● B2B offer — Supply chain & forecasting
Demand forecasting: assumptions shown, not a black box
A forecast is only useful if you know what it rests on. Your agent builds your forecasts from your sales history, your seasonality and your commercial calendars, and shows for each one its assumptions, its range and the reference periods used. Hosted in France: your volumes and your sales data stay with you. Management picks the working forecast and decides on commitments.
Hosted in France
Sales data protected
GDPR & AI Act: governed deployment
Human oversight
Updated on
Deployed in a few weeks
Demand forecasting · hosted in France
What demand should we expect for this family next quarter?
Forecast produced with its low and high range, and the reference periods used.
The assumptions are shown: seasonality observed, planned promotions, trend over the last twelve months.
One promotion from last year has been set aside: this is stated and explained.
🔗 Sourced · sales history and commercial calendar
The assumptions are shown: seasonality observed, planned promotions, trend over the last twelve months.
One promotion from last year has been set aside: this is stated and explained.
🔗 Sourced · sales history and commercial calendar
What figure should we commit to?
I provide the forecast and its range, with what makes it move from one scenario to another.
Committing to a volume with a supplier or a customer is a matter of your strategy: that decision belongs to management.
✎ Support · range provided, human commitment
Committing to a volume with a supplier or a customer is a matter of your strategy: that decision belongs to management.
✎ Support · range provided, human commitment
Local inference · no data outside the EU
History hosted in France
Sovereign by designLocal inference or hosting in France
GDPR & AI Act: governed deploymentTraceability & human oversight
TurnkeyDesigned, installed and operated for you
The supply chain department decidesThe agent prepares, never rules
✦ In brief
A Blue Lemon Agent forecasting agent builds your forecasts from your sales history, your seasonality and your commercial calendars, and shows for each one its assumptions, its range and its reference periods. No volume commitment is made automatically. It runs on local inference or is hosted in France: your sales data stays with you, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
100%
hosted in France in the target architecture
0
transfer outside the EU in the target architecture
6
forecasting uses ready to deploy
0
decision taken without human approval
These figures describe our offer, not results measured at a client. How large the gain is on your number of families forecast and the depth of your history is confirmed by a pilot.
The context
What does an AI agent bring to your demand forecasts?
A forecast whose assumptions are known can be discussed and improved; that is what makes it useful in a committee.
! The issue
A usable forecast comes with its assumptions and its range. A single figure with no context cannot be discussed; a documented forecast lets you decide in full knowledge. The agent systematically shows the reference periods used, the seasonality observed and the commercial events taken into account or set aside.
✓ Our answer
The supply chain department has forecasts that can be explained, family by family, with what makes them move. Committing to a volume with a supplier or a customer is a matter of strategy and remains its decision. Local inference or an isolated resource hosted in France: your volumes and your sales history, revealing of your commercial position, do not leave the company.
The decisive point
Your volumes and your sales data: sovereignty & compliance
Your sales history and your volumes describe your business precisely. Here is how the architecture of our agents protects them.
Local inference
The agent can run on a machine belonging to your organisation: no sales history and no volume leaves the network.
Hosting in France
Otherwise, a dedicated and isolated resource hosted in France, under French law — your sales history and your commercial calendars: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
For your volumes and your sales data, the architecture aims to reduce exposure to the Cloud Act and FISA 702; being located in France or in the European Union does not, on its own, guarantee immunity.
Isolated resource
No pooling: an environment strictly dedicated to your company and its product families.
Assumptions and ranges always shown
Every forecast states its reference periods, its assumptions and its range; encryption, role-based access and logging of the forecasts produced.
AI Act: governed deployment
The agent is strictly in support; no volume commitment is made automatically; traceability and human oversight from end to end.
What depends on the architecture chosen These points are not general guarantees: they are settled deployment by deployment, in the quotation.
- The applicable location is that of the architecture set out in the quotation and verified before commissioning.
- Local execution is announced only for the configuration explicitly described and accepted in the quotation.
- The applicable isolation depends on the deployment mode set out in the quotation; no dedicated isolation is presumed.
- Roles and permissions are configured and accepted for the identities and systems actually connected.
- The events logged, their content, their retention period and who may access them are defined for the deployment chosen.
For forecasts tied to a confidential launch or a strategic transaction, SecNumCloud and reinforced hosting are options depending on your requirements. A single architecture is designed to answer both the GDPR and extraterritorial exposure. Designed for deployment in line with the GDPR and the AI Act, after the processing, roles and context-specific risks have been assessed.
Demonstration
See the agent at work
5 real situations, taken from those that come up most often. Pick one: the exchange unfolds as it would in your organisation.
A scripted demonstration. These exchanges show how the agent behaves — its sources, its refusals, what it leaves to your teams. Nothing is sent from this page, no model is queried here, and the matters named are fictional. That is precisely what we promise your data.
The behaviours shown here — monitoring, automation rules, routing and reminders — are configured with you during deployment, from your tools, your rules and your thresholds.
The architecture points named in these exchanges — location, local execution, isolation, encryption, role-based access, logging — are not a guarantee attached to the demonstration: they are those of the architecture set out in your quotation, and verified before commissioning.
The company in this demonstration
Fictional companyLavalène — maker of ecological household cleaning products
- Sector
- Manufacture of ecological cleaning products, sold through supermarkets and specialist stores
- Headcount
- 260 staff, including 5 in supply chain: 1 director, 2 demand planners, 2 supply planners
- Market served
- 4 supermarket chains and 340 specialist stores, in France and Belgium
- Volume forecast
- 9 families, 480 active SKUs, €62M of revenue, 5 years of shipment history
- Tools in place
- ERP, monthly forecasting spreadsheet, commercial promotion calendar, retailer agreements — the agent plugs into them, nothing is replaced
- Who decides
- The supply chain director settles the working forecast; the monthly committee arbitrates the scenarios; sales commit volumes to the retailers
- Room for improvement
- The monthly forecast takes 9 days of work and comes out with no range; forecast error is measured on none of the 9 families
Lavalène has been forecasting for ten years in a spreadsheet three people know how to read, and nobody has ever measured what that forecast is worth. The agent runs on local inference on a machine at the company and reads the ERP, warehouse shipments, the promotion calendar and the retailer agreements: it builds the forecast, displays its assumptions, and always gives its range and the depth of measurement behind it. The supply chain director settles the working forecast, and commitments to retailers stay with sales. The exchanges below cover six months, from the first measurement to the review.
This company, its figures and the exchanges that follow were invented for the demonstration. They illustrate a common situation; they describe no real client.
Demand forecasting · hosted in France
I read five years of your shipments last night — 2.8 million lines, 480 SKUs, 9 families — and I started by measuring what your current forecast is worth. Nobody had ever done it.
Forecast error is the gap between what was forecast and what was actually sold, expressed as a percentage of the actual.
The starting point, measured across your last 60 months:
· Average error at 4 weeks: 27 %.
· Average error at 8 weeks: 34 %.
· The spread between families is 1 to 3: 16 % on the steadiest family, 48 % on the most promotional one. A company-wide average would have hidden exactly what needed to be seen.
The forecast I bring you for the quarter, family by family, and never a bare figure:
· Floor cleaners: 41,200 units, range 37,800 – 44,600, that is ± 8 %. Depth: 60 months of complete history.
· Laundry: 68,400 units, range 64,300 – 72,500, that is ± 6 %. Depth: 60 months.
· Glass cleaners: 22,900 units, ± 5 %. Depth: 60 months.
· Kitchen cleaners: 31,500 units, ± 7 %. Depth: 60 months.
· Baby range, launched 22 months ago: 9,800 units, ± 14 %. Depth: 22 months — the range is wider because the depth is shorter, and I say so rather than hand you a figure too precise to be true.
The time this moves: building a forecast took 60 % of the exercise; it takes 8 %. Nine days of work become a day and a half.
What I propose: I show you what makes each range as wide as it is — and on the first one, 4.5 of the 8 points close with a single piece of information you already hold. forecast_9-families-ranges.pdfError measured at 27 % at 4 weeks, depth per family
⛓ Sourced · ERP (5 years of shipments, 2.8 million lines), forecasting spreadsheet, actuals across 60 months
Forecast error is the gap between what was forecast and what was actually sold, expressed as a percentage of the actual.
The starting point, measured across your last 60 months:
· Average error at 4 weeks: 27 %.
· Average error at 8 weeks: 34 %.
· The spread between families is 1 to 3: 16 % on the steadiest family, 48 % on the most promotional one. A company-wide average would have hidden exactly what needed to be seen.
The forecast I bring you for the quarter, family by family, and never a bare figure:
· Floor cleaners: 41,200 units, range 37,800 – 44,600, that is ± 8 %. Depth: 60 months of complete history.
· Laundry: 68,400 units, range 64,300 – 72,500, that is ± 6 %. Depth: 60 months.
· Glass cleaners: 22,900 units, ± 5 %. Depth: 60 months.
· Kitchen cleaners: 31,500 units, ± 7 %. Depth: 60 months.
· Baby range, launched 22 months ago: 9,800 units, ± 14 %. Depth: 22 months — the range is wider because the depth is shorter, and I say so rather than hand you a figure too precise to be true.
The time this moves: building a forecast took 60 % of the exercise; it takes 8 %. Nine days of work become a day and a half.
What I propose: I show you what makes each range as wide as it is — and on the first one, 4.5 of the 8 points close with a single piece of information you already hold. forecast_9-families-ranges.pdfError measured at 27 % at 4 weeks, depth per family
⛓ Sourced · ERP (5 years of shipments, 2.8 million lines), forecasting spreadsheet, actuals across 60 months
± 8 % is wide. I cannot commit anything on that.
That is exactly what makes it usable: I tell you where the eight points come from, and you close half of them in two minutes.
The ± 8 % on floor cleaners, broken down:
· 4.5 points come from the March promotion, entered in the commercial calendar but unconfirmed. Depending on whether it runs, volume moves by 3,700 units.
· 2 points come from the weather effect measured over five years: an early spring pulls the cleaning peak forward by three weeks. That effect is real — it explains 2 points, not ten.
· 1.5 points are residual noise: the share nothing explains and that will stay there whatever the method.
What that means, very concretely: confirm the March promotion and the range moves from ± 8 % to ± 3.5 % within the minute. The central figure barely moves — it is the uncertainty that falls, and uncertainty is what you pay for.
What it costs you, in euros: safety stock is sized on the width of the range. On this family alone, moving from ± 8 % to ± 3.5 % releases €190,000 of tied-up stock. A ten-minute meeting with sales is worth €190,000 of cash, and this is the first time that arithmetic has been laid out here.
What I propose: I made the same breakdown across the 9 families. Three pieces of information you already hold close 11 points of range between them — confirmation of the promotions, the listing calendar of the northern chain, and the switchover date of the baby range. Give them to me and I republish everything within the hour. ranges-broken-down_9-families.pdf11 points closable, €190,000 of stock on one family
⛓ Sourced · 5 years of shipments, promotion calendar, weather records matched to sales
The ± 8 % on floor cleaners, broken down:
· 4.5 points come from the March promotion, entered in the commercial calendar but unconfirmed. Depending on whether it runs, volume moves by 3,700 units.
· 2 points come from the weather effect measured over five years: an early spring pulls the cleaning peak forward by three weeks. That effect is real — it explains 2 points, not ten.
· 1.5 points are residual noise: the share nothing explains and that will stay there whatever the method.
What that means, very concretely: confirm the March promotion and the range moves from ± 8 % to ± 3.5 % within the minute. The central figure barely moves — it is the uncertainty that falls, and uncertainty is what you pay for.
What it costs you, in euros: safety stock is sized on the width of the range. On this family alone, moving from ± 8 % to ± 3.5 % releases €190,000 of tied-up stock. A ten-minute meeting with sales is worth €190,000 of cash, and this is the first time that arithmetic has been laid out here.
What I propose: I made the same breakdown across the 9 families. Three pieces of information you already hold close 11 points of range between them — confirmation of the promotions, the listing calendar of the northern chain, and the switchover date of the baby range. Give them to me and I republish everything within the hour. ranges-broken-down_9-families.pdf11 points closable, €190,000 of stock on one family
⛓ Sourced · 5 years of shipments, promotion calendar, weather records matched to sales
And the error — how will you measure it once we are working together?
By a written protocol, published in advance, and applied even in the months when it does not suit me.
How the measurement is made:
· The forecast is frozen and timestamped at 8 weeks, then at 4 weeks before the period. A forecast that gets retouched afterwards can no longer be measured.
· It is compared to the actual out of the ERP, family by family and SKU by SKU.
· The result is published every month, with the month's error, the twelve-month rolling error, and the number of forecast/actual pairs it rests on — because an error measured over three months is not worth an error measured over sixty.
What replaying across your 60 months already gives:
· Error at 4 weeks: 27 % with your current method, 14 % with mine. Measured on 540 forecast/actual pairs — 9 families across 60 months.
· Error at 8 weeks: 34 % against 19 %. Same base.
· Across the first 4 months of live operation: 15 %, on 36 pairs only. That is consistent with the replay, and I will give it to you again at 12 months when the base is three times firmer.
What that changes in committee: you stop arguing about whether the forecast is good, you know how far it is off and on which families. A forecast whose error is known turns into a stock decision; a forecast with no known error stays an opinion.
The next step I propose: I show you the assumptions behind each of these forecasts — and you will be able to change one in a word, with the costed effect on screen.
How the measurement is made:
· The forecast is frozen and timestamped at 8 weeks, then at 4 weeks before the period. A forecast that gets retouched afterwards can no longer be measured.
· It is compared to the actual out of the ERP, family by family and SKU by SKU.
· The result is published every month, with the month's error, the twelve-month rolling error, and the number of forecast/actual pairs it rests on — because an error measured over three months is not worth an error measured over sixty.
What replaying across your 60 months already gives:
· Error at 4 weeks: 27 % with your current method, 14 % with mine. Measured on 540 forecast/actual pairs — 9 families across 60 months.
· Error at 8 weeks: 34 % against 19 %. Same base.
· Across the first 4 months of live operation: 15 %, on 36 pairs only. That is consistent with the replay, and I will give it to you again at 12 months when the base is three times firmer.
What that changes in committee: you stop arguing about whether the forecast is good, you know how far it is off and on which families. A forecast whose error is known turns into a stock decision; a forecast with no known error stays an opinion.
The next step I propose: I show you the assumptions behind each of these forecasts — and you will be able to change one in a word, with the costed effect on screen.
In the spreadsheet, nobody knows where the figures come from. That is the committee's real problem.
Every forecast arrives with its assumptions sheet, and every assumption carries its weight in units. So you know not only where the figure comes from, but what would move it.
The floor-cleaner sheet, as it is published:
· Reference base: the last 60 months of shipments, 4 months of which are neutralised and flagged as such.
· Underlying trend: + 3.1 % a year, measured over 5 years. Weight: + 1,200 units over the quarter.
· Family seasonality: coefficient 1.24 for the quarter, measured over 5 years. Weight: + 8,000 units.
· New listing at the northern chain, effective since January, taken from the signed agreement. Weight: + 2,400 units.
· March promotion, unconfirmed. Weight: + 3,700 units if it runs — and it carries 4.5 of the 8 points of range.
· Weather effect: ± 900 units depending on how early spring comes.
What this changes in committee: the discussion no longer bears on the figure, it bears on the two assumptions that weigh 11,700 units between them. That is a twenty-minute meeting instead of a morning.
The time this moves: documenting assumptions took 35 % of the exercise; it takes 4 %, and it is produced for all 9 families instead of none.
What I propose: I show you how seasonality is measured — and two of your families no longer follow the one your spreadsheet has applied since 2019. assumptions-sheet_floor-cleaners.pdf6 assumptions, each weighed in units
⛓ Sourced · 60 months of shipments, signed retailer agreements, promotion calendar, weather records
The floor-cleaner sheet, as it is published:
· Reference base: the last 60 months of shipments, 4 months of which are neutralised and flagged as such.
· Underlying trend: + 3.1 % a year, measured over 5 years. Weight: + 1,200 units over the quarter.
· Family seasonality: coefficient 1.24 for the quarter, measured over 5 years. Weight: + 8,000 units.
· New listing at the northern chain, effective since January, taken from the signed agreement. Weight: + 2,400 units.
· March promotion, unconfirmed. Weight: + 3,700 units if it runs — and it carries 4.5 of the 8 points of range.
· Weather effect: ± 900 units depending on how early spring comes.
What this changes in committee: the discussion no longer bears on the figure, it bears on the two assumptions that weigh 11,700 units between them. That is a twenty-minute meeting instead of a morning.
The time this moves: documenting assumptions took 35 % of the exercise; it takes 4 %, and it is produced for all 9 families instead of none.
What I propose: I show you how seasonality is measured — and two of your families no longer follow the one your spreadsheet has applied since 2019. assumptions-sheet_floor-cleaners.pdf6 assumptions, each weighed in units
⛓ Sourced · 60 months of shipments, signed retailer agreements, promotion calendar, weather records
Seasonality — where exactly do you get it from?
From your sales, family by family, and not from a general curve — and two of your families changed seasonality without anyone noticing.
How it is measured: I compute one coefficient per month and per family, over 5 years, after neutralising exceptional events — a promotion left in the data distorts seasonality for five years.
What that gives, and the gap with your spreadsheet:
· Seven families out of nine follow a stable seasonality, and your spreadsheet's coefficients are within 4 points of mine. Your demand planners were right, and I say so.
· Two families have moved. On glass cleaners and laundry, the peak has shifted three weeks earlier since online sales passed a quarter of volume. Measured across the last 24 months against the previous 36: the peak moved from week 14 to week 11, steadily across both years.
· Effect on the forecast: 2,100 units placed in the wrong part of the quarter, producing a shortage in March and overstock in April, every year for two years.
What I propose, and the depth behind it: compute seasonality for those two families on a rolling 24 months rather than on 5 years, so it follows the sales channel. Across the 24 months available, that setting takes the error from 21 % to 13 % on glass cleaners. The depth is 24 months: enough to decide on, and I will republish the measurement when it reaches 36.
And for the other seven families I change nothing — altering a method that works always costs more than it returns. seasonality_by-family-and-channel.pdf7 families steady, 2 peaks shifted by 3 weeks
⛓ Sourced · 60 months of shipments per family, sales split by channel, current spreadsheet coefficients
How it is measured: I compute one coefficient per month and per family, over 5 years, after neutralising exceptional events — a promotion left in the data distorts seasonality for five years.
What that gives, and the gap with your spreadsheet:
· Seven families out of nine follow a stable seasonality, and your spreadsheet's coefficients are within 4 points of mine. Your demand planners were right, and I say so.
· Two families have moved. On glass cleaners and laundry, the peak has shifted three weeks earlier since online sales passed a quarter of volume. Measured across the last 24 months against the previous 36: the peak moved from week 14 to week 11, steadily across both years.
· Effect on the forecast: 2,100 units placed in the wrong part of the quarter, producing a shortage in March and overstock in April, every year for two years.
What I propose, and the depth behind it: compute seasonality for those two families on a rolling 24 months rather than on 5 years, so it follows the sales channel. Across the 24 months available, that setting takes the error from 21 % to 13 % on glass cleaners. The depth is 24 months: enough to decide on, and I will republish the measurement when it reaches 36.
And for the other seven families I change nothing — altering a method that works always costs more than it returns. seasonality_by-family-and-channel.pdf7 families steady, 2 peaks shifted by 3 weeks
⛓ Sourced · 60 months of shipments per family, sales split by channel, current spreadsheet coefficients
And if one of my demand planners disagrees with an assumption?
They change it in a word, the effect appears costed before they even confirm, and the correction is measured like everything else. That is what makes it valuable.
How an assumption is corrected: they open the sheet, replace the value — « the March promotion will run, 25 % discount » —, and the forecast and its range recompute within the second, with the delta shown in units and in euros of stock.
What is kept: who changed what, when, the value before and after, and the error observed at the end. A human correction is measured on exactly the same footing as a machine assumption.
What four months of operation already show, and I publish both sides:
· 11 assumptions were corrected by hand by your two demand planners.
· 7 improved the error, by 3.1 points on average. All of them bore on what the data does not contain: a delisting announced verbally, an assortment change, a shortage at a competitor.
· 4 made it worse by 1.8 points. They bore on seasonality, which five years of data describe better than recollection.
· Net effect: − 1.7 points of error thanks to the human corrections. Your demand planners bring something I do not have, and now we know what it is: the field, not the curve.
What I propose: that I flag in advance the assumptions where your knowledge beats my data — there are three a month on average — and let the others run on their own. Your two demand planners move from a morning of keying to twenty minutes of judgement on what only they know.
How an assumption is corrected: they open the sheet, replace the value — « the March promotion will run, 25 % discount » —, and the forecast and its range recompute within the second, with the delta shown in units and in euros of stock.
What is kept: who changed what, when, the value before and after, and the error observed at the end. A human correction is measured on exactly the same footing as a machine assumption.
What four months of operation already show, and I publish both sides:
· 11 assumptions were corrected by hand by your two demand planners.
· 7 improved the error, by 3.1 points on average. All of them bore on what the data does not contain: a delisting announced verbally, an assortment change, a shortage at a competitor.
· 4 made it worse by 1.8 points. They bore on seasonality, which five years of data describe better than recollection.
· Net effect: − 1.7 points of error thanks to the human corrections. Your demand planners bring something I do not have, and now we know what it is: the field, not the curve.
What I propose: that I flag in advance the assumptions where your knowledge beats my data — there are three a month on average — and let the others run on their own. Your two demand planners move from a morning of keying to twenty minutes of judgement on what only they know.
Last year we ran a −30 % promotion in March. It distorts all our averages.
It is identified, dated and neutralised — and I found thirty-three other events of the same kind over five years, none of which your spreadsheet accounted for.
The 34 exceptional events, sorted:
· 12 promotional campaigns, from −15 % to −35 %.
· 9 shortage periods longer than two weeks.
· 5 delistings of SKUs in store.
· 4 new listings.
· 4 logistics incidents that pushed deliveries back by more than ten days.
What the March campaign actually did, and this is the half that was missing:
· + 187 % of volume over the three weeks of the campaign. Everyone sees that part.
· − 41 % over the following four weeks. Customers had bought ahead, and that trough was neutralised nowhere. It was therefore read as a fall in demand, and it pulled your April forecast down for two years.
· Net effect of the campaign on the quarter: + 31 %, not + 187 %. That is the figure that should decide the next one.
What neutralising changes, measured: removing the 34 events from the base and recomputing takes the error of the 9 families down by 6 points on its own.
The time this moves: neutralising exceptional events took 30 % of the exercise; it takes 6 %, and it now covers 34 events instead of the 3 or 4 anyone remembered.
What I propose: that every new event be declared as it happens — and I now show you what I already know about the promotions ahead. exceptional-events_34-dated.pdf+187 % then −41 %, net effect +31 %
⛓ Sourced · 60 months of shipments, promotion calendar, shortage history, listing agreements
The 34 exceptional events, sorted:
· 12 promotional campaigns, from −15 % to −35 %.
· 9 shortage periods longer than two weeks.
· 5 delistings of SKUs in store.
· 4 new listings.
· 4 logistics incidents that pushed deliveries back by more than ten days.
What the March campaign actually did, and this is the half that was missing:
· + 187 % of volume over the three weeks of the campaign. Everyone sees that part.
· − 41 % over the following four weeks. Customers had bought ahead, and that trough was neutralised nowhere. It was therefore read as a fall in demand, and it pulled your April forecast down for two years.
· Net effect of the campaign on the quarter: + 31 %, not + 187 %. That is the figure that should decide the next one.
What neutralising changes, measured: removing the 34 events from the base and recomputing takes the error of the 9 families down by 6 points on its own.
The time this moves: neutralising exceptional events took 30 % of the exercise; it takes 6 %, and it now covers 34 events instead of the 3 or 4 anyone remembered.
What I propose: that every new event be declared as it happens — and I now show you what I already know about the promotions ahead. exceptional-events_34-dated.pdf+187 % then −41 %, net effect +31 %
⛓ Sourced · 60 months of shipments, promotion calendar, shortage history, listing agreements
As for promotions ahead, sales never tell us in time.
Three promotions in the coming quarter appear in no shared calendar — and I found them in your own signed retailer agreements.
What I recovered: the annual agreements with your four chains provide for contractual campaigns with their periods. Three of them fall in the quarter and had been entered nowhere: two at the northern chain, one in Belgium. They represent 14 % of the quarter's volume across two families. Without them the forecast was right on the trend and wrong on the total.
What I can now cost, campaign by campaign: I measured the elasticity of your 12 past campaigns — how much volume gains for each point of discount given — by family and by discount depth.
· Laundry, 20 % discount: + 96 % of volume, range + 78 % to + 114 %, measured on 4 comparable campaigns.
· Laundry, 30 % discount: + 187 %, range + 151 % to + 223 %, measured on 3 campaigns. The depth is shorter, the range wider, and I flag it rather than hand you a smooth figure.
· Floor cleaners, 25 % discount: + 74 %, range + 61 % to + 87 %, on 5 campaigns.
· And the trough afterwards, costed too: from −28 % to −41 % over the following four weeks, depending on depth. That is the figure missing from every promotional profitability study I have read here.
What I propose: that I attach the contractual campaigns from the retailer agreements to the forecasting calendar automatically, as soon as the agreement is signed. Across the last three years that would have recovered 9 forgotten campaigns and 4 shortages that followed from them. promotional-elasticity_12-campaigns.pdf3 campaigns recovered, post-campaign trough costed
⛓ Sourced · signed retailer agreements, promotion calendar, 12 past campaigns and their volumes
What I recovered: the annual agreements with your four chains provide for contractual campaigns with their periods. Three of them fall in the quarter and had been entered nowhere: two at the northern chain, one in Belgium. They represent 14 % of the quarter's volume across two families. Without them the forecast was right on the trend and wrong on the total.
What I can now cost, campaign by campaign: I measured the elasticity of your 12 past campaigns — how much volume gains for each point of discount given — by family and by discount depth.
· Laundry, 20 % discount: + 96 % of volume, range + 78 % to + 114 %, measured on 4 comparable campaigns.
· Laundry, 30 % discount: + 187 %, range + 151 % to + 223 %, measured on 3 campaigns. The depth is shorter, the range wider, and I flag it rather than hand you a smooth figure.
· Floor cleaners, 25 % discount: + 74 %, range + 61 % to + 87 %, on 5 campaigns.
· And the trough afterwards, costed too: from −28 % to −41 % over the following four weeks, depending on depth. That is the figure missing from every promotional profitability study I have read here.
What I propose: that I attach the contractual campaigns from the retailer agreements to the forecasting calendar automatically, as soon as the agreement is signed. Across the last three years that would have recovered 9 forgotten campaigns and 4 shortages that followed from them. promotional-elasticity_12-campaigns.pdf3 campaigns recovered, post-campaign trough costed
⛓ Sourced · signed retailer agreements, promotion calendar, 12 past campaigns and their volumes
A shortage distorts the measurement too: we never know what we would have sold.
You can: I rebuilt it. Across five years your real demand was understated by 4.2 % on average, and by up to 11 % on one family.
How I estimate what was not sold: during a shortage I compare the SKU that ran out with its twin SKUs that stayed available over the same period and in the same stores — same family, same size, same positioning. The gap between their curve and yours gives the unserved demand, and I give it with its range, because it is an estimate and not a reading.
What that gives across five years' 61 shortage periods:
· Estimated unserved demand: 4.2 % of total volume, range 3.4 % to 5.1 %, across 61 measured periods.
· On kitchen cleaners: 11 %, range 8.7 % to 13.4 % — that family ran out 9 times in five years, and its forecast dropped a little further each time. A shortage that repeats ends up prophesying itself.
· Effect on the recalibrated forecast: + 3,400 units over the quarter, across three families.
What it hands you, in euros: on kitchen cleaners alone, those 11 % represent €128,000 of revenue a year your history did not show — and that your forecasts therefore never asked production for.
What I propose: plug me into service level by SKU, so that unserved demand is rebuilt continuously rather than retrospectively. You will then see the demand you have, not only the demand you managed to serve. unserved-demand_61-shortages.pdf4.2 % on average, 11 % on one family
⛓ Sourced · shortage history (61 periods), shipments of twin SKUs, service level by SKU
How I estimate what was not sold: during a shortage I compare the SKU that ran out with its twin SKUs that stayed available over the same period and in the same stores — same family, same size, same positioning. The gap between their curve and yours gives the unserved demand, and I give it with its range, because it is an estimate and not a reading.
What that gives across five years' 61 shortage periods:
· Estimated unserved demand: 4.2 % of total volume, range 3.4 % to 5.1 %, across 61 measured periods.
· On kitchen cleaners: 11 %, range 8.7 % to 13.4 % — that family ran out 9 times in five years, and its forecast dropped a little further each time. A shortage that repeats ends up prophesying itself.
· Effect on the recalibrated forecast: + 3,400 units over the quarter, across three families.
What it hands you, in euros: on kitchen cleaners alone, those 11 % represent €128,000 of revenue a year your history did not show — and that your forecasts therefore never asked production for.
What I propose: plug me into service level by SKU, so that unserved demand is rebuilt continuously rather than retrospectively. You will then see the demand you have, not only the demand you managed to serve. unserved-demand_61-shortages.pdf4.2 % on average, 11 % on one family
⛓ Sourced · shortage history (61 periods), shipments of twin SKUs, service level by SKU
The monthly committee has to decide. Decide on what, exactly?
On three complete scenarios, and on a figure nobody here had computed: at your company, a missing unit costs 4.1 times a surplus unit.
The three scenarios, across the 9 families:
· Cautious — bottom of range: 268,000 units. Average stock 27 days. Expected shortage cost: €214,000 over the quarter. Overstock cost: €31,000.
· Central — the median value: 289,000 units. Stock 33 days. Shortage: €74,000. Overstock: €68,000.
· Aggressive — top of range: 311,000 units. Stock 41 days. Shortage: €18,000. Overstock: €142,000.
The calculation I made, and it moves the break-even point: I costed what a missing unit actually costs here — lost margin, the service penalty set out in your retailer agreements, and the cost of delisting when service level falls below the contractual threshold. Then what a surplus unit costs — tied-up cash, warehouse space, markdown on dated SKUs.
· A missing unit: €4.80. A surplus unit: €1.17. A ratio of 4.1.
· Direct consequence: your optimum is not the central value, it is 297,000 units, between central and aggressive. Expected total cost: €121,000, against €142,000 at the central scenario. €21,000 a quarter, simply by putting the cursor where your own economics put it.
And the ratio is not the same everywhere: on 6 families it exceeds 4; on the 3 short-dated families it falls to 1.3, and for those the cautious scenario is the right one. A single cursor for nine families was costing €34,000 a quarter.
What I propose: you set the cursor family by family in committee, on these figures, and I prepare the retailer commitment that follows. scenarios_cautious-central-aggressive.pdfOptimum at 297,000 units, €21,000 a quarter
⛓ Sourced · forecast ranges, retailer agreements (penalties and thresholds), ERP storage and markdown costs
The three scenarios, across the 9 families:
· Cautious — bottom of range: 268,000 units. Average stock 27 days. Expected shortage cost: €214,000 over the quarter. Overstock cost: €31,000.
· Central — the median value: 289,000 units. Stock 33 days. Shortage: €74,000. Overstock: €68,000.
· Aggressive — top of range: 311,000 units. Stock 41 days. Shortage: €18,000. Overstock: €142,000.
The calculation I made, and it moves the break-even point: I costed what a missing unit actually costs here — lost margin, the service penalty set out in your retailer agreements, and the cost of delisting when service level falls below the contractual threshold. Then what a surplus unit costs — tied-up cash, warehouse space, markdown on dated SKUs.
· A missing unit: €4.80. A surplus unit: €1.17. A ratio of 4.1.
· Direct consequence: your optimum is not the central value, it is 297,000 units, between central and aggressive. Expected total cost: €121,000, against €142,000 at the central scenario. €21,000 a quarter, simply by putting the cursor where your own economics put it.
And the ratio is not the same everywhere: on 6 families it exceeds 4; on the 3 short-dated families it falls to 1.3, and for those the cautious scenario is the right one. A single cursor for nine families was costing €34,000 a quarter.
What I propose: you set the cursor family by family in committee, on these figures, and I prepare the retailer commitment that follows. scenarios_cautious-central-aggressive.pdfOptimum at 297,000 units, €21,000 a quarter
⛓ Sourced · forecast ranges, retailer agreements (penalties and thresholds), ERP storage and markdown costs
And towards the retailers, what do we commit to?
The commitment is written, costed and ready to go: volume by family, revision range, dates. Sign the mandate and it is sent in six minutes.
What the prepared commitment holds, chain by chain:
· The volume committed per family, at the cursor the committee settled.
· The revision range your agreements allow — ± 10 % at D−30 with three chains, ± 15 % with the fourth —, and how I recommend using it: commit at the optimum and keep the revision as headroom, rather than commit low and spend it in advance.
· The revision dates, entered in the calendar with their D−7 alert.
· The costed effect of each commitment level on expected service level and on any penalty.
The mandate I propose, written, capped, dated:
· Purpose: send the volume commitment to the four chains and carry it into the ERP and the supply plan.
· Cap: the volume settled in committee, family by family — no commitment above it, ever.
· Scope: the 9 families and the 4 named chains, and nothing else.
· Term: until the review on the 30th, renewed in one word.
· Withdrawal: in one word, at any moment, including from a phone.
· Trail: who settled which volume, on which scenario, at which second.
· Execution: 6 minutes after signature, against a day and a half of rekeying today.
The commitment decision stays with sales — and they take it on a finished file, with the cost of each level in front of them, instead of judging on instinct at the end of a committee. commitment-mandate_retailer-volumes.pdfCapped per family, executed in 6 minutes
✎ Framework · retailer agreements, scenarios settled in committee, mandate settings, commitment log
What the prepared commitment holds, chain by chain:
· The volume committed per family, at the cursor the committee settled.
· The revision range your agreements allow — ± 10 % at D−30 with three chains, ± 15 % with the fourth —, and how I recommend using it: commit at the optimum and keep the revision as headroom, rather than commit low and spend it in advance.
· The revision dates, entered in the calendar with their D−7 alert.
· The costed effect of each commitment level on expected service level and on any penalty.
The mandate I propose, written, capped, dated:
· Purpose: send the volume commitment to the four chains and carry it into the ERP and the supply plan.
· Cap: the volume settled in committee, family by family — no commitment above it, ever.
· Scope: the 9 families and the 4 named chains, and nothing else.
· Term: until the review on the 30th, renewed in one word.
· Withdrawal: in one word, at any moment, including from a phone.
· Trail: who settled which volume, on which scenario, at which second.
· Execution: 6 minutes after signature, against a day and a half of rekeying today.
The commitment decision stays with sales — and they take it on a finished file, with the cost of each level in front of them, instead of judging on instinct at the end of a committee. commitment-mandate_retailer-volumes.pdfCapped per family, executed in 6 minutes
✎ Framework · retailer agreements, scenarios settled in committee, mandate settings, commitment log
And if actuals drift during the month?
You know on Monday morning, not at month end. I set the forecast against actuals every week, and I have already caught four situations in four months.
What I publish every Monday: for the 9 families, the week's actual, the gap to forecast, and the projection recomputed over the rest of the month with its new range.
The threshold I propose, and why that one: alert as soon as a family drifts by more than 12 % over two consecutive weeks. A single week means nothing — across five years, 61 % of one-week gaps close the following week, and alerting on them would have had you move supply for nothing 31 times a year. Two consecutive weeks, on the other hand, confirm 84 times out of 100.
What it gave in four months:
· 6 alerts raised.
· 4 led to a supply adjustment: two order increases, two deferrals. Measured effect: 2 shortages avoided and €41,000 of overstock not built.
· 2 closed by themselves, and I told you the following Monday that there was no longer anything to do — an alert that never closes becomes an alert people ignore.
The two adjustment gestures I prepare for every alert: the costed top-up order with the referenced supplier, lead time attached, and the production deferral with its effect on the other families. You choose in two minutes because both are already built.
What I propose: you set the threshold — I have costed 8 %, 12 % and 15 % across the last five years so you choose on figures — and the alert goes to the demand planner and the supply planner at the same time. weekly-tracking_threshold-and-gestures.pdf6 alerts, 2 shortages avoided, €41,000 of overstock avoided
⛓ Sourced · weekly ERP shipments, frozen forecasts, five-year history of gaps
What I publish every Monday: for the 9 families, the week's actual, the gap to forecast, and the projection recomputed over the rest of the month with its new range.
The threshold I propose, and why that one: alert as soon as a family drifts by more than 12 % over two consecutive weeks. A single week means nothing — across five years, 61 % of one-week gaps close the following week, and alerting on them would have had you move supply for nothing 31 times a year. Two consecutive weeks, on the other hand, confirm 84 times out of 100.
What it gave in four months:
· 6 alerts raised.
· 4 led to a supply adjustment: two order increases, two deferrals. Measured effect: 2 shortages avoided and €41,000 of overstock not built.
· 2 closed by themselves, and I told you the following Monday that there was no longer anything to do — an alert that never closes becomes an alert people ignore.
The two adjustment gestures I prepare for every alert: the costed top-up order with the referenced supplier, lead time attached, and the production deferral with its effect on the other families. You choose in two minutes because both are already built.
What I propose: you set the threshold — I have costed 8 %, 12 % and 15 % across the last five years so you choose on figures — and the alert goes to the demand planner and the supply planner at the same time. weekly-tracking_threshold-and-gestures.pdf6 alerts, 2 shortages avoided, €41,000 of overstock avoided
⛓ Sourced · weekly ERP shipments, frozen forecasts, five-year history of gaps
Six months. What is it worth, measured?
Forecast error moves from 27 % to 14 % at four weeks — and I give you the depth alongside the figure, because a forecast is judged on both.
What is measured, and on what base:
· Error at 4 weeks: 27 % → 14 %, measured on 54 forecast/actual pairs — 9 families across 6 months of live operation, and consistent with the 14 % obtained on the 540-pair replay.
· Error at 8 weeks: 34 % → 19 %, same base.
· Service level: 94.1 % → 97.3 %, across 6 months of shipments.
· Stock cover: 41 days → 33 days, that is €620,000 of cash released, recorded in the monthly balance.
· Duration of the forecasting exercise: 9 days → 1.5 days. Building 60 % → 8 %, documenting assumptions 35 % → 4 %, neutralising events 30 % → 6 %.
The figure that does not flatter me: of the 9 families, 7 improved and 2 did not. Their error stays at 24 % and 22 %, against 27 % and 25 % at the start — a 3-point gain where the others gained 15.
The causes are clear, and they are not the same:
· The baby range has only 22 months of history: two annual cycles are not enough for seasonality to emerge. That is mechanical, and I have announced it from day one through a ± 14 % range.
· The professional family depends on a single customer for 61 % of its volume: it follows no seasonality, it follows an order book.
What I did with it, and what it gives:
· For the professional family, I plugged into the forward order book that customer sends you contractually. The error moves from 22 % to 12 % — but the depth is only 8 weeks, that is 2 pairs, and I will tell you again at the quarter when the base is five times wider.
· For the baby range, the range stays wide and deliberately so, and it tightens by itself: it moved from ± 14 % to ± 11 % in six months as the history lengthens. At 36 months seasonality will be readable and I will apply it.
What I propose: open forecasting to the 480 SKUs rather than the 9 families alone. I tested it on 3 families: error per SKU is 21 %, that is 7 points above family level — which is normal and is the price of detail —, and it is enough to size production runs. Say yes and I hand you all 480 on Monday. six-month-review_error-and-depth.pdf27 % → 14 % on 54 pairs, €620,000 released
⛓ Sourced · 6 months of frozen forecasts and recorded actuals, monthly stock statements, service level by SKU
What is measured, and on what base:
· Error at 4 weeks: 27 % → 14 %, measured on 54 forecast/actual pairs — 9 families across 6 months of live operation, and consistent with the 14 % obtained on the 540-pair replay.
· Error at 8 weeks: 34 % → 19 %, same base.
· Service level: 94.1 % → 97.3 %, across 6 months of shipments.
· Stock cover: 41 days → 33 days, that is €620,000 of cash released, recorded in the monthly balance.
· Duration of the forecasting exercise: 9 days → 1.5 days. Building 60 % → 8 %, documenting assumptions 35 % → 4 %, neutralising events 30 % → 6 %.
The figure that does not flatter me: of the 9 families, 7 improved and 2 did not. Their error stays at 24 % and 22 %, against 27 % and 25 % at the start — a 3-point gain where the others gained 15.
The causes are clear, and they are not the same:
· The baby range has only 22 months of history: two annual cycles are not enough for seasonality to emerge. That is mechanical, and I have announced it from day one through a ± 14 % range.
· The professional family depends on a single customer for 61 % of its volume: it follows no seasonality, it follows an order book.
What I did with it, and what it gives:
· For the professional family, I plugged into the forward order book that customer sends you contractually. The error moves from 22 % to 12 % — but the depth is only 8 weeks, that is 2 pairs, and I will tell you again at the quarter when the base is five times wider.
· For the baby range, the range stays wide and deliberately so, and it tightens by itself: it moved from ± 14 % to ± 11 % in six months as the history lengthens. At 36 months seasonality will be readable and I will apply it.
What I propose: open forecasting to the 480 SKUs rather than the 9 families alone. I tested it on 3 families: error per SKU is 21 %, that is 7 points above family level — which is normal and is the price of detail —, and it is enough to size production runs. Say yes and I hand you all 480 on Monday. six-month-review_error-and-depth.pdf27 % → 14 % on 54 pairs, €620,000 released
⛓ Sourced · 6 months of frozen forecasts and recorded actuals, monthly stock statements, service level by SKU
What do you do without being asked?
Three gestures, and all three consist of measuring and telling. Anything that commits a volume to a third party lives in a mandate you write and withdraw in one word.
· Recompute the forecast on every week of sales that comes in, with its range and its depth. 26 recomputations in six months, published every Monday at 7 a.m.
· Detect an exceptional event and flag it before neutralising it. 9 detected in six months: 3 undeclared promotions, 4 shortages, 2 logistics incidents. Neutralisation only applies after your one-click confirmation — and 8 of the 9 were confirmed the same day.
· Publish the measured error every month, family by family, including when it worsens. That is how the two lagging families came to light, and it is what allowed them to be fixed.
What the mandate covers, and what it produced: the volume commitment to the four chains and its carry-over into the ERP. 6 commitments sent in six months, 0 above the cap settled in committee, median delay of 6 minutes after signature.
What I propose adding, costed: a capped supply-adjustment mandate — for instance ± 8 % of a family's volume, never above €40,000 per adjustment, with referenced suppliers only. Across the 4 useful alerts of these six months, that would have saved an average of 5 days between signal and order, and those 5 days are worth one shortage in two. You set the percentage and the cap, and the mandate is withdrawn in one word. framework-of-gestures_and-mandates.pdf3 measuring gestures, 6 commitments, 0 breaches
✎ Framework · recomputation log, register of declared events, commitment log
· Recompute the forecast on every week of sales that comes in, with its range and its depth. 26 recomputations in six months, published every Monday at 7 a.m.
· Detect an exceptional event and flag it before neutralising it. 9 detected in six months: 3 undeclared promotions, 4 shortages, 2 logistics incidents. Neutralisation only applies after your one-click confirmation — and 8 of the 9 were confirmed the same day.
· Publish the measured error every month, family by family, including when it worsens. That is how the two lagging families came to light, and it is what allowed them to be fixed.
What the mandate covers, and what it produced: the volume commitment to the four chains and its carry-over into the ERP. 6 commitments sent in six months, 0 above the cap settled in committee, median delay of 6 minutes after signature.
What I propose adding, costed: a capped supply-adjustment mandate — for instance ± 8 % of a family's volume, never above €40,000 per adjustment, with referenced suppliers only. Across the 4 useful alerts of these six months, that would have saved an average of 5 days between signal and order, and those 5 days are worth one shortage in two. You set the percentage and the cap, and the mandate is withdrawn in one word. framework-of-gestures_and-mandates.pdf3 measuring gestures, 6 commitments, 0 breaches
✎ Framework · recomputation log, register of declared events, commitment log
One last question: our volumes and our history, where do they go?
They stay with you. I run on local inference on a machine at the company, and your sales history crosses no outside network to be computed.
Local inference means the model computes on your machine: a sales line does not leave your network to be analysed. If you would rather not host a machine, the other route is an isolated resource hosted in France, under French law, dedicated to your company — no pooling with another manufacturer.
Why this is a commercial matter and not only a technical one: your volumes by chain, your promotional elasticities and your service levels are exactly what a buying group would pay to know. Knowing that your laundry family does + 187 % at a 30 % discount and falls back 41 % afterwards means knowing your negotiating floor before you do. Those figures stay inside your walls, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity, and they train no model.
What that looks like day to day: role-based access — rights follow the job: a supply planner opens their families, not the margins or the retailer agreements —, encryption in transit and at rest, and a full log: who consulted which forecast, who corrected which assumption, who settled which volume, to the second. That log is what makes a monthly committee unarguable three months later.
And you keep control wherever you are: a web dashboard, and supervision from your phone — you approve a commitment or withdraw a mandate in one message, the night before a committee as easily as on a Sunday evening.
The next step I propose: the 9 families are covered, the 480 SKUs are ready to be. And downstream, the procurement planning agent takes my forecasts straight into your supply plans — on your three tightest families I have already costed what the pair would give: 8 fewer days of cover at constant service level, that is €240,000 more cash. Say the word and I hand you the full study on Friday. technical-framework_where-your-volumes-live.pdfLocal inference, role-based access, processing in the EU targeted
✎ Framework · deployment architecture, role-based access matrix, consultation and arbitration logs
Local inference means the model computes on your machine: a sales line does not leave your network to be analysed. If you would rather not host a machine, the other route is an isolated resource hosted in France, under French law, dedicated to your company — no pooling with another manufacturer.
Why this is a commercial matter and not only a technical one: your volumes by chain, your promotional elasticities and your service levels are exactly what a buying group would pay to know. Knowing that your laundry family does + 187 % at a 30 % discount and falls back 41 % afterwards means knowing your negotiating floor before you do. Those figures stay inside your walls, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity, and they train no model.
What that looks like day to day: role-based access — rights follow the job: a supply planner opens their families, not the margins or the retailer agreements —, encryption in transit and at rest, and a full log: who consulted which forecast, who corrected which assumption, who settled which volume, to the second. That log is what makes a monthly committee unarguable three months later.
And you keep control wherever you are: a web dashboard, and supervision from your phone — you approve a commitment or withdraw a mandate in one message, the night before a committee as easily as on a Sunday evening.
The next step I propose: the 9 families are covered, the 480 SKUs are ready to be. And downstream, the procurement planning agent takes my forecasts straight into your supply plans — on your three tightest families I have already costed what the pair would give: 8 fewer days of cover at constant service level, that is €240,000 more cash. Say the word and I hand you the full study on Friday. technical-framework_where-your-volumes-live.pdfLocal inference, role-based access, processing in the EU targeted
✎ Framework · deployment architecture, role-based access matrix, consultation and arbitration logs
Local inference · no data outside the EU
Your case is not here? That is exactly what a 15-minute conversation is for. Book the free audit →
Use cases
What does the agent actually do?
One agent, several forecasting horizons. All these uses work in support, subject to your approval.
Included in your agent
The 3 capabilities essential to this promise are included, at no extra cost.
From 563 € excl. VAT / month
Forecasts by family
Builds the forecast from your history and your seasonality.
Assumptions shown
States the reference periods, the seasonality and the events taken into account.
Explicit ranges
Gives a low and a high figure, with what makes them move.
Controls and safeguards
These 3 controls are built into the agent: they frame what it does, whatever plan you pick. They are not chosen and are not added to your order.
Human validation, exceptions and escalation
Status, safe closure and audit trail
Explain assumptions, uncertainties and the limits of the analyses
What the agent must be connected to
This connection is required for the agent to work. It concerns your information system and is scoped during the audit.
Reconcile controlled sources with data quality and lineage
Need to go further?
These agents handle a different business process, with their own owner and their own price. They are added to this one.
Does your need fall outside this?
Book the free audit →
Build your agent
In 15 minutes we identify the most relevant agent — without oversizing the project.
The gain
How many families can a team forecast in detail?
By taking on the building and the documenting, the effort shifts towards deciding between scenarios. How large the gain is depends on your volume and remains to be confirmed by a pilot.
Building a forecast
Documenting the assumptions
Setting aside exceptional events
Indicative figures, not contractual, to be confirmed by a pilot on your number of families forecast and the depth of your history. Committing to a volume with a supplier or a customer is a matter of strategy: that decision belongs to management.
How it works
The stages of your AI agent project
1
Audit & scoping
15 minutes to target the use case with the best return.
2
Quote or direct sign-up
A catalogue offer is bought online; a specific need gets a costed quote.
3
Design
We design the agent and its guardrails.
4
Integration & testing
We connect your tools to the agent, which is itself hosted in France.
5
Rollout
Going live and training your team.
6
Operation
Continuous supervision and improvement.
Pricing
One package, one agent
A demand forecasting agent (history, assumptions, ranges), installed and operated for you.
Agility
Setup + controlled subscription
7,485 € excl. VAT setup
then 563 € excl. VAT/month — you invest at installation and pay a reduced subscription. Ideal for keeping the cost under control over time.
- Installation, configuration and training for your teams
- Operation, human oversight, updates and support
- Sovereign hosting in France, a dedicated and isolated resource
The simplest
Serenity
All inclusive, no setup fee
983 € excl. VAT /month
all inclusive, immediate start. No upfront investment: a single subscription. Ideal for starting quickly and simply.
- Setup included (installation, configuration, training)
- Operation, human oversight, updates and support
- Sovereign hosting in France, managed end to end
100% Sovereign
On site, you own it
11,150 € excl. VAT setup
then 760 € excl. VAT/month · + hardware from 2,491 € (one-off purchase, in addition) — a sovereign computer installed on your premises, maintained remotely. Models run locally, your data returned at the end of the contract. 36-month commitment.
- Hardware installed on your premises (you own it)
- French / European AI models run locally
- Secure remote maintenance (Pro support included)
Not included in the packages: AI consumption (model tokens), re-invoiced at real cost with no margin, and tracked in real time in your client area. Maintenance and supervision subscription for an initial term of 12 months for the Agility package, 24 months for the Serenity package and 36 months for the 100% Sovereign package, renewable; support levels (SLA 72 h / 24 h / 4 h) optional. Bespoke development, additional integrations or exceptional volumes are quoted separately. Support Monday to Friday, 9am to 6pm. Prices exclude VAT.
AI model:
none of the AI models offered currently carries a fixed surcharge. When the selected model carries a cost, that cost is shown when you choose it, before you order, and re-invoiced at the cost incurred, with no mark-up; usage is billed at the publisher's price. Publishers' prices are published in US dollars: the amount re-invoiced is the amount in euros actually borne by Blue Lemon Agent on the publisher's invoice, at that invoice's exchange rate, with no commission or mark-up.
Included components and additional components Components included in the base offer: the Blue Lemon Agent software foundation, the AI models listed in the order journey, the standard channels (Microsoft Teams, Slack, WhatsApp Business, email, website chat, calendars, Microsoft 365 / Google Workspace, file storage, market VoIP telephony, professional social-media pages and accounts, Google Business Profile), hosting in France for the package chosen, backups, supervision, updates and support. If adapting the AI agent to your constraints, your needs or your requests requires other paid components — a third-party publisher's software licence, paid API access to one of your applications, hosting of health data, for which French law requires an HDS-certified host (art. L. 1111-8 of the French Public Health Code), SecNumCloud-qualified hosting, a speech synthesis service, particular hardware —, they are offered to you as an option or on quotation and re-invoiced at the cost incurred; nothing is committed without your written agreement. Where the artificial intelligence model you choose entails an additional cost, that cost is shown to you before you order and re-invoiced to you at the cost incurred, with no margin.
What to expect
Go-live
2 to 3 weeks
Agent designed, channels connected, team trained.
Steady state
4 to 7 weeks
After a few weeks of real use, once the agent's behaviour matches what you expect.
Indicative estimate, adjusted to the options you keep. It is not a delivery commitment.
Our commitment
Four guarantees that matter to your forecasts
Your sales history stays with youLocal inference or an isolated resource hosted in France; no history and no volume entrusted to a third party, no data used to train a model.
Data in France, under French lawYour volumes and your sales data: minimisation and location in France, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
The supply chain department keeps the decisionThe agent produces demand forecasts that can be explained, checked and altered; no approval is automated.
Human oversight & traceabilityOn your number of families forecast and the depth of your history: systematic logging and tracking, in line with the AI Act.
Going further
Related resources
Frequently asked questions
Your questions, our answers
How do I know what a forecast rests on?
Every forecast shows its reference periods, its seasonality assumptions and the commercial events taken into account or set aside.
Why a range rather than a figure?
Because a range says what can make the result move, and lets you decide how far to commit. The central figure is provided with it.
Does the agent commit volumes?
No. It provides the forecast and its range; committing to a supplier or a customer rests with management.
How much history is needed?
The deeper the history, the clearer the seasonality. What can be used is assessed during the audit, with no assumptions about what you do not have.
Is our sales data protected?
Yes. The agent is hosted in France, on local inference or an isolated resource, with the deployment objective of processing and access operated within the European Union and an architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Your history is not used to train a third-party model.
How long does it take to deploy this agent?
A few months as a rule, depending on the number of families and the quality of the history available, after a free audit then phases of design, integration and testing.
Related use cases
Other agents for the supply chain
Let's talk
Let's size up the potential in your forecasts
15 minutes to frame your families and your history — hosted in France, supervised, with no commitment.