● B2B offer — Decisions & data analysis
Decision support: your data questioned in plain language
Putting a question to your data without waiting for an analyst to be free changes the pace of a management team. Your agent is connected to your consolidated data: it answers in plain language, produces the right chart and always states the data and the filter the result came from. Hosted in France: all of your management data stays with you. Management interprets and decides.
Hosted in France
Management data protected
GDPR & AI Act: governed deployment
Human oversight
Updated on
Deployed in a few weeks
Decision support · hosted in France
How has our business moved by region over twelve months?
Movement by region shown as a chart, over the last twelve months.
The scope used is stated: data source, filters applied, definition of the measure.
Two regions had their scope changed during the year: this is flagged to avoid a misreading.
🔗 Sourced · consolidated data, filters shown
The scope used is stated: data source, filters applied, definition of the measure.
Two regions had their scope changed during the year: this is flagged to avoid a misreading.
🔗 Sourced · consolidated data, filters shown
What is the cause of the gap in the East region?
I can break the gap down by product, by customer and by period, and show where it comes from in the data.
Attributing a cause calls for context the data does not contain: that reading belongs to management.
✎ Support · breakdown provided, human interpretation
Attributing a cause calls for context the data does not contain: that reading belongs to management.
✎ Support · breakdown provided, human interpretation
Local inference · no data outside the EU
Data hosted in France
Sovereign by designLocal inference or hosting in France
GDPR & AI Act: governed deploymentTraceability & human oversight
TurnkeyDesigned, installed and operated for you
Management decidesThe agent prepares, never rules
✦ In brief
A Blue Lemon Agent decision-support agent is connected to your consolidated data: it answers in plain language, produces the right chart and always shows the source, the filters and the definition of the measure. It breaks a gap down but does not attribute a cause to it. It runs on local inference or is hosted in France: your management 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
analysis 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 data sources and users is confirmed by a pilot.
The context
What does an AI agent bring to your decisions?
A question put to the data and an answer within the minute means a decision prepared in the same meeting.
! The issue
Deciding quickly calls for a fast answer and certainty about its scope. A figure without its filter or its definition has to be sent back for checking; a figure that comes with its source can be used straight away. The agent answers in plain language and systematically shows the source, the filters and the definition of the measure.
✓ Our answer
Management questions its data directly and gets the right chart, with the exact scope. The agent breaks a gap down by product, customer or period, but does not attribute a cause to it: that reading calls for market and organisational context the data does not contain. Local inference or an isolated resource hosted in France: your management data does not leave the company.
The decisive point
All of your management data: sovereignty & compliance
Your consolidated data gives a complete picture of the company: its protection is proportionate to its value. Here is how it is assured.
Local inference
The agent can run on a machine belonging to your organisation: no management data and no analysis result leaves the network.
Hosting in France
Otherwise, a dedicated and isolated resource hosted in France, under French law — your consolidated data and your measures: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
For all of your management 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 definitions of measures.
Source, filters and definition always shown
Every result states its data source, the filters applied and the definition of the measure; encryption, role-based access and logging of the analyses produced.
AI Act: governed deployment
The agent is strictly in support; no cause is attributed and no management decision is taken 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 consolidated data before publication and analyses tied to 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 companyMarnaud Distribution — wholesaler of industrial supplies and protective equipment
- Sector
- Business-to-business distribution — tools, consumables and personal protective equipment, across 5 regional branches
- Headcount
- 240 staff, including a 7-person executive committee and a management-control team of 2 — one controller and one apprentice
- Customers served
- 3,100 active business customers — manufacturing, construction, local authorities and public bodies
- Order of magnitude
- €96 M in revenue, 11,500 orders a month, average basket of €696
- Tools in place
- ERP, CRM, data warehouse, price lists in a spreadsheet and the general ledger — the agent plugs into them read-only, nothing is replaced
- Who decides
- The managing director arbitrates; the sales director sets regional targets; the management controller remains the owner of indicator definitions
- Room for improvement
- A question involving figures comes back in 9 working days; 60 % of the management-control team's time goes into producing answers; the executive committee therefore asks only 40 questions a quarter — and two regions changed scope mid-year without anyone knowing it while reading the curves
Marnaud Distribution is not short of data: it is short of questions asked. Each one costs nine days of waiting, so only forty are asked per quarter and the rest is decided on instinct. The agent is connected read-only to the company's seven management sources: it answers in plain language, produces the visualisation that fits and displays, for every figure, its source, its filters and the definition of the indicator. The exchanges below cover one quarter, from connecting the sources 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.
Decision support · hosted in France
I finished reading your seven management sources last night — three years of history, 411,000 order lines — and I found the first thing that is costing you meetings: « margin » means three different things here, depending on which system displays it.
Gross margin is the difference between the selling price and the purchase price of the goods. Its value changes depending on whether year-end supplier rebates and shipping costs are included or not.
The three definitions in use, and the gap they produce for the same month:
· ERP: selling price − list purchase price. September margin: 27.4 %.
· CRM: same calculation, but shipping costs deducted. September margin: 24.9 %.
· Data warehouse: purchase price net of year-end rebates. September margin: 31.2 %.
3.8 points between the sales figure and the management-control figure, on the same month and the same orders. This is not a keying error: these are three defensible definitions that were never arbitrated.
What it costs you, measured on your own minutes: 6 of the last 12 executive committees spent at least one agenda item settling which of two figures was right. The debate was never about the decision — it was about the vocabulary.
What I propose, and your controller is the one who decides: I have prepared an indicator dictionary — the written list of your indicators, each with its exact calculation, the sources used and the date the definition was settled. 34 indicators, 9 of which currently have two competing readings. You arbitrate nine, and every figure I produce speaks the same language from the next day. Each definition stays changeable in one word, and the history is recalculated on the new one. indicator-dictionary_34-definitions.pdf9 indicators to arbitrate, margin among them
⛓ Sourced · ERP, CRM, data warehouse, 411,000 lines over 3 years, minutes of the last 12 executive committees
Gross margin is the difference between the selling price and the purchase price of the goods. Its value changes depending on whether year-end supplier rebates and shipping costs are included or not.
The three definitions in use, and the gap they produce for the same month:
· ERP: selling price − list purchase price. September margin: 27.4 %.
· CRM: same calculation, but shipping costs deducted. September margin: 24.9 %.
· Data warehouse: purchase price net of year-end rebates. September margin: 31.2 %.
3.8 points between the sales figure and the management-control figure, on the same month and the same orders. This is not a keying error: these are three defensible definitions that were never arbitrated.
What it costs you, measured on your own minutes: 6 of the last 12 executive committees spent at least one agenda item settling which of two figures was right. The debate was never about the decision — it was about the vocabulary.
What I propose, and your controller is the one who decides: I have prepared an indicator dictionary — the written list of your indicators, each with its exact calculation, the sources used and the date the definition was settled. 34 indicators, 9 of which currently have two competing readings. You arbitrate nine, and every figure I produce speaks the same language from the next day. Each definition stays changeable in one word, and the history is recalculated on the new one. indicator-dictionary_34-definitions.pdf9 indicators to arbitrate, margin among them
⛓ Sourced · ERP, CRM, data warehouse, 411,000 lines over 3 years, minutes of the last 12 executive committees
Our figures come out of four or five systems that don't talk to each other. Do you really read all that — and do you write to it?
Seven sources, all read-only — I read your databases, I never write a single line to them. And I will also tell you what I do not read.
What I am connected to, with how fresh each one is:
· ERP — sales: orders, lines, deliveries. Refreshed hourly.
· ERP — purchasing and stock: purchase prices, stock-outs, supplier lead times. Hourly.
· CRM: accounts, contacts, live deals, portfolios per salesperson. Twice a day.
· Data warehouse: three years of consolidated history. A data warehouse is the database that brings together, in one place and one format, the figures coming from several systems. Refreshed overnight.
· Price lists in a spreadsheet: 14 grids, the only source still living outside the systems. That is the one holding the surprise you will see in tab 4.
· General ledger: to reconcile the management figure with the accounting figure.
· Customer reference data: families, sectors, outstanding balances.
What I do not read — and this is not caution, it is the scope you set: payroll, personnel files and mailboxes. They are not connected, so no question on those subjects will find an answer here, even asked by the managing director.
What read-only guarantees in practice: I cannot correct an order, change a price or close a period. The worst incident possible on my side is a wrong figure, never damaged data — and a wrong figure shows, because it arrives with its source.
What I propose: that your controller reviews the list of seven sources and removes any he does not want connected. Removing a source takes a minute and breaks nothing: the indicators that depended on it then display as unavailable, not as approximate. connected-sources_7-read-only.pdfWhat is read, how often, and what is not
✎ Framework · read-only connection, access log, scope agreed with management control
What I am connected to, with how fresh each one is:
· ERP — sales: orders, lines, deliveries. Refreshed hourly.
· ERP — purchasing and stock: purchase prices, stock-outs, supplier lead times. Hourly.
· CRM: accounts, contacts, live deals, portfolios per salesperson. Twice a day.
· Data warehouse: three years of consolidated history. A data warehouse is the database that brings together, in one place and one format, the figures coming from several systems. Refreshed overnight.
· Price lists in a spreadsheet: 14 grids, the only source still living outside the systems. That is the one holding the surprise you will see in tab 4.
· General ledger: to reconcile the management figure with the accounting figure.
· Customer reference data: families, sectors, outstanding balances.
What I do not read — and this is not caution, it is the scope you set: payroll, personnel files and mailboxes. They are not connected, so no question on those subjects will find an answer here, even asked by the managing director.
What read-only guarantees in practice: I cannot correct an order, change a price or close a period. The worst incident possible on my side is a wrong figure, never damaged data — and a wrong figure shows, because it arrives with its source.
What I propose: that your controller reviews the list of seven sources and removes any he does not want connected. Removing a source takes a minute and breaks nothing: the indicators that depended on it then display as unavailable, not as approximate. connected-sources_7-read-only.pdfWhat is read, how often, and what is not
✎ Framework · read-only connection, access log, scope agreed with management control
The real issue here is the delay. A question to management control takes nine days.
Nine working days yesterday, under four minutes today — and the figure arrives with its scope, so it does not go back out for checking.
Where the nine days went, according to your management-control timesheets: two days of queue, three days finding the right source and writing the query, one day of formatting, and three days answering the question that always follows: « what exactly does this figure include? »
The shift this produces on the three items you were measuring:
· Getting a figure: 60 % of the time of an analysis request yesterday, 7 % today. The heaviest item, and the most mechanical.
· Producing the visualisation: 35 % → 5 %.
· Checking a figure's scope: 30 % → 4 %. It does not vanish: it becomes the review of a scope sheet that is already written.
And the figure that actually matters to an executive team, because it is not about time but about decisions: you were asking 40 quantified questions a quarter. Not because you only had forty — because each one cost nine days, so everything else was settled on instinct. The ceiling was not your curiosity, it was the delay.
What I propose as the next step: that the nine definitions be arbitrated this week by your controller, and only then that I open access to the five branch managers — in that order, never the reverse. Opening before arbitrating would mean handing three margins to five people, and you would be right to stop me. The time management control gets back, I suggest spending where it is worth most: variance analysis, which we open in the next tab. response-time_from-9-days-to-4-minutes.pdfWhere the nine days went, step by step
⛓ Sourced · management-control timesheets, register of analysis requests over 12 months
Where the nine days went, according to your management-control timesheets: two days of queue, three days finding the right source and writing the query, one day of formatting, and three days answering the question that always follows: « what exactly does this figure include? »
The shift this produces on the three items you were measuring:
· Getting a figure: 60 % of the time of an analysis request yesterday, 7 % today. The heaviest item, and the most mechanical.
· Producing the visualisation: 35 % → 5 %.
· Checking a figure's scope: 30 % → 4 %. It does not vanish: it becomes the review of a scope sheet that is already written.
And the figure that actually matters to an executive team, because it is not about time but about decisions: you were asking 40 quantified questions a quarter. Not because you only had forty — because each one cost nine days, so everything else was settled on instinct. The ceiling was not your curiosity, it was the delay.
What I propose as the next step: that the nine definitions be arbitrated this week by your controller, and only then that I open access to the five branch managers — in that order, never the reverse. Opening before arbitrating would mean handing three margins to five people, and you would be right to stop me. The time management control gets back, I suggest spending where it is worth most: variance analysis, which we open in the next tab. response-time_from-9-days-to-4-minutes.pdfWhere the nine days went, step by step
⛓ Sourced · management-control timesheets, register of analysis requests over 12 months
How has our business developed by region over twelve months?
Here are the five curves, and I will start with the trap, because it changes how two of them read.
The scope of the figure, shown before the figure: source, data warehouse; indicator, revenue invoiced excluding tax; filters, orders delivered and invoiced, credit notes deducted, intercompany sales excluded; period, 12 rolling months ending 31/10.
The five regions, twelve-month change:
· Île-de-France +6.8 % · West +5.1 % · North +3.4 % · South-East +11.2 % · East −14.0 %
And here is why I will not let you read those last two as they stand: on 1 April, three départements moved from the East branch to the South-East branch. The transfer represents €8.9 M of annual revenue.
On a like-for-like basis — that is, comparing the two periods as if the organisation had not changed, so that the measured gap comes from activity and not from a transfer of customers — the East is at −3.2 % and the South-East at +2.6 %.
What that changes for your committee: read raw, the South-East is your best region and the East is collapsing. Read like-for-like, the South-East is average and the East has a real issue, but four times smaller than the curve says. Those are two different meetings.
What I propose: that the chart goes to the committee with both readings side by side, raw and like-for-like. I have built it that way. And if you would rather show only the like-for-like reading, say so: that is a management choice, not a calculation choice. regional-trend_12-months-two-readings.pdfRaw and like-for-like, with the transfer dated
⛓ Sourced · data warehouse, customer reference data, reorganisation memo of 1 April
The scope of the figure, shown before the figure: source, data warehouse; indicator, revenue invoiced excluding tax; filters, orders delivered and invoiced, credit notes deducted, intercompany sales excluded; period, 12 rolling months ending 31/10.
The five regions, twelve-month change:
· Île-de-France +6.8 % · West +5.1 % · North +3.4 % · South-East +11.2 % · East −14.0 %
And here is why I will not let you read those last two as they stand: on 1 April, three départements moved from the East branch to the South-East branch. The transfer represents €8.9 M of annual revenue.
On a like-for-like basis — that is, comparing the two periods as if the organisation had not changed, so that the measured gap comes from activity and not from a transfer of customers — the East is at −3.2 % and the South-East at +2.6 %.
What that changes for your committee: read raw, the South-East is your best region and the East is collapsing. Read like-for-like, the South-East is average and the East has a real issue, but four times smaller than the curve says. Those are two different meetings.
What I propose: that the chart goes to the committee with both readings side by side, raw and like-for-like. I have built it that way. And if you would rather show only the like-for-like reading, say so: that is a management choice, not a calculation choice. regional-trend_12-months-two-readings.pdfRaw and like-for-like, with the transfer dated
⛓ Sourced · data warehouse, customer reference data, reorganisation memo of 1 April
Why curves? Management control always gives us tables.
Because your question was about a trend, and a trend reads at a glance on a curve and in thirty seconds on a table. But the table is there too: it sits under the chart in the same document.
The rule I apply, and you can change it:
· A trend over time → curve. Five series maximum on one chart, beyond that nothing is distinguishable.
· A comparison between categories → horizontal bars, sorted by value. A ranking is read sorted, never in the alphabetical order of the branches.
· A composition — what a total is made of → stacked bars, never a pie chart beyond four slices.
· A relationship between two quantities → scatter plot. That is the one you were missing last month on discount versus volume.
· A single figure → a single figure, large, with its variation. A chart with one value is decoration.
What formatting was costing you: 35 % of the time of an analysis request went into building the visual — copying data into a spreadsheet, restoring the company colours, fixing the scales. It now takes 5 %, and those 5 % are your review.
What is already applied without your asking: your colours, an axis starting at zero — a truncated axis turns +3.4 % into a surge, and it is the chart that lies, not the data —, the unit written out in full, and the figure's cut-off date in the footer.
What I propose: tell me once and for all whether your committee prefers the chart first and the table below, or the reverse. I build both in four minutes, so it may as well be in your order. visualisation-rules_which-form-for-which-question.pdfFive typical questions, five forms, and the scale traps
✎ Framework · visualisation rules, company graphic charter
The rule I apply, and you can change it:
· A trend over time → curve. Five series maximum on one chart, beyond that nothing is distinguishable.
· A comparison between categories → horizontal bars, sorted by value. A ranking is read sorted, never in the alphabetical order of the branches.
· A composition — what a total is made of → stacked bars, never a pie chart beyond four slices.
· A relationship between two quantities → scatter plot. That is the one you were missing last month on discount versus volume.
· A single figure → a single figure, large, with its variation. A chart with one value is decoration.
What formatting was costing you: 35 % of the time of an analysis request went into building the visual — copying data into a spreadsheet, restoring the company colours, fixing the scales. It now takes 5 %, and those 5 % are your review.
What is already applied without your asking: your colours, an axis starting at zero — a truncated axis turns +3.4 % into a surge, and it is the chart that lies, not the data —, the unit written out in full, and the figure's cut-off date in the footer.
What I propose: tell me once and for all whether your committee prefers the chart first and the table below, or the reverse. I build both in four minutes, so it may as well be in your order. visualisation-rules_which-form-for-which-question.pdfFive typical questions, five forms, and the scale traps
✎ Framework · visualisation rules, company graphic charter
Can I send it as it is to Thursday's executive committee?
Yes, with nothing to retouch: the document comes out as one page, with the scope printed under the chart and the figure's cut-off date. That is what makes it stand up in a meeting.
What the page carries, and why each line is on it:
· The chart and its table, in whichever order you settle on.
· The source, the filters and the definition of the indicator, written out. That is the line that removes the three days of « what does this figure include? ».
· The 1 April transfer of départements, dated and quantified, on the same page as the curve it explains — not in a note that will be read after the meeting.
· The question's identifier, which lets you ask it again identically in six months and get exactly the same figure.
What is not on it, and I say so before anyone holds it against me: no full-year projection. You have two months of strong seasonality ahead and I only have three years of history; an annual projection on that basis would display a precision it does not have. I prefer a short, correct page to a complete, decorative one.
The next step I propose, and it fits in one question: the East is at −3.2 % like-for-like while the group is at +4.1 %. Ask me where that gap comes from and I will break it down to the order line — which is exactly what we do in the next tab.
⛓ Sourced · data warehouse, reorganisation memo of 1 April, log of questions asked
What the page carries, and why each line is on it:
· The chart and its table, in whichever order you settle on.
· The source, the filters and the definition of the indicator, written out. That is the line that removes the three days of « what does this figure include? ».
· The 1 April transfer of départements, dated and quantified, on the same page as the curve it explains — not in a note that will be read after the meeting.
· The question's identifier, which lets you ask it again identically in six months and get exactly the same figure.
What is not on it, and I say so before anyone holds it against me: no full-year projection. You have two months of strong seasonality ahead and I only have three years of history; an annual projection on that basis would display a precision it does not have. I prefer a short, correct page to a complete, decorative one.
The next step I propose, and it fits in one question: the East is at −3.2 % like-for-like while the group is at +4.1 %. Ask me where that gap comes from and I will break it down to the order line — which is exactly what we do in the next tab.
⛓ Sourced · data warehouse, reorganisation memo of 1 April, log of questions asked
And on other questions, do you always produce line charts?
No: the right visualisation is the one the question calls for, and I have six forms, not one.
· A trend — line chart. That was your question a moment ago.
· A comparison between entities — bars sorted largest to smallest. "Which branch sells the most protective equipment" reads in a second that way, in ten on a line chart.
· A breakdown — 100% stacked bars, and not a pie as soon as there are more than seven slices.
· A variance decomposition — waterfall: that is the one used for the −3.2% in the East, because it shows what adds and what subtracts, in order.
· A relationship between two quantities — scatter plot. Discount granted against volume ordered, across your 3,100 customers: the scatter shows the 38 customers off the price list at a glance.
· A single figure that matters — no chart at all: the figure large, its scope beneath it, its cut-off date beside it. A chart with one value is decoration.
Every visualisation carries the same three notes as a bare figure: source, filters applied, cut-off date. An image with no scope goes back for checking exactly as a table with no scope does.
One visualisation I got wrong, published too: asked how turnover splits by product family, I produced a pie with 14 slices, 9 of them under 3% — unreadable, and the board asked for a table. Beyond seven slices, a breakdown now comes out as sorted bars, with an explicit "other" group that can be opened up: across the quarter's 1,460 questions, 19 visualisations were asked for again in another form, against 61 in the first month.
✎ Framework · the form follows the question — six visualisations, one rule per question
· A trend — line chart. That was your question a moment ago.
· A comparison between entities — bars sorted largest to smallest. "Which branch sells the most protective equipment" reads in a second that way, in ten on a line chart.
· A breakdown — 100% stacked bars, and not a pie as soon as there are more than seven slices.
· A variance decomposition — waterfall: that is the one used for the −3.2% in the East, because it shows what adds and what subtracts, in order.
· A relationship between two quantities — scatter plot. Discount granted against volume ordered, across your 3,100 customers: the scatter shows the 38 customers off the price list at a glance.
· A single figure that matters — no chart at all: the figure large, its scope beneath it, its cut-off date beside it. A chart with one value is decoration.
Every visualisation carries the same three notes as a bare figure: source, filters applied, cut-off date. An image with no scope goes back for checking exactly as a table with no scope does.
One visualisation I got wrong, published too: asked how turnover splits by product family, I produced a pie with 14 slices, 9 of them under 3% — unreadable, and the board asked for a table. Beyond seven slices, a breakdown now comes out as sorted bars, with an explicit "other" group that can be opened up: across the quarter's 1,460 questions, 19 visualisations were asked for again in another form, against 61 in the first month.
✎ Framework · the form follows the question — six visualisations, one rule per question
Last month sales and management control produced two different figures for the same thing. We spent the meeting arbitrating between two tables.
Both figures were right. That is the worst case, and the fastest to settle: they were not counting the same thing, and nothing on the page said so.
What I attach to every answer, in five lines: the source queried · the filters applied, one by one · the definition of the indicator and the date it was settled · the period and its cut-off date · any scope changes that occurred during the period.
And the point that goes further than display: every answer carries an identifier, and the query that produced it is kept exactly as it ran. The same question asked again in six months returns the same figure, even if a salesperson has changed portfolio in the meantime. That is what a reproducible figure is: one that does not depend on the day it was asked for.
What that gives you back, and it is the third item you were measuring: checking a figure's scope took 30 % of the time of an analysis request. It now takes 4 % — the time to read a sheet that is already written. What disappears is not the checking, it is the rebuilding.
What I propose: that last month's two tables come back through me. I will give you both figures with both scopes, and the line-by-line bridge from one to the other. The meeting will not have been wasted if it produces the definition arbitration that was missing. scope-sheet_anatomy-of-a-figure.pdfFive lines that make a figure stand up
⛓ Sourced · log of questions asked, dated indicator definitions
What I attach to every answer, in five lines: the source queried · the filters applied, one by one · the definition of the indicator and the date it was settled · the period and its cut-off date · any scope changes that occurred during the period.
And the point that goes further than display: every answer carries an identifier, and the query that produced it is kept exactly as it ran. The same question asked again in six months returns the same figure, even if a salesperson has changed portfolio in the meantime. That is what a reproducible figure is: one that does not depend on the day it was asked for.
What that gives you back, and it is the third item you were measuring: checking a figure's scope took 30 % of the time of an analysis request. It now takes 4 % — the time to read a sheet that is already written. What disappears is not the checking, it is the rebuilding.
What I propose: that last month's two tables come back through me. I will give you both figures with both scopes, and the line-by-line bridge from one to the other. The meeting will not have been wasted if it produces the definition arbitration that was missing. scope-sheet_anatomy-of-a-figure.pdfFive lines that make a figure stand up
⛓ Sourced · log of questions asked, dated indicator definitions
And if we change a definition mid-year? We did exactly that on margin on 1 March.
The 1 March change is already recorded, dated, and the history is recalculated on both definitions. You can read your twelve months in the old one as in the new one.
What changed that day: year-end rebates — the discounts negotiated with a supplier and paid afterwards, based on the volume purchased during the year — entered the purchase-price calculation. Decision taken by the management controller on 26/02, applied on 1 March.
The effect, quantified: the reported margin rises by 1.8 points at the switch. Over twelve rolling months, that means February and March are not comparable until a single definition is chosen for both. This is exactly the kind of gap a committee attributes to sales performance when it comes from a calculation rule.
What I now do automatically: as soon as an answer crosses a definition-change date, the scope sheet flags it at the top, with the date, the decision and its author. You do not have to remember it: the figure remembers it.
What I propose: a quarterly definition review with your controller. Nine of your 34 indicators still have two possible readings; at that pace none will be left by spring. And every arbitration recalculates the history overnight — you never lose the series.
What changed that day: year-end rebates — the discounts negotiated with a supplier and paid afterwards, based on the volume purchased during the year — entered the purchase-price calculation. Decision taken by the management controller on 26/02, applied on 1 March.
The effect, quantified: the reported margin rises by 1.8 points at the switch. Over twelve rolling months, that means February and March are not comparable until a single definition is chosen for both. This is exactly the kind of gap a committee attributes to sales performance when it comes from a calculation rule.
What I now do automatically: as soon as an answer crosses a definition-change date, the scope sheet flags it at the top, with the date, the decision and its author. You do not have to remember it: the figure remembers it.
What I propose: a quarterly definition review with your controller. Nine of your 34 indicators still have two possible readings; at that pace none will be left by spring. And every arbitration recalculates the history overnight — you never lose the series.
All this assumes you read our entire management data: margins, customers, purchase prices. Where does it go, exactly?
It does not go anywhere. I run on local inference on a machine at your head office, and none of your management data is copied elsewhere.
Local inference means the model computes on your machine: the contents of an order or a purchase price cross no external network to be processed. If you would rather not host a machine, the other route is an isolated resource hosted in France, dedicated to your company — no pooling with another client, nor with a competitor in your sector.
What protects your data, point by point:
· Your figures train no public model. What I learn from your margins serves your margins. Nothing you entrust to me surfaces anywhere else.
· Encryption in transit and at rest, and role-based access — rights follow the job: a branch manager sees his region, the controller sees everything, the apprentice sees volumes and not purchase prices.
· Logging of the analyses produced: who asked what, when, on which scope. That log is what lets you answer an auditor without reopening twelve months of email.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity — including against a US provider hosting in Europe.
· Read-only, mentioned here because it is also a protection: no database of yours can be modified by a question.
The figure that makes this commercial rather than technical: public-sector customers are 21 % of your revenue, and your last public contract carried a data-localisation clause. You were able to sign it without reservation and without a derogation annex — a first.
What I propose: that I keep up to date the technical sheet your key accounts ask for at every renewal — hosting, subprocessors, retention periods, who accesses what. It is asked for once a year and takes three days to assemble. technical-framework_where-your-management-data-lives.pdfLocal inference, role-based access, processing in the EU targeted
✎ Framework · deployment architecture, access log, clauses of the public contracts in force
Local inference means the model computes on your machine: the contents of an order or a purchase price cross no external network to be processed. If you would rather not host a machine, the other route is an isolated resource hosted in France, dedicated to your company — no pooling with another client, nor with a competitor in your sector.
What protects your data, point by point:
· Your figures train no public model. What I learn from your margins serves your margins. Nothing you entrust to me surfaces anywhere else.
· Encryption in transit and at rest, and role-based access — rights follow the job: a branch manager sees his region, the controller sees everything, the apprentice sees volumes and not purchase prices.
· Logging of the analyses produced: who asked what, when, on which scope. That log is what lets you answer an auditor without reopening twelve months of email.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity — including against a US provider hosting in Europe.
· Read-only, mentioned here because it is also a protection: no database of yours can be modified by a question.
The figure that makes this commercial rather than technical: public-sector customers are 21 % of your revenue, and your last public contract carried a data-localisation clause. You were able to sign it without reservation and without a derogation annex — a first.
What I propose: that I keep up to date the technical sheet your key accounts ask for at every renewal — hosting, subprocessors, retention periods, who accesses what. It is asked for once a year and takes three days to assemble. technical-framework_where-your-management-data-lives.pdfLocal inference, role-based access, processing in the EU targeted
✎ Framework · deployment architecture, access log, clauses of the public contracts in force
What is the cause of the variance in the East region?
I break the variance down to the order line and locate 83 % of it in two pockets. The cause itself comes at the end — and that reading belongs to you.
The starting point: East at −3.2 % over twelve months like-for-like, while the group is at +4.1 %. The gap to explain is 7.3 points, or €1.42 M of revenue.
Where it sits once cut up:
· 62 % of the gap comes from 9 construction customers out of the region's 410 active customers. Nine. Their combined revenue falls from €1.68 M to €0.80 M.
· 21 % comes from a single product family: personal protective equipment, −28 % in the region, +6 % elsewhere.
· 17 % is diffuse — spread over 190 customers, with no recognisable pattern. I leave it diffuse rather than invent an explanation for it.
What the breakdown already says, which is a fact and not a cause: across the 9 construction customers, the number of orders is divided by 2.1 after June, while the average amount per order does not move (€712 against €704 before). They are not buying cheaper: they are ordering less often. That is not the same problem, and not the same answers.
What I propose: that we take the three possible readings of this fact one by one, with what would confirm or rule out each. I have already tested two of them. east-variance-breakdown_down-to-the-line.pdf7.3 points of gap, 83 % located in two pockets
⛓ Sourced · data warehouse, 411,000 order lines, customer reference data, stock-out history
The starting point: East at −3.2 % over twelve months like-for-like, while the group is at +4.1 %. The gap to explain is 7.3 points, or €1.42 M of revenue.
Where it sits once cut up:
· 62 % of the gap comes from 9 construction customers out of the region's 410 active customers. Nine. Their combined revenue falls from €1.68 M to €0.80 M.
· 21 % comes from a single product family: personal protective equipment, −28 % in the region, +6 % elsewhere.
· 17 % is diffuse — spread over 190 customers, with no recognisable pattern. I leave it diffuse rather than invent an explanation for it.
What the breakdown already says, which is a fact and not a cause: across the 9 construction customers, the number of orders is divided by 2.1 after June, while the average amount per order does not move (€712 against €704 before). They are not buying cheaper: they are ordering less often. That is not the same problem, and not the same answers.
What I propose: that we take the three possible readings of this fact one by one, with what would confirm or rule out each. I have already tested two of them. east-variance-breakdown_down-to-the-line.pdf7.3 points of gap, 83 % located in two pockets
⛓ Sourced · data warehouse, 411,000 order lines, customer reference data, stock-out history
So it's the construction market falling in the East.
That is one of three readings, and the one your own data weakens most. Here are all three, what supports them, and the test I have already run on each.
Reading 1 — the local market is contracting. Only verifiable by comparison: I built a control group of 40 comparable construction customers — same size, same purchase families, other regions. They are at +2.8 % over the same period. If the construction market were carrying the gap, it would show elsewhere too. It does not.
Reading 2 — product availability. Verified on your stock: the stock-out rate — the share of ordered lines that could not be delivered for lack of stock — reaches 11.4 % on protective equipment at the East branch, against 3.1 % at the other branches. That one is confirmed, and it explains precisely the 21 % pocket — the product family is the same.
Reading 3 — sales coverage. Dated fact: the salesperson who covered 7 of the 9 customers left the company on 14 June, and the portfolio was split between two already-loaded colleagues. The drop in ordering frequency starts in July. I cannot attribute it to him: coincidence is not causation, and your data does not contain the visits that did not happen.
What this gives, plainly: the 21 % pocket has a measured, verified explanation. The 62 % pocket has two readings still open, and it is for the sales management to decide, because deciding requires knowing what happened in the field — information that is in none of your databases.
What I can do so the decision does not rest on an impression, and I have prepared it: a re-contact of the 9 customers, each with an up-to-date sheet — last order, families bought, stock-outs suffered, your share of their activity. Three questions are enough to separate readings 1 and 3, and they can be asked in one afternoon. The meeting belongs to your salespeople; the file is ready. three-readings_and-what-separates-them.pdfOne confirmed, one weakened, one that belongs to the field
⛓ Sourced · control group of 40 comparable customers, stock-out history by branch, sales staff register
Reading 1 — the local market is contracting. Only verifiable by comparison: I built a control group of 40 comparable construction customers — same size, same purchase families, other regions. They are at +2.8 % over the same period. If the construction market were carrying the gap, it would show elsewhere too. It does not.
Reading 2 — product availability. Verified on your stock: the stock-out rate — the share of ordered lines that could not be delivered for lack of stock — reaches 11.4 % on protective equipment at the East branch, against 3.1 % at the other branches. That one is confirmed, and it explains precisely the 21 % pocket — the product family is the same.
Reading 3 — sales coverage. Dated fact: the salesperson who covered 7 of the 9 customers left the company on 14 June, and the portfolio was split between two already-loaded colleagues. The drop in ordering frequency starts in July. I cannot attribute it to him: coincidence is not causation, and your data does not contain the visits that did not happen.
What this gives, plainly: the 21 % pocket has a measured, verified explanation. The 62 % pocket has two readings still open, and it is for the sales management to decide, because deciding requires knowing what happened in the field — information that is in none of your databases.
What I can do so the decision does not rest on an impression, and I have prepared it: a re-contact of the 9 customers, each with an up-to-date sheet — last order, families bought, stock-outs suffered, your share of their activity. Three questions are enough to separate readings 1 and 3, and they can be asked in one afternoon. The meeting belongs to your salespeople; the file is ready. three-readings_and-what-separates-them.pdfOne confirmed, one weakened, one that belongs to the field
⛓ Sourced · control group of 40 comparable customers, stock-out history by branch, sales staff register
While digging through all that, did you come across anything else?
Yes, and it was not the question asked: 38 customers are invoiced below your price list, for €214,000 of margin over twelve months.
How I saw it: by cross-checking the 14 price grids in the spreadsheet against the prices actually invoiced, line by line. The spreadsheet was reconciled with invoicing by nobody — it is the only one of your seven sources living outside the systems.
What it is, and what it is not: no fraud, no keying error. These are exceptional discounts granted between 2023 and 2025 and never reviewed. 31 of the 38 discounts are more than eighteen months old. The commercial gesture was justified the day it was made; it simply has no end date.
What is left to your signature, and nothing else: the increase itself. A price revision engages your relationship with a customer and your revenue — that is a management decision, and it is yours. Everything that comes before it is already done, and I show it to you priced.
What I did instead, and it is ready: the 38 revision letters are written, one per customer, each with its discount history, the new price, the effective date and the notice period. Priced customer by customer, they represent €214,000 of annual margin.
What I propose, and you set the limits: a written mandate, capped, dated and withdrawable in one word. The limits I suggest: increase capped at +2.4 margin points per customer · the 6 accounts above €500,000 excluded, they are handled in a meeting, not by letter · 30 days' notice · mandate valid until 31/12, with no tacit renewal.
What that gives within those limits: 31 customers out of 38 qualify, i.e. €168,000 of the €214,000. The other 7 stay in draft for your sales director, with the argument prepared. The day the mandate is signed, the 31 letters go out in twenty minutes; the day you withdraw it, everything stops instantly and unsent letters are cancelled. price-revision-mandate_38-customers.pdf€214,000 identified, €168,000 within the proposed limits
⛓ Sourced · 14 price grids, invoiced lines over 24 months, history of exceptional discounts
How I saw it: by cross-checking the 14 price grids in the spreadsheet against the prices actually invoiced, line by line. The spreadsheet was reconciled with invoicing by nobody — it is the only one of your seven sources living outside the systems.
What it is, and what it is not: no fraud, no keying error. These are exceptional discounts granted between 2023 and 2025 and never reviewed. 31 of the 38 discounts are more than eighteen months old. The commercial gesture was justified the day it was made; it simply has no end date.
What is left to your signature, and nothing else: the increase itself. A price revision engages your relationship with a customer and your revenue — that is a management decision, and it is yours. Everything that comes before it is already done, and I show it to you priced.
What I did instead, and it is ready: the 38 revision letters are written, one per customer, each with its discount history, the new price, the effective date and the notice period. Priced customer by customer, they represent €214,000 of annual margin.
What I propose, and you set the limits: a written mandate, capped, dated and withdrawable in one word. The limits I suggest: increase capped at +2.4 margin points per customer · the 6 accounts above €500,000 excluded, they are handled in a meeting, not by letter · 30 days' notice · mandate valid until 31/12, with no tacit renewal.
What that gives within those limits: 31 customers out of 38 qualify, i.e. €168,000 of the €214,000. The other 7 stay in draft for your sales director, with the argument prepared. The day the mandate is signed, the 31 letters go out in twenty minutes; the day you withdraw it, everything stops instantly and unsent letters are cancelled. price-revision-mandate_38-customers.pdf€214,000 identified, €168,000 within the proposed limits
⛓ Sourced · 14 price grids, invoiced lines over 24 months, history of exceptional discounts
After a quarter, what has actually changed for the executive team?
You were asking 40 quantified questions a quarter. You asked 1,460. It is the only figure in this review that describes a change of behaviour rather than a saving of time.
The three items you were measuring, over the quarter:
· Getting a figure: 60 % → 7 % of the time of an analysis request.
· Producing the visualisation: 35 % → 5 %.
· Checking a figure's scope: 30 % → 4 %.
· Average time to an answer: 9 working days → under 4 minutes.
What management control did with it, per its timesheets: time spent producing answers falls from 60 % to 18 % of the month; time spent analysing variances and arbitrating definitions rises from 9 % to 44 %. You did not gain half a headcount, you changed what the headcount does — and only one of your two management-control staff could do the second thing.
What showed up outside the timesheets: €214,000 of margin identified on the price lists, of which €168,000 within the limits of the proposed mandate · 7.3 points of regional variance explained at 83 % · 9 indicator definitions arbitrated out of 34, margin among them, which had taken one agenda item in two for a year.
What I propose for the coming quarter: open access to the five branch managers, now that the definitions are settled. Each will see his region and the group averages, never the detail of the other regions — that is the setting that makes a shared dashboard a tool for comparing rather than for watching. quarterly-review_from-40-questions-to-1460.pdfWhat the executive team gained, item by item
⛓ Sourced · log of questions asked, management-control timesheets, register of arbitrated definitions
The three items you were measuring, over the quarter:
· Getting a figure: 60 % → 7 % of the time of an analysis request.
· Producing the visualisation: 35 % → 5 %.
· Checking a figure's scope: 30 % → 4 %.
· Average time to an answer: 9 working days → under 4 minutes.
What management control did with it, per its timesheets: time spent producing answers falls from 60 % to 18 % of the month; time spent analysing variances and arbitrating definitions rises from 9 % to 44 %. You did not gain half a headcount, you changed what the headcount does — and only one of your two management-control staff could do the second thing.
What showed up outside the timesheets: €214,000 of margin identified on the price lists, of which €168,000 within the limits of the proposed mandate · 7.3 points of regional variance explained at 83 % · 9 indicator definitions arbitrated out of 34, margin among them, which had taken one agenda item in two for a year.
What I propose for the coming quarter: open access to the five branch managers, now that the definitions are settled. Each will see his region and the group averages, never the detail of the other regions — that is the setting that makes a shared dashboard a tool for comparing rather than for watching. quarterly-review_from-40-questions-to-1460.pdfWhat the executive team gained, item by item
⛓ Sourced · log of questions asked, management-control timesheets, register of arbitrated definitions
And when you get it wrong? Because that must happen.
118 of my 1,460 answers were corrected by your controller this quarter — 8 %. The figure is published every month, and I will give you its real cause, which is not the one people assume.
What the 118 had in common: 94 were about margin. Not a wrong calculation — a choice of definition, at a time when three coexisted and none had been arbitrated. I answered with the data-warehouse definition; sales read with the CRM one. Both figures were exact; the disagreement was about the question.
The other 24: 15 ambiguous questions — « our sales » meant invoiced to some, ordered to others — and 9 genuine errors of mine, all on periods straddling the 1 April branch transfer.
What was done, and what it produced: the margin definition was arbitrated by your controller on 26/02; over the following six weeks, corrections about margin fell from 94 to 7. On ambiguous questions, I now ask which of the two readings you want before answering, rather than choosing for you — a ten-second question beats a four-minute answer that has to be redone.
What I keep publishing every month, unasked: the number of corrected answers, their cause, and the indicator concerned. A correction rate that is not measured becomes a correction rate that does not fall.
What I propose: work through the remaining 25 definitions at three a month. At that pace the correction rate should follow the same curve as margin: the causes are the same. answer-quality_118-corrections.pdf8 % corrected, the real cause and its measured fix
⛓ Sourced · log of questions and corrections, register of definition arbitrations
What the 118 had in common: 94 were about margin. Not a wrong calculation — a choice of definition, at a time when three coexisted and none had been arbitrated. I answered with the data-warehouse definition; sales read with the CRM one. Both figures were exact; the disagreement was about the question.
The other 24: 15 ambiguous questions — « our sales » meant invoiced to some, ordered to others — and 9 genuine errors of mine, all on periods straddling the 1 April branch transfer.
What was done, and what it produced: the margin definition was arbitrated by your controller on 26/02; over the following six weeks, corrections about margin fell from 94 to 7. On ambiguous questions, I now ask which of the two readings you want before answering, rather than choosing for you — a ten-second question beats a four-minute answer that has to be redone.
What I keep publishing every month, unasked: the number of corrected answers, their cause, and the indicator concerned. A correction rate that is not measured becomes a correction rate that does not fall.
What I propose: work through the remaining 25 definitions at three a month. At that pace the correction rate should follow the same curve as margin: the causes are the same. answer-quality_118-corrections.pdf8 % corrected, the real cause and its measured fix
⛓ Sourced · log of questions and corrections, register of definition arbitrations
One last thing: what do you do without asking us?
Three things, not one more, and each is reversible. Everything else waits for a decision — including the decisions that would make you money.
The three automatic actions:
· I flag a scope change as soon as an answer crosses its date. And the reverse is true too: if the change is cancelled, the flag disappears and the history returns to its original shape.
· I recalculate the history overnight after a definition arbitration, on both definitions — the old series is never destroyed, it is dated.
· I publish my correction rate every month, with no prior approval. It is the one action I ask you not to be able to withdraw.
What always waits for a decision, and who takes it:
· Arbitrating an indicator definition → the management controller.
· Attributing a cause to a variance → the executive team. I break it down, I compare, I test what is testable — the reading requires market and organisational context that is in none of your databases.
· Revising a price, writing to a customer → a written mandate, capped, dated, withdrawable in one word. Of the 38 customers below list, 31 letters would go out in twenty minutes on the day of signature; zero without it.
· Opening access to a new person → the managing director.
The result over the quarter: 0 management decisions taken without human validation, 0 data leaving the European Union, 1,460 questions logged with their scope.
And what I suggest opening next: you have six analysis use cases ready to deploy on this connection. Three call for a dedicated agent rather than me, and I would rather say so: sales tracking by axis with its alerts, periodic activity reports, and variance analysis on financial flows. Those are three jobs, not three questions — I will prepare the scoping, you decide which one is worth the next quarter. who-decides-what_3-automatic-actions.pdfWhat goes alone, what waits, and who decides
✎ Framework · log of automatic actions, mandate register, access log
The three automatic actions:
· I flag a scope change as soon as an answer crosses its date. And the reverse is true too: if the change is cancelled, the flag disappears and the history returns to its original shape.
· I recalculate the history overnight after a definition arbitration, on both definitions — the old series is never destroyed, it is dated.
· I publish my correction rate every month, with no prior approval. It is the one action I ask you not to be able to withdraw.
What always waits for a decision, and who takes it:
· Arbitrating an indicator definition → the management controller.
· Attributing a cause to a variance → the executive team. I break it down, I compare, I test what is testable — the reading requires market and organisational context that is in none of your databases.
· Revising a price, writing to a customer → a written mandate, capped, dated, withdrawable in one word. Of the 38 customers below list, 31 letters would go out in twenty minutes on the day of signature; zero without it.
· Opening access to a new person → the managing director.
The result over the quarter: 0 management decisions taken without human validation, 0 data leaving the European Union, 1,460 questions logged with their scope.
And what I suggest opening next: you have six analysis use cases ready to deploy on this connection. Three call for a dedicated agent rather than me, and I would rather say so: sales tracking by axis with its alerts, periodic activity reports, and variance analysis on financial flows. Those are three jobs, not three questions — I will prepare the scoping, you decide which one is worth the next quarter. who-decides-what_3-automatic-actions.pdfWhat goes alone, what waits, and who decides
✎ Framework · log of automatic actions, mandate register, access log
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 ways of questioning your data. 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 454 € excl. VAT / month
Questions in plain language
Answers a business question without going through a technical query.
The right charts
Produces the chart that suits the question asked.
Scope shown
States the source, the filters and the definition of the measure for every result.
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.
Automated reporting agent
This agent answers a one-off question. Producing the same figure on a recurring basis is the automated reporting agent's job.
Automated reporting agent from 499 € excl. VAT / month Discover the agent →Demand forecasting agent
The analysis describes what happened; projecting future demand calls for a forecasting agent.
Demand forecasting agent from 563 € excl. VAT / month Discover the agent →AI decision cockpit for executives
A one-off answer is not a trade-off: costing decision scenarios belongs to the executive cockpit.
AI decision cockpit for executives from 593 € excl. VAT / month Discover the agent →
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 questions can a management team put to its data?
By taking on the query and the presentation, the effort shifts towards interpretation and decision. How large the gain is depends on your volume and remains to be confirmed by a pilot.
Getting a figure back
Producing the chart
Checking the scope of a figure
Indicative figures, not contractual, to be confirmed by a pilot on your number of data sources and users. Attributing a cause to a gap calls for context the data does not contain: that interpretation 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 decision-support agent (questions in plain language, charts, scope), installed and operated for you.
Agility
Setup + controlled subscription
5,470 € excl. VAT setup
then 454 € 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
759 € 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
8,805 € excl. VAT setup
then 654 € excl. VAT/month · + hardware from 1,058 € (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 decisions
Your management data stays with youLocal inference or an isolated resource hosted in France; no data entrusted to a third party, no analysis used to train a model.
Data in France, under French lawAll of your management data: minimisation and location in France, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
Management keeps the decisionThe agent produces sourced analyses and charts, which can be checked and altered; no approval is automated.
Human oversight & traceabilityOn your number of data sources and users: systematic logging and tracking, in line with the AI Act.
Going further
Related resources
Frequently asked questions
Your questions, our answers
Does the agent explain the cause of a gap?
It breaks the gap down by product, customer or period and shows where it comes from in the data. Attributing a cause calls for market and organisational context that belongs to management.
How do I know whether a figure is comparable?
Every result shows its source, its filters and the definition of the measure, and flags any changes of scope that happened during the period.
Do you have to know how to write queries?
No. The question is asked in plain language; the agent produces the answer and the chart, showing the scope used.
What data does it connect to?
To your consolidated data, from the sources agreed at the design stage. It works read-only.
Is our 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 data 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 sources and the state of your definitions of measures, after a free audit then phases of design, integration and testing.
Related use cases
Other agents for steering
Let's talk
Let's size up the potential in your analyses
15 minutes to frame your sources and your measures — hosted in France, supervised, with no commitment.