The night shift on your numbers,without a night shift.
Agents read your locations on a schedule you can check, score each one against its own normal, and bring what needs you. Most of what they do is deterministic analysis — which is exactly why the numbers hold up.
Agent registry · demo network
last full scan · today 04:00
What an agent actually does
Read, score, propose, hand over.
The same four moves, whether the agent is watching margin, demand or stock. The last one is the point: a human decides.
They read
Cloud connectors sync every few minutes, and an on-premise agent polls legacy ERP systems that have no modern API — so even a network running on 1C reads close to live.
They score
Every location is measured against its own seasonal normal for that day of the week, robustly enough that one strange day does not move the baseline, and with a money floor so a statistically odd but trivial move is never raised as a problem.
They propose
A finding becomes a decision with a target metric, a predicted effect and a measurement window — the shape the proof loop needs.
They hand over
It arrives in the Decision Inbox with the money at stake attached, and with how it was computed. The agent stops there; a person decides.
How often, exactly
Four schedules, published rather than promised.
every 5 minutes
connectors land data from POS, accounting, delivery
every hour
events become scored signals
every 24 hours
full scans, the money detectors, the decision cycle
weekly
pattern detection across the network
Published intervals, not a claim of real time. Told “instant”, an operator who then sees a four-hour-old number stops trusting the rest of the screen.
What runs on your data
Four that watch, and two you control.
None of them is a chatbot with a schedule. Each is a specific piece of analysis that would otherwise be a person with a spreadsheet.
Anomaly agent
Scores revenue and margin per location against that location's own seasonal pattern, hourly.
A quiet slide at a steady shop is caught, and a volatile one stops crying wolf.
Forecast agent
Forecasts demand per product, including the irregular and lumpy demand that breaks simple averages, and sets a reorder point at your service level.
Fewer lost sales, and less capital frozen on a shelf.
Grouping agent
Groups comparable locations by revenue scale, volatility, average check, margin and how many days they actually trade.
Fair comparisons — and the control group behind every proven result.
Replenishment agent
Builds a buffer per location and product, flags shortages and excess, and marks how much it trusts the data behind each line.
A short list of what to order today instead of a thousand rows to read.
Detectors you switch on
A catalogue of money detectors — below cost, expense creep, cash banked short, purchase cost drift, mix erosion, stock-outs, ABC movement, overstock — each a toggle for your organisation rather than a switch we flip for everyone.
You start with the core set and add the ones your business actually argues about.
Agents you describe yourself
Say what to catch, in a sentence. It is compiled into a deterministic rule and replayed over 90 days of your own history before you keep it — the section below shows the whole path.
You see how often it would have fired before you let it interrupt anyone.
an agent you describe yourself
pilot · per organisationwhat you say
“Tell me when meat write-offs pass 3% of category revenue”
what it compiles to
- metric
- waste_share
- scope
- category = meat
- test
- > 3% of category revenue
- window
- daily, per location
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
The model compiles the phrase and explains what was found. It never does the detecting, and it never invents a quantity.
Where the line is
What we mean, and don't mean, by agents.
The word is 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. Most of it is deterministic analysis and rules. The language model compiles what you asked for into a rule and explains what was found; it never does the detecting, and it never invents a quantity or an amount.
- Agents propose and, if you allow it, accept. They do not write prices or orders into your POS.
- Hands-off acceptance is opt-in and capped: a fixed list of decision types, a money ceiling, a daily limit and a kill switch.
- Agents you write yourself are a pilot, switched on per organisation and off by default. Simple threshold rules — 'tell me if this metric crosses that number' — are available to everyone.
- The replenishment agent is a pilot, and its orders leave as a CSV for your buyer rather than going to a supplier.
Agents FAQ
What people ask once they stop hearing the word 'AI'.
Honestly: they are scheduled, deterministic analytical pipelines plus a rule engine, with two places where a language model does the work — turning a phrase into a rule, and the copilot conversation. We say so rather than dress it up, because reproducibility is the point. Ask why a number is what it is and you get the same answer twice.
Yes — the figure above is the whole path: your sentence, the rule it compiles to, and the replay over 90 days of your own history that tells you how often it would have fired. Then it runs each night with the rest. It is a pilot we switch on per organisation, with a cap on how many agents one network can keep. The simpler kind — a threshold on a metric — is available to everyone without the pilot.
No. The arithmetic is not the model's job. Quantities, money, reorder points and outcomes are computed in code from your connected data, and each signal opens into a page showing how it was computed and how fresh the data behind it is.
Every few minutes, hourly, daily and weekly, depending on the job — the four schedules are drawn above. We publish the intervals instead of calling it real time.
Every live view shows when the data last landed and how many locations reported, so a gap is visible rather than silently averaged away. Where data quality is weak, the affected recommendations drop to suggestion only and say why. A window with no data at all is skipped rather than treated as a zero — no data is not the same as nothing sold.
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 measuring against the location's own volatility, and the record says so.
Talk to us
Tell us about your business. We'll call you.
Tell us a bit about your business and your data sources. We'll tailor the walkthrough, then you pick a time that works for you.