● B2B offer — Orchestration & automation
Orchestrated agent: several tools, one controlled chain
A great many business tasks come down to repeating the same chain across several tools: read here, create there, update elsewhere, then notify. Your agent runs that chain, stops at the approval points you have placed and keeps the log of every action performed. Hosted in France: the data your tools exchange stays with you. The actions that commit you wait for your decision.
Hosted in France
Exchanges between tools protected
GDPR & AI Act: governed deployment
Human oversight
Updated on
Deployed in a few weeks
Orchestrated agent · hosted in France
Handle the new requests under our usual procedure.
Chain run: request read, record created in the business tool, documents attached, manager notified.
Every action is logged with its tool, its timestamp and its result.
One action in the chain is marked as committing: it is waiting for your approval.
🔗 Sourced · log of actions, tool by tool
Every action is logged with its tool, its timestamp and its result.
One action in the chain is marked as committing: it is waiting for your approval.
🔗 Sourced · log of actions, tool by tool
Which one, and why?
Sending to the client: you defined it as an approval point at the design stage.
Everything before it is reversible and recorded; anything that leaves the company waits for a human act.
✎ Support · committing action pending
Everything before it is reversible and recorded; anything that leaves the company waits for a human act.
✎ Support · committing action pending
Local inference · no data outside the EU
Actions 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 business department decidesThe agent prepares, never rules
✦ In brief
A Blue Lemon Agent orchestrated agent runs your chains across several business tools — read, create, update, notify — stops at the approval points you have placed and keeps the log of every action with its tool and its timestamp. Anything that leaves the company always waits for a human act. It runs on local inference or is hosted in France: the data exchanged 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
orchestration 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 tools connected and chains is confirmed by a pilot.
The context
What does an orchestrated agent bring to your business tasks?
A chain repeated several times a day is worth running identically each time, and worth leaving a record of each step.
! The issue
A recurring business task combines several tools and a precise order. Done by hand it is consistent but costly; automated without traceability it becomes hard to pick up again. The agent runs it identically, logs every action with its tool and its result, and stops where you decided.
✓ Our answer
The business department gets a chain that is reliable and fully traceable, with approval points wherever the action commits the company. What comes before is reversible and documented; anything that leaves the company waits for a human act. Local inference or an isolated resource hosted in France: the data your tools exchange does not leave the company.
The decisive point
The data your tools exchange: sovereignty & compliance
An agent that acts inside your tools must be strictly bounded and fully traceable. Here is how it is.
Local inference
The agent can run on a machine belonging to your organisation: no data exchanged between your tools leaves the network.
Hosting in France
Otherwise, a dedicated and isolated resource hosted in France, under French law — your business tools and your task chains: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
For the data your tools exchange, 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 organisation and its business tools.
Every action logged
Tool, timestamp and result are kept for every action; encryption, role-based access and a scope of action limited to what you have authorised.
AI Act: governed deployment
The agent is strictly in support; no committing action is performed without human approval; 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 chains touching health, financial or confidential data, 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 companyVergnaud Technique — installation and servicing of professional refrigeration equipment
- Sector
- Commercial refrigeration and professional kitchens — installation, servicing under contract, emergency breakdown cover
- Headcount
- 52 staff, including 6 in sales administration — the team connected to the agent —, 18 field engineers and 4 in the design office
- Client base
- 340 clients under a service contract across four counties — institutional catering, central kitchens, food retail
- Volume
- 640 call-out requests, 410 parts orders and 96 warranty claims a month; €580 average invoice per visit
- Tools in place
- Email, ERP, document management, route planner, the portals of the three approved suppliers and a CRM — the agent plugs into them, nothing is replaced
- Who decides
- The sales administration manager places the validation points and signs the mandates; the technical director settles emergencies; the owner sets the spending caps
- Room for improvement
- Running the sequence takes 65 % of the 40 minutes a call-out request costs; 17 requests a month get no answer within 24 hours; a parts order waits 2.4 days before it is placed
Vergnaud Technique is not looking to work differently: it repeats the same sequences in five tools every day, and would like to repeat them identically without spending its days on them. The agent runs on local inference on a machine at the company and plugs into the mailbox, the ERP, document management, the route planner and the supplier portals: it executes, it logs, it stops where sales administration decided it should. The exchanges below span one quarter, from mapping the sequences 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.
Orchestrated agent · hosted in France
I have read three months of logs from your five tools — 5,292 chains of actions — and out of them come six sequences you replay identically, 1,764 times a month.
A sequence is a series of actions always in the same order and across several tools: read here, create there, update elsewhere, notify afterwards. Your teams know them by heart; they were written down nowhere, and that is exactly why they cost so much.
The gap I measured, and it is the gap that decides the gain: a call-out request costs 40 minutes to your sales administration team — the department that takes the requests in, opens the files, orders the parts and triggers invoicing. 26 of those 40 minutes — 65 % — are running the sequence itself: opening the message, creating the file in the ERP, attaching the documents, hunting for the engineer's slot, writing the acknowledgement. The remaining 14 minutes are your trade: judging the urgency, calling the client, deciding who goes out.
The six sequences, by monthly volume:
· Incoming call-out request — 640 a month, 40 min each.
· Parts order to a supplier — 410 a month, 20 min, of which 9 minutes of pure retyping from one tool into another.
· Manufacturer warranty claim — 96 a month, 24 min, of which 6 minutes building the trace.
· Call-out quote on the price list — 210 a month.
· Round completion and invoicing trigger — 380 a month.
· Service contract renewal — 28 a month.
What the first one gives you back once connected: running it drops from 26 minutes to 3 — 65 % of the cycle down to 8 %. The 14 minutes of judgement stay with you, and that is exactly right: they are what your clients pay for.
What I propose: start with the call-out request, the biggest volume and the easiest to bound. You read the map of the six sequences and cross out whatever must not be touched — it is your organisation, I have only written it down. six-sequences_costed.pdf1,764 runs a month, minute by minute
⛓ Sourced · logs of the five tools over three months, 5,292 chains of actions, ERP and mailbox timestamps
A sequence is a series of actions always in the same order and across several tools: read here, create there, update elsewhere, notify afterwards. Your teams know them by heart; they were written down nowhere, and that is exactly why they cost so much.
The gap I measured, and it is the gap that decides the gain: a call-out request costs 40 minutes to your sales administration team — the department that takes the requests in, opens the files, orders the parts and triggers invoicing. 26 of those 40 minutes — 65 % — are running the sequence itself: opening the message, creating the file in the ERP, attaching the documents, hunting for the engineer's slot, writing the acknowledgement. The remaining 14 minutes are your trade: judging the urgency, calling the client, deciding who goes out.
The six sequences, by monthly volume:
· Incoming call-out request — 640 a month, 40 min each.
· Parts order to a supplier — 410 a month, 20 min, of which 9 minutes of pure retyping from one tool into another.
· Manufacturer warranty claim — 96 a month, 24 min, of which 6 minutes building the trace.
· Call-out quote on the price list — 210 a month.
· Round completion and invoicing trigger — 380 a month.
· Service contract renewal — 28 a month.
What the first one gives you back once connected: running it drops from 26 minutes to 3 — 65 % of the cycle down to 8 %. The 14 minutes of judgement stay with you, and that is exactly right: they are what your clients pay for.
What I propose: start with the call-out request, the biggest volume and the easiest to bound. You read the map of the six sequences and cross out whatever must not be touched — it is your organisation, I have only written it down. six-sequences_costed.pdf1,764 runs a month, minute by minute
⛓ Sourced · logs of the five tools over three months, 5,292 chains of actions, ERP and mailbox timestamps
40 minutes for one request sounds like a lot. Where does that figure come from?
From your own timestamps, not from a market ratio — and I give you the minute-by-minute breakdown so that you can contradict it.
How I measure: I take the time the message lands in the team's mailbox and the time of the last action of the sequence in the ERP, stripping out nights, weekends and any gap longer than two hours — a file left aside does not count as work. What remains is the time actually spent.
The 40 minutes in detail, averaged over 1,920 requests:
· Reading the request and finding the client in the ERP — 6 min
· Creating the call-out file and qualifying it — 7 min
· Attaching the documents to the file in document management — 5 min
· Hunting for a slot in the route planner — 8 min
· Writing and sending the acknowledgement to the client — 4 min ← the committing action; it will stay with you
· Sub-total for running it: 30 min, of which 26 minutes reproducible identically — the 4 minutes of the acknowledgement leave the company, so they are not in the gain.
· Judging the urgency, calling the client, choosing the engineer — 10 min, your trade.
The figure that will speak to you more than the minutes: your six people in sales administration spend 427 hours a month on call-out requests alone. That is 47 % of their combined working time — and last year you absorbed 18 % more requests without hiring. The point is not to do less, it is to take the next step up.
What I propose to settle it: I replay the sequence dry on 40 real requests from last week, writing nothing into your tools, and we compare my record with your timestamps. If my figure is wrong, it is better to know before anything is connected. call-out-request_minute-by-minute.pdf40 min broken down, measurement method included
⛓ Sourced · 1,920 timestamped requests over three months, mailbox and ERP cross-checked
How I measure: I take the time the message lands in the team's mailbox and the time of the last action of the sequence in the ERP, stripping out nights, weekends and any gap longer than two hours — a file left aside does not count as work. What remains is the time actually spent.
The 40 minutes in detail, averaged over 1,920 requests:
· Reading the request and finding the client in the ERP — 6 min
· Creating the call-out file and qualifying it — 7 min
· Attaching the documents to the file in document management — 5 min
· Hunting for a slot in the route planner — 8 min
· Writing and sending the acknowledgement to the client — 4 min ← the committing action; it will stay with you
· Sub-total for running it: 30 min, of which 26 minutes reproducible identically — the 4 minutes of the acknowledgement leave the company, so they are not in the gain.
· Judging the urgency, calling the client, choosing the engineer — 10 min, your trade.
The figure that will speak to you more than the minutes: your six people in sales administration spend 427 hours a month on call-out requests alone. That is 47 % of their combined working time — and last year you absorbed 18 % more requests without hiring. The point is not to do less, it is to take the next step up.
What I propose to settle it: I replay the sequence dry on 40 real requests from last week, writing nothing into your tools, and we compare my record with your timestamps. If my figure is wrong, it is better to know before anything is connected. call-out-request_minute-by-minute.pdf40 min broken down, measurement method included
⛓ Sourced · 1,920 timestamped requests over three months, mailbox and ERP cross-checked
And the 17 requests a month that get no answer within 24 hours — do they come from that too?
They do, and it is the most expensive figure on your list — the only one that shows up in your revenue.
What your CRM says over twelve months: 17 requests a month cross the 24-hour line with no first answer. They are not spread at random: 11 out of 17 arrive between 4 pm and 7 pm, when sales administration is processing the completed rounds, and 4 out of 17 arrive on Friday afternoon. This is not a diligence problem, it is a queue problem.
What it costs, using your own figures: of the 204 requests that went past 24 hours last year, 38 never turned into a visit — the client had called someone else. At €580 average invoice, that is €22,040 of business that never happened.
What the connected sequence changes: the file is created and the acknowledgement is drafted within three minutes, at any hour, including Friday at 6 pm. The acknowledgement is not sent without you — it leaves the company, so it waits for a human gesture or a mandate, and we will come to that. But it is written, complete, and it goes out on the first glance.
The benchmark I propose to track: requests with no answer within 24 hours, recorded every Monday. 17 today. I commit to a target of 6 after the first month, and if we are not there, the record will tell you before I do.
The next step I propose: let us connect the call-out request alone, for three weeks, with the log open. The other five sequences can wait until that one has convinced you.
⛓ Sourced · CRM over twelve months, 204 requests past 24 hours, outcomes in the ERP
What your CRM says over twelve months: 17 requests a month cross the 24-hour line with no first answer. They are not spread at random: 11 out of 17 arrive between 4 pm and 7 pm, when sales administration is processing the completed rounds, and 4 out of 17 arrive on Friday afternoon. This is not a diligence problem, it is a queue problem.
What it costs, using your own figures: of the 204 requests that went past 24 hours last year, 38 never turned into a visit — the client had called someone else. At €580 average invoice, that is €22,040 of business that never happened.
What the connected sequence changes: the file is created and the acknowledgement is drafted within three minutes, at any hour, including Friday at 6 pm. The acknowledgement is not sent without you — it leaves the company, so it waits for a human gesture or a mandate, and we will come to that. But it is written, complete, and it goes out on the first glance.
The benchmark I propose to track: requests with no answer within 24 hours, recorded every Monday. 17 today. I commit to a target of 6 after the first month, and if we are not there, the record will tell you before I do.
The next step I propose: let us connect the call-out request alone, for three weeks, with the log open. The other five sequences can wait until that one has convinced you.
⛓ Sourced · CRM over twelve months, 204 requests past 24 hours, outcomes in the ERP
Concretely, this morning, what did you do in our tools?
31 requests handled end to end between 6 am and 9 am, 5 tools touched, 214 actions logged, and two files waiting for you.
For each of the 31, in this order:
· Mailbox — I read the request and identify the client by contract number or, failing that, by the sending address and the name of the site.
· ERP — I create the call-out file, attach it to the contract, and qualify the equipment from the asset list already recorded. I never create a second client record: if the match is doubtful, the file goes on hold and you see it.
· Document management — I attach the documents received to the file, renamed according to your naming plan: contract number, date, nature.
· Route planner — I propose the slot of the engineer whose round passes closest, and I hold the option without confirming it.
· Mailbox — I write the acknowledgement to the client with the file number and the proposed slot. It waits: this is a validation point.
The time: 3 minutes per request against 26 by hand — 8 % of the cycle instead of 65 %. On this one morning, 11 hours and 53 minutes that will not be taken out of your six people's day.
The two files waiting for you, and why:
· Berthaud central kitchen — the request concerns a piece of equipment missing from the recorded asset list. Either it was never entered, or it is not under contract: either way that is a commercial decision, not an execution step. I have drafted both versions of the acknowledgement.
· Marcelin Group — the message covers three sites at once. I have prepared three separate files; I will not open them without you, because three files mean three chargeable visits.
What I propose: look at those two and nothing else. The other 29 do not need you, and that is exactly what you asked me for. connection_what-the-agent-touches.pdf5 tools, rights granted, rights refused
⛓ Sourced · action log of 12/09, 214 timestamped actions across 5 tools
For each of the 31, in this order:
· Mailbox — I read the request and identify the client by contract number or, failing that, by the sending address and the name of the site.
· ERP — I create the call-out file, attach it to the contract, and qualify the equipment from the asset list already recorded. I never create a second client record: if the match is doubtful, the file goes on hold and you see it.
· Document management — I attach the documents received to the file, renamed according to your naming plan: contract number, date, nature.
· Route planner — I propose the slot of the engineer whose round passes closest, and I hold the option without confirming it.
· Mailbox — I write the acknowledgement to the client with the file number and the proposed slot. It waits: this is a validation point.
The time: 3 minutes per request against 26 by hand — 8 % of the cycle instead of 65 %. On this one morning, 11 hours and 53 minutes that will not be taken out of your six people's day.
The two files waiting for you, and why:
· Berthaud central kitchen — the request concerns a piece of equipment missing from the recorded asset list. Either it was never entered, or it is not under contract: either way that is a commercial decision, not an execution step. I have drafted both versions of the acknowledgement.
· Marcelin Group — the message covers three sites at once. I have prepared three separate files; I will not open them without you, because three files mean three chargeable visits.
What I propose: look at those two and nothing else. The other 29 do not need you, and that is exactly what you asked me for. connection_what-the-agent-touches.pdf5 tools, rights granted, rights refused
⛓ Sourced · action log of 12/09, 214 timestamped actions across 5 tools
You write into the ERP. What stops me ending up with 400 duplicate files or an unfortunate deletion?
Two things, and they are in the architecture, not in my good intentions: a narrower access scope than yours, and a key that makes every action replayable without ever duplicating it.
The scope, tool by tool — this is role-based access: rights follow the job, and mine is narrower than that of any of your users.
· ERP: create a call-out file, update it, read the asset list and the contracts. No deletion. No price changes. No access to the payroll module, even though it lives in the same software.
· Document management: file and classify. No deletion, no overwriting — a document filed under the same name creates a version, it erases nothing.
· Route planner: propose and hold an option. No confirmation, no moving an existing appointment.
· Mailbox: read the team inbox, write drafts. Sending anything outside is a validation point or a mandate, never a standing right.
· Supplier portals: read the catalogue and stock levels. Ordering is under mandate, and the mandate does not exist yet.
Now the duplicates, and that is the real technical question: every run carries a uniqueness key — an identifier built from the original request: the sender, the timestamp of the message and its subject. Before writing, I check that no file already carries that key. If the sequence is restarted, it picks up where it stopped and re-creates nothing.
The figure that proves it, over the quarter: 168 runs were interrupted — a portal outage, an unreadable message, a tool under maintenance. 168 restarts. Zero duplicate files in the ERP, zero documents filed twice. You can check it yourself: the reconciliation report is attached.
What I propose: keep me on that scope for the whole pilot. If a right is missing, it can be requested, logged and withdrawn — that is healthier than a broad access granted once and forgotten. restarts_168-stops-zero-duplicates.pdfWhere it stops, where it resumes, what is never replayed
✎ Framework · rights matrix by tool, reconciliation report on the 168 restarts
The scope, tool by tool — this is role-based access: rights follow the job, and mine is narrower than that of any of your users.
· ERP: create a call-out file, update it, read the asset list and the contracts. No deletion. No price changes. No access to the payroll module, even though it lives in the same software.
· Document management: file and classify. No deletion, no overwriting — a document filed under the same name creates a version, it erases nothing.
· Route planner: propose and hold an option. No confirmation, no moving an existing appointment.
· Mailbox: read the team inbox, write drafts. Sending anything outside is a validation point or a mandate, never a standing right.
· Supplier portals: read the catalogue and stock levels. Ordering is under mandate, and the mandate does not exist yet.
Now the duplicates, and that is the real technical question: every run carries a uniqueness key — an identifier built from the original request: the sender, the timestamp of the message and its subject. Before writing, I check that no file already carries that key. If the sequence is restarted, it picks up where it stopped and re-creates nothing.
The figure that proves it, over the quarter: 168 runs were interrupted — a portal outage, an unreadable message, a tool under maintenance. 168 restarts. Zero duplicate files in the ERP, zero documents filed twice. You can check it yourself: the reconciliation report is attached.
What I propose: keep me on that scope for the whole pilot. If a right is missing, it can be requested, logged and withdrawn — that is healthier than a broad access granted once and forgotten. restarts_168-stops-zero-duplicates.pdfWhere it stops, where it resumes, what is never replayed
✎ Framework · rights matrix by tool, reconciliation report on the 168 restarts
Parts orders are where we lose the most time retyping. Can you do something without committing money yet?
Yes: all the work except the final click. The month's 410 orders are prepared, checked and filed as drafts on the portals; all that is left is sending them, which commits your money.
What retyping costs you today: an order takes 20 minutes, 9 of them pure copying — the part number read in the engineer's report, retyped into the portal, then copied into the ERP, then carried over into the call-out file. 45 % of the time of an order spent recopying what is already written elsewhere.
What the sequence does:
· it reads the part number in the engineer's report, including when it is photographed on the label of the part;
· it finds it in the supplier's contract catalogue and checks the price against your negotiated list;
· it looks at your own stock before ordering — of the month's 410, 37 parts were on the shelf, that is €4,180 of orders that had no reason to be placed;
· it builds the purchase order, attaches it to the call-out file and updates the ERP;
· it files the order as a draft on the portal and stops there.
The time: 9 minutes of data entry fall to 1 minute 15 — 45 % of the cycle down to 6 %. Across 410 orders, 53 hours a month.
The figure that is not a time saving, and weighs more: your orders currently wait 2.4 days between the engineer finding the fault and someone approving the order. 63 visits a month are held up waiting for a part, and 9 clients have cited that delay to renegotiate their service contract over the last twelve months, and 3 left — three contracts at €4,800 a year, so €14,400 of recurring revenue that will not come back.
What I propose, and it is the real decision of this tab: those drafts can go out on their own, within limits that you write. Not a request for trust on my part: a mandate, bounded and withdrawable. I will put it to you in the next tab, costed and ready to sign. parts-orders_410-prepared.pdf37 parts already in stock, €4,180 avoided
⛓ Sourced · the month's 410 orders, round reports, contract catalogue of the 3 suppliers, ERP stock levels
What retyping costs you today: an order takes 20 minutes, 9 of them pure copying — the part number read in the engineer's report, retyped into the portal, then copied into the ERP, then carried over into the call-out file. 45 % of the time of an order spent recopying what is already written elsewhere.
What the sequence does:
· it reads the part number in the engineer's report, including when it is photographed on the label of the part;
· it finds it in the supplier's contract catalogue and checks the price against your negotiated list;
· it looks at your own stock before ordering — of the month's 410, 37 parts were on the shelf, that is €4,180 of orders that had no reason to be placed;
· it builds the purchase order, attaches it to the call-out file and updates the ERP;
· it files the order as a draft on the portal and stops there.
The time: 9 minutes of data entry fall to 1 minute 15 — 45 % of the cycle down to 6 %. Across 410 orders, 53 hours a month.
The figure that is not a time saving, and weighs more: your orders currently wait 2.4 days between the engineer finding the fault and someone approving the order. 63 visits a month are held up waiting for a part, and 9 clients have cited that delay to renegotiate their service contract over the last twelve months, and 3 left — three contracts at €4,800 a year, so €14,400 of recurring revenue that will not come back.
What I propose, and it is the real decision of this tab: those drafts can go out on their own, within limits that you write. Not a request for trust on my part: a mandate, bounded and withdrawable. I will put it to you in the next tab, costed and ready to sign. parts-orders_410-prepared.pdf37 parts already in stock, €4,180 avoided
⛓ Sourced · the month's 410 orders, round reports, contract catalogue of the 3 suppliers, ERP stock levels
Placing an order, sending a quote: those commit the company. I assume that is where you stop?
No. I stop as long as nothing is written down — which is not the same thing. An action that commits your money does not need a refusal, it needs a mandate.
A mandate, here, is a written authority by which you allow me to act in your name within limits you set: what I may do, up to how much, until when, and how you take it back. It is exactly what you already sign for your buyers — I am asking to be treated like them, no better and no worse.
What the mandate does not remove: the human gesture. It moves it upstream and puts it in writing — one signature covering a named scope, instead of twenty-eight approvals to give one by one every morning. Anything outside those bounds still waits for a person, and you see it arrive in your list.
What I have prepared, ready to sign, in four lines:
· Scope — order parts from the three approved suppliers, from the contract catalogue only, at the price on your negotiated list, and only for a call-out file that is already open. Outside those four bounds, nothing goes out and you see it.
· Cap — €400 per order, €6,000 per week, 30 orders per day. Those three figures cover 87 % of your orders over the last twelve months; the remaining 13 % keep waiting for your approval, and that is what you want: they are the compressors and the complete units.
· Date — mandate from 1 September to 31 December, with a review on 30 September, figures in hand. A mandate with no end date is not a mandate, it is an abdication.
· Withdrawal — one message and it falls, within the minute. Orders already placed stay visible, each with its time, amount, file and the line of the mandate that authorised it. You have nothing to justify in order to withdraw it.
What stays outside, and I write it down in black and white: any supplier not approved, any part outside the catalogue, any price above the list, any order without an open file, and any negotiation — a price is discussed between two people.
What I propose: read it, cross out what bothers you, lower the caps if you would rather start cautiously. A €200 mandate suits me perfectly well: it will cover 61 % of your orders instead of 87 %, and we will raise it on 30 September if the record is good. purchasing-mandate_scope-cap-review-withdrawal.pdf4 bounds, 3 caps, one date, one word to withdraw it
✎ Framework · draft purchasing mandate, 12 months of orders analysed by amount and by supplier
A mandate, here, is a written authority by which you allow me to act in your name within limits you set: what I may do, up to how much, until when, and how you take it back. It is exactly what you already sign for your buyers — I am asking to be treated like them, no better and no worse.
What the mandate does not remove: the human gesture. It moves it upstream and puts it in writing — one signature covering a named scope, instead of twenty-eight approvals to give one by one every morning. Anything outside those bounds still waits for a person, and you see it arrive in your list.
What I have prepared, ready to sign, in four lines:
· Scope — order parts from the three approved suppliers, from the contract catalogue only, at the price on your negotiated list, and only for a call-out file that is already open. Outside those four bounds, nothing goes out and you see it.
· Cap — €400 per order, €6,000 per week, 30 orders per day. Those three figures cover 87 % of your orders over the last twelve months; the remaining 13 % keep waiting for your approval, and that is what you want: they are the compressors and the complete units.
· Date — mandate from 1 September to 31 December, with a review on 30 September, figures in hand. A mandate with no end date is not a mandate, it is an abdication.
· Withdrawal — one message and it falls, within the minute. Orders already placed stay visible, each with its time, amount, file and the line of the mandate that authorised it. You have nothing to justify in order to withdraw it.
What stays outside, and I write it down in black and white: any supplier not approved, any part outside the catalogue, any price above the list, any order without an open file, and any negotiation — a price is discussed between two people.
What I propose: read it, cross out what bothers you, lower the caps if you would rather start cautiously. A €200 mandate suits me perfectly well: it will cover 61 % of your orders instead of 87 %, and we will raise it on 30 September if the record is good. purchasing-mandate_scope-cap-review-withdrawal.pdf4 bounds, 3 caps, one date, one word to withdraw it
✎ Framework · draft purchasing mandate, 12 months of orders analysed by amount and by supplier
We will sign it as it stands, with the €400 cap. From when does it run?
It already does. Mandate recorded at 9:12; by 9:26 the day's 31 orders had been placed. Fourteen minutes.
What those 31 orders cover: 28 under €400, placed under mandate; 3 above — €612, €890 and €1,340 — prepared and left for your approval, exactly as planned. None was split to slip under the cap: splitting is the first way a mandate can be turned around, I have blocked it in my own rules and the log shows it.
What changed in the field, in one morning:
· Time to place an order: 2.4 days → 14 minutes.
· 63 visits a month were waiting for a part; the 2.4 days they used to lose at the start are given back to the client, without a single engineer changing anything about the way they work.
· 4 parts arrived the very next day because the order went out before the supplier's 11 am cut-off. Nobody knew about that cut-off: it is in the portal's terms, and I read them.
What the log records for each of the 28: the time, the supplier, the part number, the amount, the call-out file, and the line of the mandate that authorises it. An order with no mandate line cannot exist — that is not a rule of prudence, it is the condition for writing.
The first measure I will give you back: on 30 September, four figures — number of orders under mandate, total amount, orders escalated for your approval, and discrepancies found. If there is a discrepancy, it will be in that table before it is in a supplier complaint.
The next step I propose: the same arrangement on quotes, where the gain is no longer time but revenue. Let me show you the calculation. mandate-at-work_31-orders-14-minutes.pdf28 under mandate, 3 escalated, 0 splitting
⛓ Sourced · action log of 01/09, 31 orders, mandate no. 1 recorded at 9:12
What those 31 orders cover: 28 under €400, placed under mandate; 3 above — €612, €890 and €1,340 — prepared and left for your approval, exactly as planned. None was split to slip under the cap: splitting is the first way a mandate can be turned around, I have blocked it in my own rules and the log shows it.
What changed in the field, in one morning:
· Time to place an order: 2.4 days → 14 minutes.
· 63 visits a month were waiting for a part; the 2.4 days they used to lose at the start are given back to the client, without a single engineer changing anything about the way they work.
· 4 parts arrived the very next day because the order went out before the supplier's 11 am cut-off. Nobody knew about that cut-off: it is in the portal's terms, and I read them.
What the log records for each of the 28: the time, the supplier, the part number, the amount, the call-out file, and the line of the mandate that authorises it. An order with no mandate line cannot exist — that is not a rule of prudence, it is the condition for writing.
The first measure I will give you back: on 30 September, four figures — number of orders under mandate, total amount, orders escalated for your approval, and discrepancies found. If there is a discrepancy, it will be in that table before it is in a supplier complaint.
The next step I propose: the same arrangement on quotes, where the gain is no longer time but revenue. Let me show you the calculation. mandate-at-work_31-orders-14-minutes.pdf28 under mandate, 3 escalated, 0 splitting
⛓ Sourced · action log of 01/09, 31 orders, mandate no. 1 recorded at 9:12
Quotes are more sensitive: that is a price leaving the house. What do you propose?
A second mandate, narrower than the first, and the figure that justifies it comes from your last 24 months — not from a market average.
What your history says, and it is the reason this mandate exists: across 4,830 call-out quotes, those sent within the hour are accepted 61 % of the time; those sent the next day or later, 46 %. Fifteen points apart, on your own client base.
The honest reading of that gap, because it does not read itself: part of it is speed, part of it is that urgent jobs naturally go out faster and get turned down less. So I stripped out the emergencies: on non-urgent quotes alone, the gap falls to 9 points — 58 % against 49 %. That is the figure I keep, not the first one.
What it is worth, calculated on 9 points: 210 quotes a month, 19 more accepted, at €580 average invoice — around €11,000 of extra monthly revenue, at unchanged quote volume and without a single extra discount.
The mandate I propose, and it is deliberately tight:
· Scope — quotes drawn at the exact price of your current list, discount at zero, for services in the catalogue, for clients under contract, with no open dispute and no overrun of their outstanding balance — the amount the client already owes you and has not yet paid.
· Cap — €900 per quote, 25 quotes a day. Above that, your approval.
· Date — two months, review at six weeks on the measured acceptance rate.
· Withdrawal — one word, immediate effect, and quotes already sent stay viewable one by one.
What I will never do under this mandate, even if you ask me to: grant a discount, change a list price or promise a call-out deadline. A price and a deadline are commercial commitments: they are decided, not executed. You want pricing flexibility? Add a line to the list — I will apply it the same day.
The decision stays with the owner: you set the cap and the duration, I give you the record every week. quote-mandate_9-points-and-its-limits.pdfThe calculation, the adjusted gap, the 3 things the mandate forbids
⛓ Sourced · 4,830 quotes over 24 months, acceptance rate by sending delay, emergencies isolated
What your history says, and it is the reason this mandate exists: across 4,830 call-out quotes, those sent within the hour are accepted 61 % of the time; those sent the next day or later, 46 %. Fifteen points apart, on your own client base.
The honest reading of that gap, because it does not read itself: part of it is speed, part of it is that urgent jobs naturally go out faster and get turned down less. So I stripped out the emergencies: on non-urgent quotes alone, the gap falls to 9 points — 58 % against 49 %. That is the figure I keep, not the first one.
What it is worth, calculated on 9 points: 210 quotes a month, 19 more accepted, at €580 average invoice — around €11,000 of extra monthly revenue, at unchanged quote volume and without a single extra discount.
The mandate I propose, and it is deliberately tight:
· Scope — quotes drawn at the exact price of your current list, discount at zero, for services in the catalogue, for clients under contract, with no open dispute and no overrun of their outstanding balance — the amount the client already owes you and has not yet paid.
· Cap — €900 per quote, 25 quotes a day. Above that, your approval.
· Date — two months, review at six weeks on the measured acceptance rate.
· Withdrawal — one word, immediate effect, and quotes already sent stay viewable one by one.
What I will never do under this mandate, even if you ask me to: grant a discount, change a list price or promise a call-out deadline. A price and a deadline are commercial commitments: they are decided, not executed. You want pricing flexibility? Add a line to the list — I will apply it the same day.
The decision stays with the owner: you set the cap and the duration, I give you the record every week. quote-mandate_9-points-and-its-limits.pdfThe calculation, the adjusted gap, the 3 things the mandate forbids
⛓ Sourced · 4,830 quotes over 24 months, acceptance rate by sending delay, emergencies isolated
Our supplier is refusing a warranty claim: it says the part was declared out of time. We have nothing to prove otherwise.
You do have everything — and it was in the log before the dispute even existed. File 24-1187, refrigeration unit at the Berthaud central kitchen.
What the log shows, to the second:
· 03/07, 2:22 pm — engineer's report filed in document management, photograph of the rating plate included.
· 03/07, 2:24 pm — warranty claim filed on the supplier's portal, acknowledgement number GR-88214, copy of the receipt archived.
· 03/07, 2:25 pm — ERP file updated, status “warranty declared”.
· Supplier's contractual deadline: 15 days after the fault is found. Fault found on 02/07. Claim filed on day 1.
What I propose to do with that trace: I have drafted the challenge, with the timestamped portal receipt attached, and quoted the deadline clause of your framework contract — yours, not some general text I would be quoting from memory. €2,340 of parts and labour. It goes out as soon as you have read it.
And the subject goes beyond this file: across the 96 warranty claims a month, building the trace took 6 minutes each — 25 % of the time of a claim. It now takes 45 seconds, 3 %, because it is no longer a task: it is a by-product of running the sequence. We do not log afterwards, we log while acting.
The commercial figure of this tab: over the last twelve months, 7 warranty claims were refused for want of proof of filing on time — €14,900 that you carried yourselves. With the log, that ground for refusal disappears.
What I propose next: a monthly record of warranty claims filed, with the supplier's acknowledgement beside each. A warranty claim with no archived acknowledgement is a warranty claim that does not exist. action-log_file-24-1187.pdfEvery action, its tool, its time, its authority
⛓ Sourced · action log of file 24-1187, portal receipt GR-88214, supplier framework contract
What the log shows, to the second:
· 03/07, 2:22 pm — engineer's report filed in document management, photograph of the rating plate included.
· 03/07, 2:24 pm — warranty claim filed on the supplier's portal, acknowledgement number GR-88214, copy of the receipt archived.
· 03/07, 2:25 pm — ERP file updated, status “warranty declared”.
· Supplier's contractual deadline: 15 days after the fault is found. Fault found on 02/07. Claim filed on day 1.
What I propose to do with that trace: I have drafted the challenge, with the timestamped portal receipt attached, and quoted the deadline clause of your framework contract — yours, not some general text I would be quoting from memory. €2,340 of parts and labour. It goes out as soon as you have read it.
And the subject goes beyond this file: across the 96 warranty claims a month, building the trace took 6 minutes each — 25 % of the time of a claim. It now takes 45 seconds, 3 %, because it is no longer a task: it is a by-product of running the sequence. We do not log afterwards, we log while acting.
The commercial figure of this tab: over the last twelve months, 7 warranty claims were refused for want of proof of filing on time — €14,900 that you carried yourselves. With the log, that ground for refusal disappears.
What I propose next: a monthly record of warranty claims filed, with the supplier's acknowledgement beside each. A warranty claim with no archived acknowledgement is a warranty claim that does not exist. action-log_file-24-1187.pdfEvery action, its tool, its time, its authority
⛓ Sourced · action log of file 24-1187, portal receipt GR-88214, supplier framework contract
Who decides what, exactly, now that there are mandates? I need it to be clear for the team.
Three regimes, nine named validation points, and one page your team can pin to the wall.
Regime 1 — what runs on its own (internal, reversible, no commitment): creating and updating a call-out file, filing a document, proposing a slot, updating a status, preparing a draft. None of that leaves the company, and all of it can be undone.
Regime 2 — what runs under mandate (committing, but bounded in writing): ordering a part under €400 from an approved supplier; sending a quote under €900 at list price. Two mandates, two caps, two review dates, withdrawable on a word.
Regime 3 — the nine validation points, which wait for a person: the acknowledgement to a client on equipment outside the asset list · a quote above €900 or with a discount · an order above €400 · confirming a client appointment · triggering an invoice · closing a warranty claim · opening a new client · any message to a client in dispute · any action on a file flagged by the sales administration manager.
The three things I do without asking, and they protect the relationship rather than the figure:
· Blocking an order for a part already in stock — 37 times last month, €4,180.
· Holding back anything addressed to a client in an open dispute, whatever the mandate says. And the reverse is true: once the dispute is closed, sending resumes on its own, without you having to think about it.
· Stopping the sequence when two tools contradict each other — a contract terminated in the CRM but active in the ERP, for instance. I would rather stop and show you both screens.
What I propose: that the sales administration manager reads those nine points and removes or adds some. One validation point too many costs a minute; one validation point missing costs a client. who-decides-what_9-points-and-2-mandates.pdfThree regimes, one page to pin up
✎ Framework · register of validation points, mandates no. 1 and no. 2, blocking rules
Regime 1 — what runs on its own (internal, reversible, no commitment): creating and updating a call-out file, filing a document, proposing a slot, updating a status, preparing a draft. None of that leaves the company, and all of it can be undone.
Regime 2 — what runs under mandate (committing, but bounded in writing): ordering a part under €400 from an approved supplier; sending a quote under €900 at list price. Two mandates, two caps, two review dates, withdrawable on a word.
Regime 3 — the nine validation points, which wait for a person: the acknowledgement to a client on equipment outside the asset list · a quote above €900 or with a discount · an order above €400 · confirming a client appointment · triggering an invoice · closing a warranty claim · opening a new client · any message to a client in dispute · any action on a file flagged by the sales administration manager.
The three things I do without asking, and they protect the relationship rather than the figure:
· Blocking an order for a part already in stock — 37 times last month, €4,180.
· Holding back anything addressed to a client in an open dispute, whatever the mandate says. And the reverse is true: once the dispute is closed, sending resumes on its own, without you having to think about it.
· Stopping the sequence when two tools contradict each other — a contract terminated in the CRM but active in the ERP, for instance. I would rather stop and show you both screens.
What I propose: that the sales administration manager reads those nine points and removes or adds some. One validation point too many costs a minute; one validation point missing costs a client. who-decides-what_9-points-and-2-mandates.pdfThree regimes, one page to pin up
✎ Framework · register of validation points, mandates no. 1 and no. 2, blocking rules
And the day a tool goes down in the middle of the night? What do we wake up to?
To a three-line record and nothing to catch up on. It happened four times this quarter, and you only found out by reading the report.
What happens exactly: the sequence stops at the current step, never in the middle of a write; the reason is logged; the work already done stays done. The restart point is the last confirmed step: the run picks up from there, not from the beginning.
The quarter's 168 stops, out of 5,292 runs — 3.2 %:
· 112 came from me, and I publish those first: three of your clients write the part number on a scanned form attached to the message, and I was only reading the body of the message. I fixed that on 22 July; since then the stop rate has gone from 3.2 % to 0.8 % over four weeks.
· 41 came from one supplier's portal being unavailable between 2 am and 4 am — its maintenance window, which I have since written into my execution calendar: the night's orders now go out at 4:10 am.
· 15 came from genuinely unusable messages — a blurred photograph, an empty message. Those go up to the team the same morning, with the original request in front of them, and that is the right outcome: nobody can guess what is not written.
What the 168 did not produce: zero duplicate files, zero orders placed twice, zero documents filed twice — the uniqueness key of each run forbids it.
What I propose: an alert to your manager as soon as there are three stops on the same tool in the same day. One stop is an incident, three stops are a signal, and a signal has to leave the log and reach a person.
⛓ Sourced · the quarter's 168 stops, causes matched one by one, non-duplication report
What happens exactly: the sequence stops at the current step, never in the middle of a write; the reason is logged; the work already done stays done. The restart point is the last confirmed step: the run picks up from there, not from the beginning.
The quarter's 168 stops, out of 5,292 runs — 3.2 %:
· 112 came from me, and I publish those first: three of your clients write the part number on a scanned form attached to the message, and I was only reading the body of the message. I fixed that on 22 July; since then the stop rate has gone from 3.2 % to 0.8 % over four weeks.
· 41 came from one supplier's portal being unavailable between 2 am and 4 am — its maintenance window, which I have since written into my execution calendar: the night's orders now go out at 4:10 am.
· 15 came from genuinely unusable messages — a blurred photograph, an empty message. Those go up to the team the same morning, with the original request in front of them, and that is the right outcome: nobody can guess what is not written.
What the 168 did not produce: zero duplicate files, zero orders placed twice, zero documents filed twice — the uniqueness key of each run forbids it.
What I propose: an alert to your manager as soon as there are three stops on the same tool in the same day. One stop is an incident, three stops are a signal, and a signal has to leave the log and reach a person.
⛓ Sourced · the quarter's 168 stops, causes matched one by one, non-duplication report
After a quarter, what has it actually given us back?
245 hours a month in sales administration, and above all two figures that show up somewhere other than a timesheet.
On the three areas the management was tracking:
· Running the sequence: 65 % → 8 % of the time of a call-out request. The biggest gain, and the most mechanical.
· Data entry across the tools: 45 % → 6 % of the time of a parts order.
· Building the trace: 25 % → 3 % of the time of a warranty claim.
What those hours became: your six people did not work less, they absorbed the 18 % rise in requests without hiring, and the 24-hour client callback went from 17 requests left waiting a month to 2. At €580 average invoice, the 15 requests recovered each month are around €8,700 of revenue that happens instead of going elsewhere.
And quotes, which are the most profitable area: sending time down from 1.8 days to 11 minutes, acceptance rate on non-urgent quotes from 49 % to 57 % over the quarter. Eight points out of the nine announced — I promised you nine, I am delivering eight, and I would rather put it to you that way.
The figure that does not flatter me, and what follows it: 168 runs stopped out of 5,292 — 3.2 %. 112 came from me, for a silly reason: three clients write their part numbers on a scanned form that I was not reading. Fixed on 22 July, the rate is 0.8 % over the last four weeks. No stop produced a duplicate, and none was discovered by a client: they all came out of the morning record.
The framework measure, finally: 0 committing actions run outside a mandate or outside a validation, across 5,292 runs and 21,400 logged actions. That is not a cautious setting, it is how I am wired — and the log proves it line by line. quarterly-review_what-sales-admin-regained.pdf245 h a month, 3.2 % of stops and their fix
⛓ Sourced · timesheets, action log for the quarter, CRM, quote record
On the three areas the management was tracking:
· Running the sequence: 65 % → 8 % of the time of a call-out request. The biggest gain, and the most mechanical.
· Data entry across the tools: 45 % → 6 % of the time of a parts order.
· Building the trace: 25 % → 3 % of the time of a warranty claim.
What those hours became: your six people did not work less, they absorbed the 18 % rise in requests without hiring, and the 24-hour client callback went from 17 requests left waiting a month to 2. At €580 average invoice, the 15 requests recovered each month are around €8,700 of revenue that happens instead of going elsewhere.
And quotes, which are the most profitable area: sending time down from 1.8 days to 11 minutes, acceptance rate on non-urgent quotes from 49 % to 57 % over the quarter. Eight points out of the nine announced — I promised you nine, I am delivering eight, and I would rather put it to you that way.
The figure that does not flatter me, and what follows it: 168 runs stopped out of 5,292 — 3.2 %. 112 came from me, for a silly reason: three clients write their part numbers on a scanned form that I was not reading. Fixed on 22 July, the rate is 0.8 % over the last four weeks. No stop produced a duplicate, and none was discovered by a client: they all came out of the morning record.
The framework measure, finally: 0 committing actions run outside a mandate or outside a validation, across 5,292 runs and 21,400 logged actions. That is not a cautious setting, it is how I am wired — and the log proves it line by line. quarterly-review_what-sales-admin-regained.pdf245 h a month, 3.2 % of stops and their fix
⛓ Sourced · timesheets, action log for the quarter, CRM, quote record
We would like to go further: the full client dispute, from after-sales through to accounts. Can you do that?
Part of it, yes, and I will tell you frankly which part — because the right answer here is not me, it is the offer next door, and it costs you less than making me do what I do badly.
What I do very well: a sequence that is repetitive, inside one team, with written rules. Your six sequences are the example: they all look alike, every day.
What a full client dispute is: a file that crosses sales administration, the technical department, accounts and management, with regulatory deadlines, credit notes, a commercial decision in the middle and a different path depending on the amount. That is no longer a sequence, it is a process — a chain of steps crossing several departments, where each one waits on a decision from the one before.
The path I propose, and it starts with what I can do tomorrow:
· I take the three parts of the dispute that are already repetitive: building the claim file, gathering the evidence from the trace, preparing the credit note as a draft. Across the quarter's 41 disputes, that is 62 % of the time spent.
· The rest — the arbitration, the signed credit note, the legal follow-up — belongs to process automation, which knows how to hold a file across several departments and several weeks. That is an offer distinct from mine, and I point you to it rather than pretending.
· And if you connect a second agent — a telephone reception agent, say — do not run them side by side: there is an agent platform that provides the shared foundation, the access rights and a single log. Two agents without a foundation means two logs to reconcile by hand.
What I propose concretely: I build the three dispute sequences within a fortnight, and the 62 % table will serve as your file for deciding whether the rest is worth a process offer. The decision will be taken on a figure, not on a promise. scope_what-the-agent-does-and-what-extends-it.pdf62 % of the dispute tomorrow, the rest via the dedicated offer
⛓ Sourced · the quarter's 41 disputes, time per step, scope of the orchestration and process offers
What I do very well: a sequence that is repetitive, inside one team, with written rules. Your six sequences are the example: they all look alike, every day.
What a full client dispute is: a file that crosses sales administration, the technical department, accounts and management, with regulatory deadlines, credit notes, a commercial decision in the middle and a different path depending on the amount. That is no longer a sequence, it is a process — a chain of steps crossing several departments, where each one waits on a decision from the one before.
The path I propose, and it starts with what I can do tomorrow:
· I take the three parts of the dispute that are already repetitive: building the claim file, gathering the evidence from the trace, preparing the credit note as a draft. Across the quarter's 41 disputes, that is 62 % of the time spent.
· The rest — the arbitration, the signed credit note, the legal follow-up — belongs to process automation, which knows how to hold a file across several departments and several weeks. That is an offer distinct from mine, and I point you to it rather than pretending.
· And if you connect a second agent — a telephone reception agent, say — do not run them side by side: there is an agent platform that provides the shared foundation, the access rights and a single log. Two agents without a foundation means two logs to reconcile by hand.
What I propose concretely: I build the three dispute sequences within a fortnight, and the 62 % table will serve as your file for deciding whether the rest is worth a process offer. The decision will be taken on a figure, not on a promise. scope_what-the-agent-does-and-what-extends-it.pdf62 % of the dispute tomorrow, the rest via the dedicated offer
⛓ Sourced · the quarter's 41 disputes, time per step, scope of the orchestration and process offers
Last question, the one our food-retail client asks us: where does all this actually go?
Through a machine in your own workshop, and nowhere else. What your tools exchange — your prices, your margins, your client list — does not leave the company.
Local inference means the model computes on your machine: a part number, a supplier's negotiated price or a client's name crosses no outside 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 installer, nor with another contracting client.
What that means in your trade, line by line:
· Your negotiated price lists serve you and only you. What I learn about your supplier prices feeds no public model and surfaces nowhere else.
· Encryption in transit and at rest, and role-based access narrower than that of your own users: I have never had access to the payroll module of your ERP, even though it is in the same software.
· A complete, retained log: who asked for what, when, with what result and on whose authority. It is that log that answers your contracting client — and an inspection.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity, including for the data your five tools exchange.
The figure that makes this commercial rather than technical: 2 of your contracting clients — a food-retail buying group and the central kitchen of a hospital — already require hosting within the European Union in their framework contract, and the last institutional-catering tender you answered made it a scored criterion. You were able to tick the box without reservation, with the evidence to back it.
What I propose to close the quarter: I keep up to date the technical sheet your contracting clients ask for at renewal — hosting, subprocessors, retention periods, who has access to what, and the list of the two mandates in force with their review dates. It is requested once a year and takes three days to hunt down. And the next step, if you want it: let us connect the round completion, 380 runs a month, the last big volume still done by hand. technical-framework_where-your-exchanges-travel.pdfLocal inference, processing in the EU targeted, 2 mandates in force
✎ Framework · deployment architecture, rights matrix, framework contracts of the contracting clients
Local inference means the model computes on your machine: a part number, a supplier's negotiated price or a client's name crosses no outside 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 installer, nor with another contracting client.
What that means in your trade, line by line:
· Your negotiated price lists serve you and only you. What I learn about your supplier prices feeds no public model and surfaces nowhere else.
· Encryption in transit and at rest, and role-based access narrower than that of your own users: I have never had access to the payroll module of your ERP, even though it is in the same software.
· A complete, retained log: who asked for what, when, with what result and on whose authority. It is that log that answers your contracting client — and an inspection.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity, including for the data your five tools exchange.
The figure that makes this commercial rather than technical: 2 of your contracting clients — a food-retail buying group and the central kitchen of a hospital — already require hosting within the European Union in their framework contract, and the last institutional-catering tender you answered made it a scored criterion. You were able to tick the box without reservation, with the evidence to back it.
What I propose to close the quarter: I keep up to date the technical sheet your contracting clients ask for at renewal — hosting, subprocessors, retention periods, who has access to what, and the list of the two mandates in force with their review dates. It is requested once a year and takes three days to hunt down. And the next step, if you want it: let us connect the round completion, 380 runs a month, the last big volume still done by hand. technical-framework_where-your-exchanges-travel.pdfLocal inference, processing in the EU targeted, 2 mandates in force
✎ Framework · deployment architecture, rights matrix, framework contracts of the contracting clients
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 business chains. All these uses work in support, subject to your approval.
Included in your agent
The 4 capabilities essential to this promise are included, at no extra cost.
From 711 € excl. VAT / month
Running the chains
Reads, creates, updates and notifies in the tools you have connected.
Approval points
Stops at the actions you have defined as committing.
Full log
Keeps the tool, timestamp and result of every action performed.
Sovereign AI
The hosting and confidentiality foundation the agent rests on.
Controls and safeguards
This control is built into the agent: it frames what the agent does, whatever plan you pick. It is not chosen and is not added to your order.
Data protection, rights and human oversight
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.
Integration with existing tools without double entry
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 chains can a team hand over?
By taking on the repetitive execution, the effort shifts towards decisions and the unusual cases. How large the gain is depends on your volume and remains to be confirmed by a pilot.
Running the chain
Keying into the different tools
Building the record
Indicative figures, not contractual, to be confirmed by a pilot on your number of tools connected and chains. The actions you have defined as committing, and anything that leaves the company, wait for human approval.
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
An orchestrated agent (chains, approval points, log), installed and operated for you.
Agility
Setup + controlled subscription
9,430 € excl. VAT setup
then 711 € 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
1,236 € 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
13,385 € excl. VAT setup
then 912 € 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 automation
Your tools' data stays with youLocal inference or an isolated resource hosted in France; no data entrusted to a third party, no data used to train a model.
Data in France, under French lawThe data your tools exchange: minimisation and location in France, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
The business department keeps the decisionThe agent produces chains that are run and recorded, which can be checked and altered; no approval is automated.
Human oversight & traceabilityOn your number of tools connected and chains: systematic logging and tracking, in line with the AI Act.
Going further
Related resources
Frequently asked questions
Your questions, our answers
Does the agent really act inside our tools?
Yes, within the scope you have authorised tool by tool, and never beyond it. Every action is logged with its result.
Which actions wait for approval?
The ones you define as committing at the design stage — typically anything that leaves the company. The rest chain together without intervention.
What happens after an interruption?
The chain stops at the step concerned, the reason is logged, and it resumes from that stopping point rather than from the beginning: nothing is done twice.
How does this differ from process automation?
This agent runs business chains within a team; process automation orchestrates a complete case crossing several departments and systems.
Is the data exchanged 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.
How long does it take to deploy this agent?
A few months as a rule, depending on the number of tools connected and how complex the chains are, after a free audit then phases of design, integration and testing.
Related use cases
Going further
Let's talk
Let's size up the potential in your chains
15 minutes to frame your tools and your approval points — hosted in France, supervised, with no commitment.