not added to any role other than Admin automatically: an Admin grants it to the roles that should decide.

Every AI answer, in the assistants and in summaries, passes through output filters before anyone reads it. Card numbers, API keys, private keys and passwords are replaced with a removed marker; an answer that names another organization on AWRA is withheld entirely; a price that is not in the sources the answer was based on is dropped with its sentence; a certification or guarantee the sources do not make is dropped; and an instruction to do something unsafe, such as switching off MFA or sharing a password, is dropped. Each catch is logged, and AI usage can be reviewed by module, role and model.