Accounting Insights reads your ledger the way a controller would if they had time to do it every morning: where cash actually moved, which entries are unusually large for your workspace, and what changed since yesterday. Every figure is a query against your own data with a threshold you set. The language model is handed the answer and writes the paragraph — it is not asked what it thinks.
Written by the model, from the four figures above
“Operating cash is positive and carrying the period; the negative investing line is a single equipment purchase rather than a trend.”
If no provider is configured, the amber panel is simply absent. The four figures above it do not change.
A list of nouns is easy to write and impossible to evaluate. These are the four computations behind the Accounting card, described closely enough that you can decide whether they would tell you anything you do not already know.
Movement is classified as operating, investing or financing from the accounts it touched, and a net figure is taken across the three. The split is the point: a healthy net number made entirely of a loan drawdown and a healthy net number made entirely of trading are the same number and completely different situations, and a single total hides which one you are in.
“Large” is not a fixed sum, because a fixed sum is either noise for one workspace or silence for another. An entry is flagged when it exceeds a multiple of your own typical entry — shipping at 1.5× and adjustable. A workspace whose entries are usually six figures is not told that a six-figure entry is unusual.
Where the pattern of posting departs from the workspace's own recent behaviour, the departure is surfaced as an alert carrying a label and a link straight to the record. The link is the part that matters. An anomaly you have to go and find is one you look at on the third occasion it appears, by which time it is not news.
The last movements on each side, in one place, so the question “what has actually happened since I last looked” has an answer that does not require running a report. Unglamorous, and the thing most people open the card for.
The card carries a headline score and, next to it, a confidence. They are separate on purpose. The score says how the module looks; the confidence says how much data that judgement was formed on. A workspace two days into its first month gets a low confidence rather than a confident verdict on four transactions — which is the failure that teaches a finance team to ignore a dashboard permanently.
This is the single most important thing to understand about Accounting Insights, and it is the thing most “AI analytics” pages are careful not to say. Every figure on the page — every count, every total, every flagged record — is computed by an ordinary database query against your own data, with thresholds you control. A language model is used for exactly one job: turning that computed result into a paragraph a person can read quickly.
Counts, totals, ages and comparisons, run against your workspace under the same tenant scoping as every other screen. This step is deterministic: run it twice on unchanged data and you get the same answer twice. If the engine says nine invoices are overdue, nine rows matched.
What makes a request stale, a cost a spike, an entry large or a balance overdue is a number in your settings, not a judgement in a model. The defaults are listed below. Change one and the flags change with it, immediately and predictably.
Only then is a short plain-language summary generated from those facts. It is written on a background job so the page never waits for it, and it fails soft: if no provider is configured or the request fails, the summary is simply not there and every number on the page is unaffected.
The summary is stored with a fingerprint of the exact figures it was generated from. When those figures move, the fingerprint no longer matches and the summary is regenerated rather than served stale. This closes the failure that makes cached AI commentary dangerous: a confident paragraph describing last week, sitting directly above this week's numbers, with nothing on the screen indicating the two disagree.
Results are cached per workspace for a window you set — fifteen minutes by default — so the page is fast without being out of date, and a refresh action clears it when you want the current picture immediately.
These are the actual settings behind the flags on this page. They are per workspace, they are yours to change, and the defaults are chosen to be defensible rather than dramatic — an engine tuned to flag everything on day one is an engine nobody reads by day three.
| Setting | Ships as | What it decides |
|---|---|---|
| accounting_large_entry_multiplier | 1.5 | How far above your own typical entry a posting has to be before it is flagged as large. Raise it in a workspace with naturally variable entries; lower it if you want a tighter watch. |
| cache_minutes | 15 | How long a computed result is reused before it is worked out again. The refresh action clears it immediately when you want the current picture. |
| modules_enabled | accounting: true | Whether the accounting module contributes to insights at all. A workspace that does not keep its books here can switch it off rather than read an empty card. |
Every flag carries the way to act on it. A finding is not a number on a dashboard you then go hunting for — it is rendered with a label and a link to the screen where the work is done. An insight you cannot act on from where you are reading it is a report, and you already have reports.
The test of an insights card is not whether it is clever. It is whether a controller who opens it before the first meeting closes it having done something. Here is that, concretely.
Operating is carrying the period and investing is deeply negative. The split makes it a one-line question rather than a morning of digging: was the equipment purchase approved, and is it one thing or the start of a pattern?
It is 2.3× the workspace's typical posting, comfortably past the 1.5 threshold. The alert carries a link, so the controller is looking at the journal entry itself rather than searching for it by amount.
A quarterly insurance premium. Nothing to do — but it was checked in ninety seconds by the person who should check it, on the day it posted rather than at month end when the context has gone.
Two large customer receipts landed yesterday. That is the number the ten o'clock actually wanted, and it did not require running a report to get it.
Notice what did not happen: nothing was decided by a model. The engine found a large entry because a number exceeded a multiple, and a person decided it was an insurance premium. That division of labour is deliberate, and it is why the summary paragraph can go missing without the card losing its value.
What AWRA OpsHub does today
More we can add to your workspace
Where we point you to a specialist
The middle column is work on top of an engine that already computes, caches, thresholds and links. If cash forecasting or duplicate-payment detection is the row that would change your month, tell us.
Revenue momentum, overdue invoices, average days to pay and who is slow.
Stockout risk on a horizon you set, dead stock, overstock and margin problems.
Late orders, stale requests, cost spikes and how concentrated your supply is.
All four modules on one card, and the workspace-level narrative across them.
The arithmetic rules underneath these figures — and why nothing is converted.
The detailed view behind the overdue signal, by age bucket and owner.
Help Center
Run inventory, procurement, assets, sales, and field work with approved AWRA guidance for setup, migration, integrations, security, pricing, and support.
