● B2B offer — Knowledge & unified search
Company knowledge: one search, all your sources
The information exists: it is spread across the intranet, the file stores, the business tools and the team spaces in Microsoft Teams and Slack. Your agent offers a single search across the lot — queried from Microsoft Teams, Slack or email — strictly respecting each employee's access rights, and cites for every answer the source and the original location. Hosted in France: all of your internal content stays with you.
Hosted in France
Internal content protected
GDPR & AI Act: governed deployment
Human oversight
Updated on
Deployed in a few weeks
Company knowledge · hosted in France
Who has worked on this kind of case before, and where are the deliverables?
Four sources concerned: one project space, two file folders and an intranet page.
For each: the exact location and the date last changed.
These results correspond to the spaces you have access to.
🔗 Sourced · internal sources, within your access rights
For each: the exact location and the date last changed.
These results correspond to the spaces you have access to.
🔗 Sourced · internal sources, within your access rights
There must be other documents elsewhere.
I only look at the spaces your rights allow: the others are no more accessible to me than to you.
Widening your access rests with the administrators of each space — the agent never opens an access itself.
✎ Support · access rights unchanged
Widening your access rests with the administrators of each space — the agent never opens an access itself.
✎ Support · access rights unchanged
Local inference · no data outside the EU
Sources hosted in France
Sovereign by designLocal inference or hosting in France
GDPR & AI Act: governed deploymentTraceability & human oversight
TurnkeyDesigned, installed and operated for you
Your teams decideThe agent prepares, never rules
✦ In brief
A Blue Lemon Agent knowledge management agent unifies search across all your internal sources — intranet, file stores, business tools — while respecting each employee's access rights and citing the original location of every result. It never opens an access itself. It runs on local inference or is hosted in France: your internal content 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
search 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 sources connected and users is confirmed by a pilot.
The context
What does an AI agent bring to your company's knowledge?
What the company already knows is its most profitable asset — provided an employee can find it from a single place.
! The issue
A company's knowledge is complete but scattered: each tool holds part of the answer. A unified search brings those parts together, provided it respects the access rights of whoever is asking. The agent applies each employee's exact scope and cites the original location of every result.
✓ Our answer
Every employee questions all the sources they are entitled to, from a single point of entry, and gets back to the original location in one click. Access rights stay managed by the administrators of each space: the agent never widens them. Local inference or an isolated resource hosted in France: all of your internal content does not leave the company.
The decisive point
All of your internal content: sovereignty & compliance
A unified search touches the company's entire body of information: respecting access rights is the central guarantee. Here is how it is upheld.
Local inference
The agent can run on a machine belonging to your organisation: no internal content leaves the network.
Hosting in France
Otherwise, a dedicated and isolated resource hosted in France, under French law — your scattered information sources: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
For all of your internal content, 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 workspaces.
Each person's rights, applied to the letter
An employee sees only what their rights already allow; encryption, role-based access and logging of every search.
AI Act: governed deployment
The agent is strictly in support; no access is opened and no content is changed 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 executive spaces, employment data and confidential projects, 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 companyVentaline — design office and manufacturer of ventilation systems for commercial buildings
- Sector
- Design, manufacture and maintenance of ventilation systems for commercial buildings — in-house design office and after-sales service
- Headcount
- 240 staff, including 38 in the design office, 26 in after-sales, 12 in purchasing and 9 in quality; 186 people hold an account on at least one of the connected sources
- Market served
- Commercial property owners, social landlords and hospitals — 410 sites delivered over ten years, 190 of them still under warranty or under a maintenance contract
- Volume
- 1.9 million files on the servers, 12 years of intranet, 74,000 team discussion threads — and 2,900 information searches a month, all logged
- Tools in place
- Intranet, two file repositories, the design office's document management system, the after-sales ticketing tool and the team messaging platform — the agent plugs into them read-only, nothing is replaced and nothing is migrated
- Who decides
- The head of information systems settles the list of connected sources; the administrators of each space remain the only people who can open access; the design office director designates what counts as the reference version
- Room for improvement
- 60 % of the time spent on a search goes into querying the sources one after another; tracing where a document came from takes another 30 %; and 35 % of searches end with a question put to a colleague — a successful search takes 16 minutes on average
Ventaline is not short of information: it has twelve years of it, spread across six tools that do not talk to each other. What the company already knows is its most profitable asset, provided a member of staff can find it from a single place — and without seeing one byte more than their rights already allow. The agent runs on local inference on a machine at Ventaline and reads the six sources without ever writing to them. The exchanges below span a quarter, from the initial indexing 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.
Company knowledge · hosted in France
I indexed your six sources over the weekend — 1.9 million files, twelve years of intranet, 74,000 discussion threads — and I have pulled out a figure your access logs were carrying without anyone ever reading it.
Indexing means reading the content of every document so it can then be found by what it says, and not by the file name or the folder it sits in. Nothing is copied, nothing is moved, nothing is renamed: your current paths work exactly as they did yesterday.
The gap I measured, and it is the one that decides the gain: out of your 2,900 searches this month, 1,276 were re-run word for word on a second source in under ten minutes — 44 %. A search re-run identically elsewhere is a search that did not succeed the first time. Your staff search well: they search six times.
What that costs, across three items you already measure:
· Querying the sources one after another: 60 % of the time spent on a search.
· Tracing where the document found came from, and whether it governs: another 30 %.
· 35 % of searches end with a question put to a colleague — the most expensive item of all, because it is paid twice: by the person searching and by the person interrupted.
What I am putting in place: a single point of entry across the six sources, read-only, where every answer arrives with its exact location and its last-modified date. A successful search goes from 16 minutes to 5.
The next step, and it is the first thing to settle: the map of the six sources is drawn up, with for each one the number of spaces and the number of people entitled to them today. Hand it to your head of information systems: she settles the connection order, and I start with the sources she names. map-of-the-six-sources_what-is-indexed.pdf1.9 M files, 6 sources, 0 content copied
⛓ Sourced · access logs of the six sources over 30 days, inventory of spaces and their rights
Indexing means reading the content of every document so it can then be found by what it says, and not by the file name or the folder it sits in. Nothing is copied, nothing is moved, nothing is renamed: your current paths work exactly as they did yesterday.
The gap I measured, and it is the one that decides the gain: out of your 2,900 searches this month, 1,276 were re-run word for word on a second source in under ten minutes — 44 %. A search re-run identically elsewhere is a search that did not succeed the first time. Your staff search well: they search six times.
What that costs, across three items you already measure:
· Querying the sources one after another: 60 % of the time spent on a search.
· Tracing where the document found came from, and whether it governs: another 30 %.
· 35 % of searches end with a question put to a colleague — the most expensive item of all, because it is paid twice: by the person searching and by the person interrupted.
What I am putting in place: a single point of entry across the six sources, read-only, where every answer arrives with its exact location and its last-modified date. A successful search goes from 16 minutes to 5.
The next step, and it is the first thing to settle: the map of the six sources is drawn up, with for each one the number of spaces and the number of people entitled to them today. Hand it to your head of information systems: she settles the connection order, and I start with the sources she names. map-of-the-six-sources_what-is-indexed.pdf1.9 M files, 6 sources, 0 content copied
⛓ Sourced · access logs of the six sources over 30 days, inventory of spaces and their rights
The set includes the HR space and the executive space. A workshop supervisor must not stumble onto a salary table because someone plugged in a search engine.
He will not stumble onto it, and not because I decided so: it is your rights that answer, not mine.
The access perimeter is the exact list of spaces a person can already open with their account. I have no other one: I query the six sources ON BEHALF of whoever asks the question.
The proof runs on three accounts, same question — “what is the year-end bonus policy?”:
· The workshop supervisor receives the intranet note open to everyone, dated. Nothing else exists for him.
· The head of human resources also receives the two internal notes from the HR space and the company agreement.
· The managing director also receives the budget framing note from the executive space.
Three answers to the same question: each person sees what they already saw, and now they find it.
The three guarantees that hold this sealing, and you can have them written into the contract:
· Opening an access stays with the administrator of each space, and with them alone — it is a technical power I do not have, not an instruction I comply with.
· An extract stays inside the perimeter of whoever is asking, summaries included. The summary of a confidential document is a confidential document.
· When the answer sits in a closed space, you get the path: “a document answers your question in the HR space; its owner is the head of human resources.” Ten seconds, instead of three follow-ups and an answer the next day.
What I propose: that your head of information systems reviews the 41 spaces and 186 accounts before go-live. Out of the 41, 7 carry rights broader than your internal policy provides for, and unified search is what finally puts them side by side. The list of seven is ready, proposed tightening and review date included: a benefit you had not asked for, and it lands before go-live. access-perimeters_41-spaces-186-accounts.pdf3 accounts, one question, three answers
⛓ Sourced · account directory, declared rights of the 41 spaces, internal authorisation policy
The access perimeter is the exact list of spaces a person can already open with their account. I have no other one: I query the six sources ON BEHALF of whoever asks the question.
The proof runs on three accounts, same question — “what is the year-end bonus policy?”:
· The workshop supervisor receives the intranet note open to everyone, dated. Nothing else exists for him.
· The head of human resources also receives the two internal notes from the HR space and the company agreement.
· The managing director also receives the budget framing note from the executive space.
Three answers to the same question: each person sees what they already saw, and now they find it.
The three guarantees that hold this sealing, and you can have them written into the contract:
· Opening an access stays with the administrator of each space, and with them alone — it is a technical power I do not have, not an instruction I comply with.
· An extract stays inside the perimeter of whoever is asking, summaries included. The summary of a confidential document is a confidential document.
· When the answer sits in a closed space, you get the path: “a document answers your question in the HR space; its owner is the head of human resources.” Ten seconds, instead of three follow-ups and an answer the next day.
What I propose: that your head of information systems reviews the 41 spaces and 186 accounts before go-live. Out of the 41, 7 carry rights broader than your internal policy provides for, and unified search is what finally puts them side by side. The list of seven is ready, proposed tightening and review date included: a benefit you had not asked for, and it lands before go-live. access-perimeters_41-spaces-186-accounts.pdf3 accounts, one question, three answers
⛓ Sourced · account directory, declared rights of the 41 spaces, internal authorisation policy
All of this assumes you read our drawings, our contracts and our internal exchanges. Where exactly do they end up?
They stay with you. I run on local inference on a Ventaline machine, and your six sources are copied nowhere.
Local inference means the model computes on your machine: the text of a drawing or a contract crosses no outside network to be processed. If you would rather not administer a machine, the other route is an isolated resource hosted in France, dedicated to your company alone — no sharing.
What that lets you write in a tender file:
· Your content trains no model, neither ours nor a third party's. What I learn from Ventaline serves Ventaline.
· I read your six sources and stop there: the technical account I read through has no write permission — sturdier than a promise, and your administrator checks it in three clicks.
· Encryption in transit and at rest, and role-based access — rights follow the job: an after-sales technician opens site files, not distribution contracts.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
· A complete log: who searched for what, when, and what the system returned.
The figure that makes this commercial rather than technical: of your 190 sites still under contract, 34 belong to healthcare establishments or public bodies whose specifications already require hosting in France. The last tender you answered made it a scored criterion — and you ticked “under review”.
What I propose: that I keep up to date the technical sheet your clients ask for at renewal — hosting, subcontractors, retention periods, who accesses what. It costs you three days a year to reconstruct; it will be permanently current, and you will tick “yes” at the next tender. technical-framework_where-your-content-lives.pdfLocal inference, read-only, processing in the EU targeted
✎ Framework · deployment architecture, rights of the technical read account, specifications of the 34 contracts concerned
Local inference means the model computes on your machine: the text of a drawing or a contract crosses no outside network to be processed. If you would rather not administer a machine, the other route is an isolated resource hosted in France, dedicated to your company alone — no sharing.
What that lets you write in a tender file:
· Your content trains no model, neither ours nor a third party's. What I learn from Ventaline serves Ventaline.
· I read your six sources and stop there: the technical account I read through has no write permission — sturdier than a promise, and your administrator checks it in three clicks.
· Encryption in transit and at rest, and role-based access — rights follow the job: an after-sales technician opens site files, not distribution contracts.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
· A complete log: who searched for what, when, and what the system returned.
The figure that makes this commercial rather than technical: of your 190 sites still under contract, 34 belong to healthcare establishments or public bodies whose specifications already require hosting in France. The last tender you answered made it a scored criterion — and you ticked “under review”.
What I propose: that I keep up to date the technical sheet your clients ask for at renewal — hosting, subcontractors, retention periods, who accesses what. It costs you three days a year to reconstruct; it will be permanently current, and you will tick “yes” at the next tender. technical-framework_where-your-content-lives.pdfLocal inference, read-only, processing in the EU targeted
✎ Framework · deployment architecture, rights of the technical read account, specifications of the 34 contracts concerned
We are looking for the acoustic calculation note for the Loudes Hospital site. Three people looked for it yesterday, none of them found it.
It exists in three versions, across two different sources — and I will tell you which one governs.
What comes back:
· Acoustic calculation note, revision C, dated 14/09/2023 — design office file repository, folder Sites / 2023 / LOUDES / 04-Acoustics. This is the one that governs: it is the revision attached to the as-built file — the record handed to the client at delivery; it describes the works as actually built, and it is the document referred to in a dispute.
· Revision B, dated 02/06/2023, same folder — a study version, prior to the modification of the level 3 air-handling units. It is still circulating by e-mail: it is the one your three colleagues opened yesterday.
· Revision A, dated 11/04/2023, in the design office's document management system — the tool where the design office files its documents with their successive revisions. Never issued.
Every result carries three things, and that is where the gain sits: the exact location, the last-modified date, and the reason I rank it first. Tracing where a document came from and whether it governs took 30 % of the time spent on a search; it now takes 4 % — and those 4 % are one click, not an investigation.
The fact I bring up without being asked: revision B is attached to 11 e-mails sent to the project manager between June and September 2023. If a question arises today about the acoustics of level 3, that is the version your counterpart has in front of them, not yours. So you have a point to prepare before the meeting, and you have it today rather than in the room.
What I propose: that the design office director marks “governs” on the revisions attached to the as-built files of the 190 sites still under contract. Half a day of his time, and the most frequent question in your searches disappears — 412 of the month's 2,900 searches asked “which is the right version”. search-result_loudes-acoustic-note.pdf3 versions, 1 that governs, 11 e-mails concerned
⛓ Sourced · design office file repository, document management system, Loudes site e-mails
What comes back:
· Acoustic calculation note, revision C, dated 14/09/2023 — design office file repository, folder Sites / 2023 / LOUDES / 04-Acoustics. This is the one that governs: it is the revision attached to the as-built file — the record handed to the client at delivery; it describes the works as actually built, and it is the document referred to in a dispute.
· Revision B, dated 02/06/2023, same folder — a study version, prior to the modification of the level 3 air-handling units. It is still circulating by e-mail: it is the one your three colleagues opened yesterday.
· Revision A, dated 11/04/2023, in the design office's document management system — the tool where the design office files its documents with their successive revisions. Never issued.
Every result carries three things, and that is where the gain sits: the exact location, the last-modified date, and the reason I rank it first. Tracing where a document came from and whether it governs took 30 % of the time spent on a search; it now takes 4 % — and those 4 % are one click, not an investigation.
The fact I bring up without being asked: revision B is attached to 11 e-mails sent to the project manager between June and September 2023. If a question arises today about the acoustics of level 3, that is the version your counterpart has in front of them, not yours. So you have a point to prepare before the meeting, and you have it today rather than in the room.
What I propose: that the design office director marks “governs” on the revisions attached to the as-built files of the 190 sites still under contract. Half a day of his time, and the most frequent question in your searches disappears — 412 of the month's 2,900 searches asked “which is the right version”. search-result_loudes-acoustic-note.pdf3 versions, 1 that governs, 11 e-mails concerned
⛓ Sourced · design office file repository, document management system, Loudes site e-mails
Half the answers are written down nowhere: they float around in the team discussion threads.
I read them, I bring them back, and I label them for what they are: a lead, not a reference.
On the question “why were the level 3 air-handling units modified at Loudes?”, no document answers. Three messages in the #loudes-site thread, from 22 to 24 August 2023, between the account manager and the production manager, do answer it: a ceiling-height constraint found on site, and the project manager's verbal agreement the following day.
How I present them, and why it matters: a team message is dated and signed, but nobody has reviewed it. It gives you the explanation and the name of the person who knows it, and I always mark it apart from a contractual document. Among those three messages, the project manager's agreement is confirmed by no document in the file: it is the kind of gap that one e-mail settles today, and you have it in front of you five years before it could be invoked.
What reading the threads gives you straight away: 35 % of your searches used to end with a question put to a colleague. 6 % now do — the ones where you really do need to talk to someone. The rest find the answer AND the name of the person who wrote it, without interrupting them.
The move that turns a thread into an asset, and I have already prepared it: the note “modification of level 3 air-handling units” is drafted from those three messages, with their dates and their authors. The account manager approves it in three lines or corrects it with a word — two minutes —, it joins the site file and becomes a reference.
What I propose next: I have spotted 46 of the same kind across the 190 sites under contract, today alive only in the memory of four people. 46 notes × 2 minutes of approval = half a day of human work, for twelve years of explanations put in writing. Tell me which sites to start with and I will bring them in that order. team-threads_46-explanations-to-fix.pdfWhat a message proves, and what it does not
⛓ Sourced · team messaging, #loudes-site thread, documents in the site file
On the question “why were the level 3 air-handling units modified at Loudes?”, no document answers. Three messages in the #loudes-site thread, from 22 to 24 August 2023, between the account manager and the production manager, do answer it: a ceiling-height constraint found on site, and the project manager's verbal agreement the following day.
How I present them, and why it matters: a team message is dated and signed, but nobody has reviewed it. It gives you the explanation and the name of the person who knows it, and I always mark it apart from a contractual document. Among those three messages, the project manager's agreement is confirmed by no document in the file: it is the kind of gap that one e-mail settles today, and you have it in front of you five years before it could be invoked.
What reading the threads gives you straight away: 35 % of your searches used to end with a question put to a colleague. 6 % now do — the ones where you really do need to talk to someone. The rest find the answer AND the name of the person who wrote it, without interrupting them.
The move that turns a thread into an asset, and I have already prepared it: the note “modification of level 3 air-handling units” is drafted from those three messages, with their dates and their authors. The account manager approves it in three lines or corrects it with a word — two minutes —, it joins the site file and becomes a reference.
What I propose next: I have spotted 46 of the same kind across the 190 sites under contract, today alive only in the memory of four people. 46 notes × 2 minutes of approval = half a day of human work, for twelve years of explanations put in writing. Tell me which sites to start with and I will bring them in that order. team-threads_46-explanations-to-fix.pdfWhat a message proves, and what it does not
⛓ Sourced · team messaging, #loudes-site thread, documents in the site file
Our people do not search with the same words. The design office writes “air-handling unit”, the workshop says “box”, and after-sales types the commercial reference.
That is the leading cause of your failed searches, and it is fixed with your own documents — not with a dictionary bought elsewhere.
What I built: a glossary of 340 equivalences, drawn from your twelve years of documents. Each equivalence holds because both terms designate the same reference in your parts lists, never because they look alike.
· “air-handling unit” = “box” = “AHU” = the 14 commercial references in the range — four ways of naming one machine, four populations of staff.
· 218 in-house acronyms expanded, of which 31 carry two meanings depending on the department. “PD” means “peak duty” in the design office and “price demand” in purchasing: I return the two families of results separately, each labelled with its department, and everyone recognises theirs at first glance.
· 12 site names changed along the way — Loudes Hospital was called “Loudes GH” for two years, then “Western Trust — Loudes site”. Three names, one site: your searches used to bring back a third of it, now they bring back all of it.
The result measured over the month: querying the sources one after another goes from 60 % to 8 % of the time spent on a search, and the share of searches that succeed first time rises from 56 % to 84 %.
What I advise against, and I put a number on it: adding “general trade” synonyms taken from outside your walls. On a trial with 900 generic equivalences, searches succeeding first time fell back to 71 %: a term that is not yours brings back documents that are not the answer. 340 true equivalences beat 900 plausible ones.
What I propose next: that any member of staff can correct the glossary with a word, and that I report each month the ten most-searched terms that return nothing — they are what writes the rest of the glossary, which grows on your material and stays yours. in-house-glossary_340-equivalences.pdf31 double-meaning acronyms, 12 renamed sites
⛓ Sourced · product parts lists, twelve years of documents, log of terms searched with no result
What I built: a glossary of 340 equivalences, drawn from your twelve years of documents. Each equivalence holds because both terms designate the same reference in your parts lists, never because they look alike.
· “air-handling unit” = “box” = “AHU” = the 14 commercial references in the range — four ways of naming one machine, four populations of staff.
· 218 in-house acronyms expanded, of which 31 carry two meanings depending on the department. “PD” means “peak duty” in the design office and “price demand” in purchasing: I return the two families of results separately, each labelled with its department, and everyone recognises theirs at first glance.
· 12 site names changed along the way — Loudes Hospital was called “Loudes GH” for two years, then “Western Trust — Loudes site”. Three names, one site: your searches used to bring back a third of it, now they bring back all of it.
The result measured over the month: querying the sources one after another goes from 60 % to 8 % of the time spent on a search, and the share of searches that succeed first time rises from 56 % to 84 %.
What I advise against, and I put a number on it: adding “general trade” synonyms taken from outside your walls. On a trial with 900 generic equivalences, searches succeeding first time fell back to 71 %: a term that is not yours brings back documents that are not the answer. 340 true equivalences beat 900 plausible ones.
What I propose next: that any member of staff can correct the glossary with a word, and that I report each month the ten most-searched terms that return nothing — they are what writes the rest of the glossary, which grows on your material and stays yours. in-house-glossary_340-equivalences.pdf31 double-meaning acronyms, 12 renamed sites
⛓ Sourced · product parts lists, twelve years of documents, log of terms searched with no result
A client is calling about a site delivered in 2017. The account manager of the time has left, we cannot find anything, and he is on the line.
The file is rebuilt: 34 documents across five sources, in forty seconds, and I am giving you the reading order so you can answer during the call.
The three facts that answer his question, first:
· Handover accepted on 08/06/2017, with two reservations cleared on 21/07/2017 — the clearance record is in the file repository.
· Equipment good-working-order warranty: two years from handover, so expired since 08/06/2019. The maintenance contract, however, is still running — tacit renewal, next expiry 31/12.
· The fault he describes has already been dealt with: three tickets in your after-sales tool in 2021 and 2022 on the same unit, including one motor replacement invoiced and one adjustment under contract.
What it changes in the call: your technician says “we have already been out three times on that unit, the last in March 2022, and you are covered by the maintenance contract”, instead of “I will call you back”. Same information, nine minutes earlier — and that is the difference between a reassured client and a client who calls your management.
The item it empties, measured over your month: 35 % of searches used to end with a question to a colleague; 6 % now do. For after-sales taken alone, it was 61 % — the department most dependent on other people's memory, because it works on sites it did not build.
What I bring up unasked, and it is worth revenue: of the 190 sites still under contract, 23 have their maintenance expiring within four months, and 7 have received no visit for more than eighteen months. Those are the ones a competitor takes over at renewal. The dated list is ready: your after-sales manager settles the calling order, and I regenerate it every month. 2017-site-file_34-documents-rebuilt.pdf5 sources, 3 interventions found, 23 expiries
⛓ Sourced · site file repository, after-sales ticketing tool, maintenance contracts, intranet
The three facts that answer his question, first:
· Handover accepted on 08/06/2017, with two reservations cleared on 21/07/2017 — the clearance record is in the file repository.
· Equipment good-working-order warranty: two years from handover, so expired since 08/06/2019. The maintenance contract, however, is still running — tacit renewal, next expiry 31/12.
· The fault he describes has already been dealt with: three tickets in your after-sales tool in 2021 and 2022 on the same unit, including one motor replacement invoiced and one adjustment under contract.
What it changes in the call: your technician says “we have already been out three times on that unit, the last in March 2022, and you are covered by the maintenance contract”, instead of “I will call you back”. Same information, nine minutes earlier — and that is the difference between a reassured client and a client who calls your management.
The item it empties, measured over your month: 35 % of searches used to end with a question to a colleague; 6 % now do. For after-sales taken alone, it was 61 % — the department most dependent on other people's memory, because it works on sites it did not build.
What I bring up unasked, and it is worth revenue: of the 190 sites still under contract, 23 have their maintenance expiring within four months, and 7 have received no visit for more than eighteen months. Those are the ones a competitor takes over at renewal. The dated list is ready: your after-sales manager settles the calling order, and I regenerate it every month. 2017-site-file_34-documents-rebuilt.pdf5 sources, 3 interventions found, 23 expiries
⛓ Sourced · site file repository, after-sales ticketing tool, maintenance contracts, intranet
In purchasing, we would like to know what the company already knows about a supplier before placing another order. Today, we ask around.
On the supplier Delmont Ventilation, the company knows eleven things spread across five sources — and three of them appear in no purchasing document.
What your purchasing team had: the framework contract expiring in five months, the tiered price list, the order history — €2.1 M over three years, 41 orders.
What the company knew elsewhere:
· Quality: 4 non-conformities opened in eighteen months on series 200 motors, of which 2 with a late-delivery penalty applied — a penalty already applied is a pricing argument already documented.
· Design office: a 2024 note recommends no longer specifying them above 8,000 m³/h, for want of an acoustic test record at that flow rate — it bounds your forecast volumes.
· After-sales: 22 days of average spare-parts lead time, against 9 days at their direct competitor. That is the costliest item in maintenance, and it has never entered a negotiation.
The call, I put it on the table with figures: those eleven facts weigh 2 penalties applied, 4 non-conformities opened and 13 days of lead time more than the direct competitor, on €2.1 M of purchases. The full comparison fits on one page and I am missing exactly two inputs: your forecast volumes for the coming year, and the prices submitted at the last tender. Those are not analyses, they are two files — send them over and you have the comparison tomorrow morning. Renewing or re-tendering belongs to your head of purchasing, and the decision will be made in ten minutes on figures.
What I propose next: the same sheet for your 60 active suppliers, regenerated at every renewal — of the six drawn up as a trial, four already carried a fact known to one department and unknown to purchasing. And that quality non-conformities reach purchasing as soon as they are opened: that is a connection, not a project. supplier-sheet_what-the-company-already-knows.pdf11 facts, 5 sources, 3 unknown to purchasing
⛓ Sourced · purchasing framework contracts, quality tool, design office notes, after-sales tickets
What your purchasing team had: the framework contract expiring in five months, the tiered price list, the order history — €2.1 M over three years, 41 orders.
What the company knew elsewhere:
· Quality: 4 non-conformities opened in eighteen months on series 200 motors, of which 2 with a late-delivery penalty applied — a penalty already applied is a pricing argument already documented.
· Design office: a 2024 note recommends no longer specifying them above 8,000 m³/h, for want of an acoustic test record at that flow rate — it bounds your forecast volumes.
· After-sales: 22 days of average spare-parts lead time, against 9 days at their direct competitor. That is the costliest item in maintenance, and it has never entered a negotiation.
The call, I put it on the table with figures: those eleven facts weigh 2 penalties applied, 4 non-conformities opened and 13 days of lead time more than the direct competitor, on €2.1 M of purchases. The full comparison fits on one page and I am missing exactly two inputs: your forecast volumes for the coming year, and the prices submitted at the last tender. Those are not analyses, they are two files — send them over and you have the comparison tomorrow morning. Renewing or re-tendering belongs to your head of purchasing, and the decision will be made in ten minutes on figures.
What I propose next: the same sheet for your 60 active suppliers, regenerated at every renewal — of the six drawn up as a trial, four already carried a fact known to one department and unknown to purchasing. And that quality non-conformities reach purchasing as soon as they are opened: that is a connection, not a project. supplier-sheet_what-the-company-already-knows.pdf11 facts, 5 sources, 3 unknown to purchasing
⛓ Sourced · purchasing framework contracts, quality tool, design office notes, after-sales tickets
And when you find nothing? It is bound to happen.
It happens, I count it, and I publish it to you: of the month's 2,900 searches, 221 came back with no result — 7.6 %. That is the figure that does not flatter me, and it is better that it comes from me.
The cause, measured rather than assumed: 197 of the 221 concerned documents predating 2019, scanned at the time as plain images. An image of a page has no text: the document is there, it opens, and it stays invisible to a search. The other 24 concerned documents that do not exist — and knowing in ten seconds that a document does not exist beats hunting for it for three days: those 24 failures are already a gain.
What I did about it, and it is measured: I ran text recognition over the pre-2019 holdings — text recognition reads the words inside an image of a page: the scanned document becomes one you can read, search and quote — that is 22,400 documents in eleven nights, without touching a single original.
The following month: 63 searches with no result out of 3,000 — 2.1 %. And of those 63, 47 concern documents that genuinely do not exist.
The rule that makes everything else usable: a gap is a named gap. A missing test record stays missing, it never becomes “probably compliant”. I tell you what is missing, where I looked and who would have it — three lines that save the reader half a day. A guessed value would be indistinguishable from a read value, and it is confidence in the other 2,837 answers that is at stake there.
What I propose next: that I report each month the ten most frequent searches that return nothing. This month, three of them concern 2015-2018 acoustic test records your design office believed it had archived — that is management information, and it reaches you before the client asks for it. searches-returning-nothing_221-then-63.pdf7.6 % → 2.1 % after text recognition on 22,400 documents
⛓ Sourced · search log, pre-2019 scanned holdings, following month's figures
The cause, measured rather than assumed: 197 of the 221 concerned documents predating 2019, scanned at the time as plain images. An image of a page has no text: the document is there, it opens, and it stays invisible to a search. The other 24 concerned documents that do not exist — and knowing in ten seconds that a document does not exist beats hunting for it for three days: those 24 failures are already a gain.
What I did about it, and it is measured: I ran text recognition over the pre-2019 holdings — text recognition reads the words inside an image of a page: the scanned document becomes one you can read, search and quote — that is 22,400 documents in eleven nights, without touching a single original.
The following month: 63 searches with no result out of 3,000 — 2.1 %. And of those 63, 47 concern documents that genuinely do not exist.
The rule that makes everything else usable: a gap is a named gap. A missing test record stays missing, it never becomes “probably compliant”. I tell you what is missing, where I looked and who would have it — three lines that save the reader half a day. A guessed value would be indistinguishable from a read value, and it is confidence in the other 2,837 answers that is at stake there.
What I propose next: that I report each month the ten most frequent searches that return nothing. This month, three of them concern 2015-2018 acoustic test records your design office believed it had archived — that is management information, and it reaches you before the client asks for it. searches-returning-nothing_221-then-63.pdf7.6 % → 2.1 % after text recognition on 22,400 documents
⛓ Sourced · search log, pre-2019 scanned holdings, following month's figures
What do you do on your own, without being asked? That is our executive committee's first question.
Three actions, not one more, and each can be withdrawn with a word. Everything else waits for your request.
· I index recent deposits every night. A document filed today is findable tomorrow. And the reverse is true too: a document deleted from a source leaves the index at the same hour — I keep no copy of a document you have erased.
· I follow rights when they change. Someone moving department sees their perimeter follow the same day; a departure closes access within the hour.
· I report the ten most-searched terms that return nothing, once a month, to your head of information systems. It is my only unprompted send, and it carries terms, never a searcher's name.
And four guarantees hold by construction, which is what makes the rest signable: your six sources stay read-only, opening an access belongs to your space administrators, a document is deleted only by you, and ranking people stays outside my perimeter.
On that last point, the reason is mechanical: the day “searches per employee” becomes a tracked indicator, it becomes a target; people search for the counter, and the tool that showed you what the company does not know shows only the reaction to the counter.
What I measure instead: the questions asked, the sources that answer them, and the ones that never answer. This month, one of your six sources served an answer only 14 times out of 2,900 searches — a document governance question, asked without putting a single name on it.
And if you want an indicator by department — not by person —, I produce it: it is your decision. The three documents that make it sound are drafted: prior information to staff, a written purpose, a retention period. Set the start date and it runs from the following month. what-the-agent-does-alone_3-actions.pdf3 reversible actions, 4 non-configurable prohibitions
✎ Framework · settings of the three automatic actions, indexing log, usage report by source
· I index recent deposits every night. A document filed today is findable tomorrow. And the reverse is true too: a document deleted from a source leaves the index at the same hour — I keep no copy of a document you have erased.
· I follow rights when they change. Someone moving department sees their perimeter follow the same day; a departure closes access within the hour.
· I report the ten most-searched terms that return nothing, once a month, to your head of information systems. It is my only unprompted send, and it carries terms, never a searcher's name.
And four guarantees hold by construction, which is what makes the rest signable: your six sources stay read-only, opening an access belongs to your space administrators, a document is deleted only by you, and ranking people stays outside my perimeter.
On that last point, the reason is mechanical: the day “searches per employee” becomes a tracked indicator, it becomes a target; people search for the counter, and the tool that showed you what the company does not know shows only the reaction to the counter.
What I measure instead: the questions asked, the sources that answer them, and the ones that never answer. This month, one of your six sources served an answer only 14 times out of 2,900 searches — a document governance question, asked without putting a single name on it.
And if you want an indicator by department — not by person —, I produce it: it is your decision. The three documents that make it sound are drafted: prior information to staff, a written purpose, a retention period. Set the start date and it runs from the following month. what-the-agent-does-alone_3-actions.pdf3 reversible actions, 4 non-configurable prohibitions
✎ Framework · settings of the three automatic actions, indexing log, usage report by source
We would also like to connect the after-sales tickets and the team messaging. There is client data and exchanges between staff in there: our management will not sign blind.
It will sign in full knowledge: everything is drawn up in draft, costed, in the form of a written mandate, capped, dated and withdrawable with a word.
What it brings in, because that is what decides: of the month's 221 searches with no result, 74 had their answer in an after-sales ticket or a team thread. Rebuilding an old file would go from 9 minutes to 2, and 380 more searches a month would succeed.
The mandate fits in five lines:
· Exact perimeter: the after-sales tickets and 19 team channels out of 74 — those named after a site. The other 55 stay out of the index, department, social, executive channels and the works council channel included.
· Even within the 19: direct messages between individuals stay out of the index, in any channel.
· Duration: review after three months, with the figures. Without an explicit decision at the review, the connection stops — it is renewal that requires a signature, not stopping.
· Prior information to staff: the text is ready and states what is indexed, what stays out, whom to contact.
· Withdrawal: a word from your head of information systems, de-indexing done in twenty minutes — nothing to dismantle, since nothing was copied.
What I advise against, with figures: connecting everything at once. Across the 55 non-site channels, only 4 % of messages carry reusable trade information: the 19 site channels already bring the 380 searches that count, while the other 55 would add fewer than 30 a month. The tighter perimeter gives most of the gain and a mandate signable on the spot.
Your management has everything it needs to decide: mandate drafted, staff note attached, review calendar at D+30 and D+90. You sign, and the connection is live the next morning. connection-mandate_2-sources-capped.pdf19 channels out of 74, review at 3 months, withdrawal in 20 minutes
✎ Framework · inventory of the 74 channels, log of searches returning nothing, draft mandate and information note
What it brings in, because that is what decides: of the month's 221 searches with no result, 74 had their answer in an after-sales ticket or a team thread. Rebuilding an old file would go from 9 minutes to 2, and 380 more searches a month would succeed.
The mandate fits in five lines:
· Exact perimeter: the after-sales tickets and 19 team channels out of 74 — those named after a site. The other 55 stay out of the index, department, social, executive channels and the works council channel included.
· Even within the 19: direct messages between individuals stay out of the index, in any channel.
· Duration: review after three months, with the figures. Without an explicit decision at the review, the connection stops — it is renewal that requires a signature, not stopping.
· Prior information to staff: the text is ready and states what is indexed, what stays out, whom to contact.
· Withdrawal: a word from your head of information systems, de-indexing done in twenty minutes — nothing to dismantle, since nothing was copied.
What I advise against, with figures: connecting everything at once. Across the 55 non-site channels, only 4 % of messages carry reusable trade information: the 19 site channels already bring the 380 searches that count, while the other 55 would add fewer than 30 a month. The tighter perimeter gives most of the gain and a mandate signable on the spot.
Your management has everything it needs to decide: mandate drafted, staff note attached, review calendar at D+30 and D+90. You sign, and the connection is live the next morning. connection-mandate_2-sources-capped.pdf19 channels out of 74, review at 3 months, withdrawal in 20 minutes
✎ Framework · inventory of the 74 channels, log of searches returning nothing, draft mandate and information note
Our data protection officer will ask what is logged and how long it is kept.
She will have the answer on one page, already written in the form she expects — so that she corrects it rather than writes it.
The log, here, is the time-stamped trace of every search: who asked, what, when, and what the system returned.
What is recorded: the account, the time stamp to the second, the question asked, the sources queried, the identifiers of the documents returned — their location, and that alone. Your documents stay where they are.
For how long: 12 rolling months, then automatic deletion. That is a setting: your officer sets the period, I apply it, and the change is dated in the log.
Who reads it: your head of information systems and your data protection officer, nobody else — a line manager stays out of it, the condition for the log to remain a security tool and not one for monitoring people.
What it is for, and this is where it pays for itself:
· Answering a staff member's access request: what concerns them, extracted in one operation.
· Proving, in the event of an incident, who consulted what. It is the only proof that carries weight with a client or an insurer, and you now have it.
· Feeding your record of processing activities: purpose, legal basis, categories, recipients, retention — entry drafted, your officer corrects it.
On the European regulation on artificial intelligence: the agent is in support — it answers, it does not act. No automated decision, no ranking of people, no writing to any system. Human supervision is the only way the system works.
What I propose: a one-hour quarterly review, where I arrive with the agenda filled in: sources added, periods modified, accounts closed. Where compliance checked once a year is discovered in breach, yours is corrected in ten minutes every three months. log-and-compliance_what-is-traced.pdf12 rolling months, 2 readers, 0 content copied
✎ Framework · log schema, configured retention periods, draft record-of-processing entry
The log, here, is the time-stamped trace of every search: who asked, what, when, and what the system returned.
What is recorded: the account, the time stamp to the second, the question asked, the sources queried, the identifiers of the documents returned — their location, and that alone. Your documents stay where they are.
For how long: 12 rolling months, then automatic deletion. That is a setting: your officer sets the period, I apply it, and the change is dated in the log.
Who reads it: your head of information systems and your data protection officer, nobody else — a line manager stays out of it, the condition for the log to remain a security tool and not one for monitoring people.
What it is for, and this is where it pays for itself:
· Answering a staff member's access request: what concerns them, extracted in one operation.
· Proving, in the event of an incident, who consulted what. It is the only proof that carries weight with a client or an insurer, and you now have it.
· Feeding your record of processing activities: purpose, legal basis, categories, recipients, retention — entry drafted, your officer corrects it.
On the European regulation on artificial intelligence: the agent is in support — it answers, it does not act. No automated decision, no ranking of people, no writing to any system. Human supervision is the only way the system works.
What I propose: a one-hour quarterly review, where I arrive with the agenda filled in: sources added, periods modified, accounts closed. Where compliance checked once a year is discovered in breach, yours is corrected in ten minutes every three months. log-and-compliance_what-is-traced.pdf12 rolling months, 2 readers, 0 content copied
✎ Framework · log schema, configured retention periods, draft record-of-processing entry
After a quarter, what has it given us back? I need a figure I can defend in committee.
531 hours a month, that is 2 h 50 per person across 186 users — and the calculation is yours, it fits in two lines.
The calculation, so you can redo it without me: 2,900 searches a month, 16 minutes on average before, 5 minutes after, measured on the logs of both periods. 11 minutes × 2,900 = 531 hours. At your fully loaded hourly cost of €38 — gross salary plus employer contributions, brought back to the hour worked —, that comes to €20,100 a month. That figure only holds if the time is put back to work, and it has been: what follows shows it.
On the three items you were measuring, to the unit:
· Querying the sources one after another: 60 % → 8 % of the time spent on a search.
· Tracing the original location and the version that governs: 30 % → 4 %.
· Interrupting a colleague to find a document: 35 % → 6 %.
What those hours became, according to your own records: after-sales brought its first-response time down from 4 h 10 to 1 h 20, and the design office answered two more tenders over the quarter, one of them won — it used to turn down four a quarter for lack of time. That is the only figure in this list that shows up in your revenue, and it comes from time given back, not from one more tool.
The figure that does not flatter me, published with the others: 221 searches with no result in the first month out of 2,900 — 7.6 %, brought down to 2.1 % after text recognition on the older holdings. 63 failures a month remain, 47 of them on documents that do not exist.
And the measure you had on none of your six sources: 0 access opened, 0 document modified, 0 content out of your walls, across 8,700 searches traced over the quarter. Take those four figures into committee: they are what makes the 531 hours hold. quarterly-review_531-hours-a-month.pdf60→8, 30→4, 35→6, and the calculation you can redo
⛓ Sourced · logs of both periods, after-sales response-time records, tender response tracking
The calculation, so you can redo it without me: 2,900 searches a month, 16 minutes on average before, 5 minutes after, measured on the logs of both periods. 11 minutes × 2,900 = 531 hours. At your fully loaded hourly cost of €38 — gross salary plus employer contributions, brought back to the hour worked —, that comes to €20,100 a month. That figure only holds if the time is put back to work, and it has been: what follows shows it.
On the three items you were measuring, to the unit:
· Querying the sources one after another: 60 % → 8 % of the time spent on a search.
· Tracing the original location and the version that governs: 30 % → 4 %.
· Interrupting a colleague to find a document: 35 % → 6 %.
What those hours became, according to your own records: after-sales brought its first-response time down from 4 h 10 to 1 h 20, and the design office answered two more tenders over the quarter, one of them won — it used to turn down four a quarter for lack of time. That is the only figure in this list that shows up in your revenue, and it comes from time given back, not from one more tool.
The figure that does not flatter me, published with the others: 221 searches with no result in the first month out of 2,900 — 7.6 %, brought down to 2.1 % after text recognition on the older holdings. 63 failures a month remain, 47 of them on documents that do not exist.
And the measure you had on none of your six sources: 0 access opened, 0 document modified, 0 content out of your walls, across 8,700 searches traced over the quarter. Take those four figures into committee: they are what makes the 531 hours hold. quarterly-review_531-hours-a-month.pdf60→8, 30→4, 35→6, and the calculation you can redo
⛓ Sourced · logs of both periods, after-sales response-time records, tender response tracking
The design office has 400,000 drawings with their successive revisions, and our staff ask the same twenty questions about leave and expenses every day. Do you take that on too?
Three dedicated agents do better than I do on those three subjects, they plug into the same six sources with the same rights, and I will tell you which takes what.
· Your 400,000 revised drawings belong to a technical document agent. A versioned set is one where every document exists in several successive revisions and where only the latest applicable one governs. I tell you where the drawing is and which revision governs; that agent compares two revisions, lists what changed between them and ties each drawing to its parts list. Across your 190 sites under contract, that is the question asked 640 times a quarter.
· Your twenty daily questions belong to an internal FAQ agent. Of the month's 2,900 searches, 470 concern twelve internal-life subjects — leave (112), expenses (94), remote working (71), badges (48) and eight others. I answer by quoting the note and its date; a dedicated intranet agent answers in plain language, keeps the history of the rule's versions and says what has changed since last year.
· Training your new joiners belongs to an internal training agent. Your 26 after-sales technicians recruited over two years take five months to learn where to look: that is learning procedures, not a search-engine matter.
Why I tell you rather than take it all on: an agent that accepts everything becomes average at everything, and your design office would pay for it first. All three share the six sources, the same rights and the same log — you connect nothing twice, and you redo no compliance file.
What I propose: start with the internal FAQ. It is the quickest to bring into service, and it takes 470 searches a month out of the document channel — which mechanically improves what I give you back on the other 2,430. Settle the order of the three and I will set the calendar this week. the-three-handovers_who-takes-what.pdf640 revision questions, 470 internal questions, 26 joiners
⛓ Sourced · breakdown of the 2,900 searches by nature, log of internal questions, after-sales recruitment records
· Your 400,000 revised drawings belong to a technical document agent. A versioned set is one where every document exists in several successive revisions and where only the latest applicable one governs. I tell you where the drawing is and which revision governs; that agent compares two revisions, lists what changed between them and ties each drawing to its parts list. Across your 190 sites under contract, that is the question asked 640 times a quarter.
· Your twenty daily questions belong to an internal FAQ agent. Of the month's 2,900 searches, 470 concern twelve internal-life subjects — leave (112), expenses (94), remote working (71), badges (48) and eight others. I answer by quoting the note and its date; a dedicated intranet agent answers in plain language, keeps the history of the rule's versions and says what has changed since last year.
· Training your new joiners belongs to an internal training agent. Your 26 after-sales technicians recruited over two years take five months to learn where to look: that is learning procedures, not a search-engine matter.
Why I tell you rather than take it all on: an agent that accepts everything becomes average at everything, and your design office would pay for it first. All three share the six sources, the same rights and the same log — you connect nothing twice, and you redo no compliance file.
What I propose: start with the internal FAQ. It is the quickest to bring into service, and it takes 470 searches a month out of the document channel — which mechanically improves what I give you back on the other 2,430. Settle the order of the three and I will set the calendar this week. the-three-handovers_who-takes-what.pdf640 revision questions, 470 internal questions, 26 joiners
⛓ Sourced · breakdown of the 2,900 searches by nature, log of internal questions, after-sales recruitment records
Last question, the one that will come up in committee: if we want to stop in a year, how stuck are we?
You are out in half a day, and you can verify it before signing rather than after.
What there is to dismantle, the day you stop:
· The index. It is deleted, and it contained none of your documents — only what is needed to find them where they are. Your 1.9 million files have not moved by a single byte: same folders, same names, same rights.
· The log. It is handed to you in an open format, or destroyed — you choose, and the choice is settled at go-live, not on the way out.
· The glossary of 340 equivalences and the 46 approved notes. They are yours: they are made of your own material, and they leave with you in a format readable without us. That is the only asset this deployment will have created, and its place is with you.
What makes the exit that short, and it is checkable with any vendor: no migration on the way in, therefore no migration on the way out. Your six tools stay in place, no member of staff has changed how they file anything, and no format belongs to us. A system you cannot leave is a system you should not have installed — and you are entitled to have that sentence written into the contract.
What I propose so that it does not stay a sentence: a dry-run exit at the end of the first quarter, half a day: we switch off, we check that your six sources work exactly as before, we switch back on. You will know what the promise is worth before committing to a second year — and it is the strongest argument you can take into committee. technical-framework_where-your-content-lives.pdfReversibility: 0 migration in, 0 migration out
✎ Framework · index architecture, export formats for the glossary and the log, dry-run exit protocol
What there is to dismantle, the day you stop:
· The index. It is deleted, and it contained none of your documents — only what is needed to find them where they are. Your 1.9 million files have not moved by a single byte: same folders, same names, same rights.
· The log. It is handed to you in an open format, or destroyed — you choose, and the choice is settled at go-live, not on the way out.
· The glossary of 340 equivalences and the 46 approved notes. They are yours: they are made of your own material, and they leave with you in a format readable without us. That is the only asset this deployment will have created, and its place is with you.
What makes the exit that short, and it is checkable with any vendor: no migration on the way in, therefore no migration on the way out. Your six tools stay in place, no member of staff has changed how they file anything, and no format belongs to us. A system you cannot leave is a system you should not have installed — and you are entitled to have that sentence written into the contract.
What I propose so that it does not stay a sentence: a dry-run exit at the end of the first quarter, half a day: we switch off, we check that your six sources work exactly as before, we switch back on. You will know what the promise is worth before committing to a second year — and it is the strongest argument you can take into committee. technical-framework_where-your-content-lives.pdfReversibility: 0 migration in, 0 migration out
✎ Framework · index architecture, export formats for the glossary and the log, dry-run exit protocol
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, all your information sources. All these uses work in support, subject to your approval.
Included in your agent
The 5 capabilities essential to this promise are included, at no extra cost.
From 708 € excl. VAT / month
Unified search
Questions the intranet, the file stores and the business tools all at once.
Access rights respected
Strictly applies each employee's scope, without ever widening it.
Original location cited
Refers back to the exact source and its date last changed.
Business document management
For a large, versioned body of business documents, a dedicated agent takes it on.
Internal FAQ
For employees' everyday questions, a dedicated intranet agent takes over.
Controls and safeguards
These 4 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
Explicit handling of questions with no answer and of conflicting sources
Inherited access rights, versions and freshness of the corpus
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.
Measurement of accuracy, coverage and time saved
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 much time does a company spend looking for what it already knows?
By unifying the search, the effort shifts towards using the information found. How large the gain is depends on your volume and remains to be confirmed by a pilot.
Searching source by source
Identifying the original location
Asking a colleague to find a document
Indicative figures, not contractual, to be confirmed by a pilot on your number of sources connected and users. The agent never opens an access: widening an employee's scope rests with the administrators of each space.
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 knowledge management agent (multi-source, rights, traceability), installed and operated for you.
Agility
Setup + controlled subscription
8,105 € excl. VAT setup
then 708 € 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
1,158 € excl. VAT /month
all inclusive, immediate start. No upfront investment: a single subscription. Ideal for starting quickly and simply.
- Setup included (installation, configuration, training)
- Operation, human oversight, updates and support
- Sovereign hosting in France, managed end to end
100% Sovereign
On site, you own it
11,780 € excl. VAT setup
then 905 € 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 body of information
Your internal content stays with youLocal inference or an isolated resource hosted in France; no content entrusted to a third party, no data used to train a model.
Data in France, under French lawAll of your internal content: minimisation and location in France, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
Your teams keep the decisionThe agent produces a unified search across all your sources, which can be checked and altered; no approval is automated.
Human oversight & traceabilityOn your number of sources connected and users: systematic logging and tracking, in line with the AI Act.
Going further
Related resources
Frequently asked questions
Your questions, our answers
Can an employee see more than usual?
No. The agent applies the existing access rights exactly: it shows nothing that is not already accessible to the person asking.
What sources can it connect to?
Intranet, file stores, business tools and common workspaces. The list is settled with you at the design stage.
Are the results traceable?
Yes: every result states the source, the exact location and the date last changed, and every search is logged.
Does the agent change documents?
No. It works read-only on the sources connected.
Do we have to learn a new tool to use it?
No. Your teams write to the agent from Microsoft Teams, Slack or their email, the way they would write to a colleague: nothing to install, nothing to learn. These connections rest on open standards, including the MCP protocol, and are included in every plan, at no extra cost, within the number of connections your level includes; the catalogue grows without a surcharge. You supervise the agent from a web dashboard.
Is our content 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 content 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 to connect and how fine-grained your access rights are, after a free audit then phases of design, integration and testing.
Related use cases
Other agents for your teams
Let's talk
Let's size up the potential in your company knowledge
15 minutes to frame your sources and your access rights — hosted in France, supervised, with no commitment.