Every store, read every day. Every signal has an owner.
Detectors score each store against its own normal after every sync and in a daily scan. Each signal goes to the AI employee who owns it, and reaches you only as a decision, with the evidence inside.
Your watchers
deterministic rules and their firing history
Write-offs above 3% of sales
write_offs / revenue > 3% · daily · Ana
Store margin under 25%
store_revenue / store_cost < 1.33 · daily · Ana
Marql agents — the built-in sensor layer, tuned by us
Overstock — frozen cash
days of cover ≥ 1.5× target DIO, latest snapshot
Projected stockout
days of cover ≤ lead time + safety
Below-cost sales
revenue < cost per SKU, 14d window
Demand shift
unit velocity sustained 2+ weeks, de-seasonalised, promo/calendar excluded
Promo shortlist
in-stock SKUs scored on margin headroom × elasticity × cover, weekly
Event ahead
holiday inside its planning lead time, with last year's category moves
POS↔books revenue check
POS vs revenue accounts gap > 2%, last closed month
Purchase-cost creep
unit cost ≥ +8% vs 3-month median
Duplicate supplier invoices
same supplier+amount, ≥2 invoices, 45d
23 built-in detectors · each switched on for your organisation
A signal is evidence, not a notification.
It goes to whoever is responsible for that kind of problem, and reaches a person only once it is worth a decision.
Read
Connectors bring data in every five minutes; an older ERP such as 1C is read by an on-site agent or overnight. Forecasts and store groups are rebuilt each night in your own time zone.
Score
Each store is scored against its own normal for that weekday, robust to one odd day, with a money floor so a trivial move is never raised. Every signal carries a confidence: High, Medium or Low.
Route
Every signal type has one owner among your AI employees. Some hand over: a below-cost sale during a promotion goes to the marketer rather than to operations.
Decide
A person sees signals only inside a decision on Today: grouped, filtered (at least €50 a week, under 14 days old, no duplicates) and ranked by benefit. When the condition clears, the signal closes itself.
65 signal types · one owner eachA role you have not hired hands its signals to the owner or COO.
AI supply employee9
Projected stockout · overstock · purchase-cost creep · late delivery
AI operations manager17
Revenue pace · below-cost sale · shrinkage · cash banking gap · data health
AI finance manager19
Books against the till · expense creep · margin mix · sliding SKU margin
AI marketer3
Promo shortlist · event ahead · promo measured
AI category manager3
Demand shift · assortment class drop · category shift
Owner or COO14
Network patterns · store behind its peers · your own rules
Inside every decision, the evidence.
Review prices and suppliers for low-margin dairy SKUs
This decision groups 3 signals. This card is about Westside only.
Sliding SKU marginHigh
6 dairy SKUs · −4.2 pp · 28d vs prior 28d
Purchase-cost creepHigh
one dairy supplier · unit cost +9% vs 3-month median
Margin dropMedium
dairy · 3 weeks below the store's own normal
How it’s computed
- Method
- Margin against the store's own 8-week normal, robust to one odd day
- Window
- 21 days
- Sources
- POS · ERP cost prices
- Missing
- Nothing
Confidence: HighMeasurable
Written by marql from POS and ERP, 6 min ago
Built-in detectors, and rules of your own.
Each is a specific piece of analysis that would otherwise be a person with a spreadsheet. None of them is a chatbot with a schedule.
Each store against its own normal
Revenue and margin are scored against that store's own seasonal pattern after every sync. A flat rule either misses a quiet slide at a steady store or cries wolf at a volatile one; this does neither.
A real drop is caught, and an ordinary Tuesday is left alone.
MonTueWedThuFriSatSun
Tuesday, an ordinary day
Tuesdays run about 5% below the week's average here, and this one sits in the middle of its own normal. The tightened flat rule fires on it anyway — and on Wednesday and Thursday.
Saturday, 16% below its own normal
Saturdays run 16% above the average here, so a drop of the same size lands at an ordinary Tuesday's height. A flat −20% rule never reaches it, and tightening one until it does costs three false alarms a week.
A revenue signal fires at 3σ from this location's own normal and at least 15% below it. Saturday is 16% and about 12σ. Illustrative figures.
The books, checked against the till
Nine checks on your accounting, switched on per organisation: duplicate supplier invoices, gaps in invoice numbering, VAT rates outside the allowed set, overdue receivables, payments to dormant counterparties, and three reconciliations of the books against POS revenue, cash and stock.
The leak a sales report cannot show: an invoice paid twice, or books that disagree with the till.
Reconciliations · last closed month
POS ↔ books revenuegap 0.8% · limit 2%
Till cash ↔ booksgap 9.6% · limit 8%
Inventory ↔ booksgap 4.1% · limit 12%
Signal → AI finance manager: till cash 9.6% below what the books received.
Duplicate supplier invoicessame supplier + amount, 45dOn
Invoice numbering gaps≥3 missing numbers in a series, 92dOn
VAT rate anomaliesrate outside the allowed setOn
Overdue receivables> 30d and ≥ €500 per entityOff
Payments to dormant counterparties≥ €250 after 180d silenceOff
Negative cash registerrunning balance < 0, 45dOff
Rules you describe yourself
Write what to watch in a sentence. It is compiled into a deterministic rule and replayed over the last 90 days of your own data, so you see how often it would have fired before it interrupts anyone.
You set the threshold on your own history, not on a guess.
an agent you describe yourself
pilot · per organisationwhat you say
“Tell me when write-offs pass 3% of a location's revenue”
what it compiles to
- metric
- spoilage
- scope
- per location
- test
- > 3% of revenue
- window
- daily
built only from the metric catalogue · never free-form code
replayed over your last 90 days
would have fired 7 times
about once a fortnight, at two locations
Keep itDiscard
The model compiles the phrase and explains what was found. It never does the detecting, and it never invents a quantity.
What we mean, and don't mean, by detection.
The words are used loosely everywhere. Here is the version you can hold us to, and the reason we prefer it that way.
This is not a swarm of autonomous AIs.
Detection is deterministic code and rules. The language model compiles your sentence into a rule and explains what was found; it never detects, and never invents an amount.
The 23 built-in detectors are switched on per organisation.
Every network has the core checks on revenue, margin, stock-outs and data health from the first day.
Detectors propose and a person decides.
Nothing is written into your POS. What each AI employee may do on its own is set by the owner or COO, per type of action.
Rules you write yourself are a pilot, switched on per organisation, up to 20 per network.
Simple threshold alerts on a metric are available to everyone.
Where data is missing, the decision waits: a gap shows as one data-health line on Today instead of a number built on nothing.
What people ask once they stop hearing the word 'AI'.
No. Your AI employees are the roles that act on signals. The detectors underneath are scheduled, deterministic analysis plus a rule engine. A language model does two jobs only: turning your sentence into a rule, and explaining a finding in plain words. Ask why a number is what it is and you get the same answer twice.
Data arrives every five minutes, and the realtime checks run after each sync. The full scan runs once a day, forecasts and store groups are rebuilt overnight in your time zone, patterns weekly, and your AI employees read the results from 06:00 local time. We publish the intervals instead of calling it real time.
The core checks on revenue, margin, stock-outs and data health run for everyone. The 23 built-in detectors — overstock, projected stock-outs, below-cost sales, demand shifts, promo shortlists, the books checks and others — are switched on for your organisation, and each shows when it last fired and how often in 30 days.
Yes. The figure above is the whole path: your sentence, the rule it compiles to, and the replay over your last 90 days that shows how often it would have fired. It is a pilot we switch on per organisation, up to 20 rules per network. A threshold alert on a metric is available to everyone without the pilot.
Every view shows when data last landed and how many stores reported. Decisions that depend on missing data are held back and counted in one line, «Data gaps», rather than shown on a guess. A window with no data is skipped, not treated as zero.
The watching works the same: each location is scored against its own history, which is what most of the detection uses anyway. What weakens is comparison. With too few comparable locations there is no control group, so proof falls back to the location's own volatility, and the record says so.
Ready to see
the euros it found?
The walkthrough is built around the POS and accounting you already run. Leave your details and pick a time.