AssembloERP
We sell an AI-ERP, which makes this an awkward article to write honestly. We are going to do it anyway, because the hype is doing real damage: owners buy expecting an oracle and get a moving average with a nicer chart.
The three things that genuinely work
1. Demand forecasting — for items with a pattern
Give a model two years of clean sales history for an item you sell every month, and it will beat your spreadsheet's three-month average. It picks up seasonality, weekday effects and the slow drift you stopped noticing.
Where it fails: new items, project-based orders, and anything driven by a single large customer's decision. If your demand is five orders a year, no model on earth can help. A human who knows the customer will do better, every time.
2. Anomaly detection — the underrated one
This is where the money actually is, and it gets the least attention. A model watching your transactions can flag:
- A purchase rate 14% above the last three orders for the same item
- Consumption on a job that is out of line with its BOM
- A customer whose payment days have quietly slipped from 30 to 52
- Scrap on one machine drifting up over three weeks
None of this is clever. It is arithmetic run consistently on everything, which is precisely what humans stop doing after the second busy week. It is also how the leaks in job work and stock variance surface before they become quarterly surprises.
3. Asking questions in plain language
"Which customers bought less this quarter than last?" is a real question that traditionally required a report request and a two-day wait. Language models answer it against your own data now, and that genuinely changes who can get answers — the owner, at 9pm, without asking anyone.
The caveat is unglamorous: it answers from your data. If two item codes mean the same part, it will confidently give you half the picture.
Where it gets it wrong
| Situation | What happens | What to do instead |
|---|---|---|
| Under 12 months of history | Forecast is noise wearing a confidence interval | Use a simple average; revisit in a year |
| Lumpy project demand | Model smooths away the reality | Human judgement plus a pipeline view |
| Duplicate item codes | Splits one item's history in two | Clean the master first |
| A one-off spike (a tender, a shutdown) | Learned as if it repeats | Mark outliers so they are excluded |
| Prices changing sharply | Value-based forecasts drift | Forecast quantity, apply price separately |
The prerequisite nobody sells
Every useful thing above depends on the same foundation: one code per item, transactions entered the day they happen, and rejections recorded as data rather than remembered. AI does not clean your data. It amplifies whatever discipline you already have.
Which leads to the least commercial advice in this article: if your masters are a mess, spend the first month on that and switch the forecasting on later. It will work better, and you will trust it — which matters more than accuracy, because a forecast nobody believes changes no decisions.
What we actually claim
Our copilot forecasts demand where there is history to forecast from, flags anomalies across purchases, consumption and collections, and answers plain questions about your own numbers. On a business with clean data and two years of history, that is genuinely useful. On six months of messy data, it will politely tell you it does not have enough to be confident — which we consider a feature.
If you want to know which of the three applies to your situation, tell us what you sell and how long you have been recording it. See also: what the copilot does inside the product.
