Artificial intelligence

AI is only as good as the data beneath it.

A language model cannot match an invoice against a supplier catalogue if there is no item master where both come together. FoodNotify has that item master. On top of it run demand forecasting, order suggestions, document capture and an MCP server for AI assistants. All in production, none of it on a roadmap.

Book a demoSee the MCP server ↓
170+sites with the forecast measured daily
90.9 %forecast summed over three days
1,000+suppliers with data in the system
Forecast vs. actual35 days, one site
actually soldforecast
Real measurement from one site, anonymised: 35 of the 90 holdout days the model never saw. Accuracy there is 90.9 per cent per day.

In production

Four AI features your sites use today.

Each one works on the same item master. That is why they calculate with your pack sizes, your prices and your recipes, not with averages from the internet.

In production

Demand forecast in every Smart List

The model learns from your sales data, plus weekday, season, public holidays, weather and events. It forecasts the quantities sold per site and breaks them down to items through your recipes, right inside the list you order from.

Across more than 170 sites we measure every night how well the forecast matches the quantities actually sold: 87.7 % on a single day, volume-weighted, over 90 days the model never saw. An order covers several days, and summed over three days it is 90.9 %, over seven days 92.7 %.Ongoing measurement in live operations, not a lab result. As of 20 September 2026.
In production

An order suggestion you confirm

Forecast, stock, delivery window and minimum quantities turn into a pre-filled basket: the right item, the right quantity, the right supplier. You adjust or confirm. Nothing leaves without your approval.

That cuts overstock and emergency orders, and saves the time that goes into copying numbers and recounting today.
In production

Invoices and delivery notes by camera

BEAM reads your documents and proposes the goods receipt: items, quantities, prices, pack sizes. Differences against the order and the catalogue price are flagged, and uncertain fields go to a human for review.

Price changes surface as the document is read, not at month end.
In production

MCP server for Claude and ChatGPT

Connect FoodNotify to your AI assistant once and ask your own data in plain language. The assistant calls our tools and gets real values back, not an estimate.

Read-only, with your permissions. Today the focus is recipe data: ingredients, quantities, costing, allergens and nutrition.

MCP server

Ask your data instead of hunting for reports.

This is how it looks in the assistant. Pick a question; the answer comes from an example business.

FoodNotify connected via MCP · read-only access

Example data for illustration. The assistant only sees what your account has rights to, and writes nothing back. Switching between several accounts is currently in testing.

How the forecast is built

Traceable, not magic.

1 · Data

POS data per item and day, recipes, stock, supplier catalogues. Plus the calendar: weekday, school holidays, public holidays, weather and events at the site.

2 · Model

A gradient boosting model per chain learns from the history of all sites and spots patterns a weekly average misses: shifts caused by weather, a public holiday or an event.

3 · Decision

The forecast turns into a quantity per item, calculated for your delivery window and your pack size. The suggestion sits in the Smart List; the approval stays with you.

Limits

What our AI deliberately does not do.

In a business where allergens and margins matter, the more important question is not what an AI can do, but where it stops.

×
It does not order by itselfEvery suggestion waits for a human to approve it. There is no automatic mode quietly ordering goods.
×
It does not invent allergensDeclarable information comes from supplier and manufacturer data. Where a value is missing, it says so instead of letting a model fill the gap.
×
No black boxFor every quantity you see what it is based on: consumption, stock, delivery window. And you see continuously how good the forecast was.
×
No access beyond your permissionsThe MCP server reads what your account may see, and writes nothing back.
×
No training on your dataYour data is stored in Germany on Hetzner servers and is not used to train language models.

See AI in your own operation

20 minutes with your numbers: the forecast in the Smart List, the order suggestion, document capture and the MCP server inside your assistant.

Book a demoCalculate your savings