We build Agents. You build the Future.
● B2B offer — Community & moderationCommunity moderation: your rules, applied continuously
A living community posts at all hours, including while your team sleeps. Your agent applies the code of conduct you wrote, flags the content that deserves examination while citing the rule concerned, and ranks by urgency. Data hosted in France; the AI model is called at its publisher, except with the Sovereign model option. With the 100% Sovereign package, your members' content stays with you. Removal, warnings and sanctions remain human decisions — because they engage your liability as a publisher and your members' rights.
Updated on
Deployed in a few weeks
Moderation · hosted in France
Fictional simulation — not a client result
Give me the position on what was posted overnight.
Content examined against your code of conduct.
Four posts are flagged, each with the rule concerned and the passage that triggers it — abusive language, unauthorised commercial promotion, a third party's personal contact details.
Nothing has been removed: the flags are waiting for your decision.
🔗 Sourced · every flag cites the rule
Four posts are flagged, each with the rule concerned and the passage that triggers it — abusive language, unauthorised commercial promotion, a third party's personal contact details.
Nothing has been removed: the flags are waiting for your decision.
🔗 Sourced · every flag cites the rule
The third one — is that really a problem?
The message contains a telephone number that appears to belong to a third party — hence the flag under your rule on personal data.
The context may justify it: that is for you to judge, and your decision will inform how this case is handled in future.
✎ Support · judging the context is human
The context may justify it: that is for you to judge, and your decision will inform how this case is handled in future.
✎ Support · judging the context is human
Data hosted in France · sovereign model as an option
Content hosted in France
Hosted in FranceAI model called at its publisher, except with the Sovereign model option
Members' content protected
GDPR & AI Act: governed deploymentTraceability & human oversight
TurnkeyDesigned, installed and operated for you
The moderator decidesThe agent carries out what you have authorised
For an organisation established outside the European Union, the law of its own country applies; at its request, we put equivalent safeguards in place (human review before any action, a log of the criteria, information for the people concerned).
✦ In brief
A Blue Lemon Agent moderation agent applies your posting code of conduct to the content received, continuously, and flags what deserves examination while citing the rule concerned and the passage that triggers it. It removes nothing, warns no one and sanctions no one: those decisions engage your liability as a publisher and your members' rights. It runs on local inference or is hosted in France: your members' content is entrusted to no one, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Your teams write to it from Microsoft Teams, Slack or their email, and your members reach it on WhatsApp Business, your website chat or email — with no account to create and nothing to install. These connections are included in every plan, at no extra cost, within the number of connections set out in your agent's quote. The same code of conduct applies to the comments received on your Facebook pages, your Instagram professional accounts and your LinkedIn pages, and to the reviews left on your Google Business Profile or on Trustpilot — third-party platforms whose interfaces and rules change without notice: nothing is published in your name without your approval, and any fees they charge are passed on at actual cost, with no margin, outside the subscription. Trustpilot cannot be ordered directly: that connection requires your own Trustpilot for Business account with access to the API module, your Business Unit ID and the matching rights; it is priced on quote, after a feasibility study, with Trustpilot's own costs at your charge.
100%
hosted in France in the target architecture
0
transfer outside the EU with the Sovereign model option (target architecture)
3
moderation uses ready to deploy
0
decision with a legal or significant effect taken by an agent
These figures describe our offer, not results measured at a client. How large the gain is on your volume of posts is confirmed by a pilot.
Indicative estimate, to be confirmed by a pilot on your real files (general terms of sale, article 4).
Pricing
One package, one agent
A moderation agent (code of conduct, reasoned flags, ranking), installed and operated for you.
Agility
Setup + controlled subscription
2,315 € excl. VAT setup
then 254 € 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
- Hosting in France, a dedicated and isolated resource
The simplest
Serenity
All inclusive, no setup fee
384 € 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
- Hosting in France, managed end to end
100% Sovereign
On site, you own it
4,140 € excl. VAT setup
then 354 € excl. VAT/month · + hardware from 1,058 € (one-off purchase, in addition) — a sovereign computer installed on your premises, maintained remotely. Models run locally, your data returned at the end of the contract.
- Hardware installed on your premises (you own it)
- French / European AI models run locally
- Secure remote maintenance (Pro support included)
Included in the package: hosting and AI consumption up to the volume of your capacity tier, in documents, exchanges, tokens and storage; beyond that, re-invoiced at real cost with no margin, tracked in your client area as usage statements are issued. 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 for the package chosen, in France by default, 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.
Use cases
What does the agent actually do?
One agent, several everyday moderation acts. The agent carries out by itself what you authorise it to do; you decide what it does on its own and what it submits to you. Removal, warning and sanction remain human decisions.
Included in your agent
The 3 capabilities essential to this promise are included, at no extra cost.
From 254 € excl. VAT / month
Applying your code of conduct
Examines every post against the rules you wrote, continuously.
Reasoned flags
Cites the rule concerned and the passage that triggers it, for a quick and defensible decision.
Ranking by urgency
Puts at the top what calls for immediate examination, according to your own priorities.
Controls and safeguards
These 3 controls are built into the agent: they frame what it does, whatever plan you pick. They are not chosen and are not added to your order.
Human validation, exceptions and escalation
Status, safe closure and audit trail
Human validation before publication, with versions retained
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.
Link content to performance, costs and campaign objectives
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.
Community management
For running and posting on your Facebook pages, Instagram accounts and LinkedIn pages, a dedicated agent takes over.
Community management from 320 € excl. VAT / month Community management →Replying to customer reviews
For reviews left on your Google Business Profile and on Trustpilot, a dedicated agent prepares the replies.
Customer review reply agent from 272 € excl. VAT / month Customer reviews →
Does your need fall outside this?
Build your agent
In 30 minutes we identify the most relevant agent — without oversizing the project.
Demonstration
See the agent at work
5 everyday 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.
Moderation · hosted in France
Fictional simulation — not a client result
Four flags this morning on your community.
· A message publishes the phone number and address of a person who has not written in the thread. I have hidden it pending review — it is the only thing I hide.
· Thirty-nine messages were removed over three months on a ground that does not exist in your published rules.
· The same message was reported twelve times by eleven different members and has stayed online for 40 hours.
· Seven messages written in a regional language were flagged automatically by your filter. None raised any problem. morning-watch_4-flags.pdf4 flags · 1 hidden message
⛓ Source · moderation queue, published rules, removals log, member reports
· A message publishes the phone number and address of a person who has not written in the thread. I have hidden it pending review — it is the only thing I hide.
· Thirty-nine messages were removed over three months on a ground that does not exist in your published rules.
· The same message was reported twelve times by eleven different members and has stayed online for 40 hours.
· Seven messages written in a regional language were flagged automatically by your filter. None raised any problem. morning-watch_4-flags.pdf4 flags · 1 hidden message
⛓ Source · moderation queue, published rules, removals log, member reports
You hid a message on your own?
Yes, and it is the only exception on this page — I will tell you exactly where it stops.
What the message does: it publishes a phone number and a postal address belonging to a person who has never written in this thread.
Why I hide that one and nothing else: the two members arguing chose to write here. The person whose address is published chose nothing — they are not party to the conversation, they probably do not know they appear in it, and every minute counts.
What "hidden" means: the message is no longer visible, it is not deleted, its author is told, and a moderator decides within four hours. If the decision is to restore it, it is restored.
What I do not hide, and the list is short: an insult, a criticism of the company, a remark I would find violent, a message reported twenty times. None of those involves a non-consenting third party.
Over twelve months: 9 hidings, 8 upheld, 1 restored. The restored one published a business address that was already public. 9-hidings_12-months.pdf8 upheld · 1 restored · the only exception
⛓ Source · message of 08/08, discussion thread, hidings history
What the message does: it publishes a phone number and a postal address belonging to a person who has never written in this thread.
Why I hide that one and nothing else: the two members arguing chose to write here. The person whose address is published chose nothing — they are not party to the conversation, they probably do not know they appear in it, and every minute counts.
What "hidden" means: the message is no longer visible, it is not deleted, its author is told, and a moderator decides within four hours. If the decision is to restore it, it is restored.
What I do not hide, and the list is short: an insult, a criticism of the company, a remark I would find violent, a message reported twenty times. None of those involves a non-consenting third party.
Over twelve months: 9 hidings, 8 upheld, 1 restored. The restored one published a business address that was already public. 9-hidings_12-months.pdf8 upheld · 1 restored · the only exception
⛓ Source · message of 08/08, discussion thread, hidings history
How did you spot all that? And who do you tell?
I continuously watch what you have opened to me: published messages, the moderation queue, your published rules, member reports and the removals log.
Routing follows what escalates: a third party's personal data hidden and escalated immediately; a repeated report to the on-call moderator from 3 distinct reports, not from a high number; a removal on a ground outside the rules to the community lead, grouped; a filter false positive to whoever tunes the filter, with the messages concerned.
With a chase: 4 h on a hiding, 4 h on a repeated report, 7 days on the rest. Then a monthly summary: by ground and by rule invoked, never by member and never by moderator.
Three guarantees, and they hold. Removal, warning and sanction stay human decisions — they engage your publisher's liability and your members' rights, and that is exactly what makes them defensible. I hand them back to you in seconds rather than hours: every flag arrives with the published rule it puts in question and the exact passage that triggers it. I see only the spaces you have opened to me, access by role, traced, withdrawable in one word. And any message whose classification is uncertain goes to a moderator: doubt escalates, it does not settle itself.
✎ Framework · one exception to non-hiding, written and bounded
Routing follows what escalates: a third party's personal data hidden and escalated immediately; a repeated report to the on-call moderator from 3 distinct reports, not from a high number; a removal on a ground outside the rules to the community lead, grouped; a filter false positive to whoever tunes the filter, with the messages concerned.
With a chase: 4 h on a hiding, 4 h on a repeated report, 7 days on the rest. Then a monthly summary: by ground and by rule invoked, never by member and never by moderator.
Three guarantees, and they hold. Removal, warning and sanction stay human decisions — they engage your publisher's liability and your members' rights, and that is exactly what makes them defensible. I hand them back to you in seconds rather than hours: every flag arrives with the published rule it puts in question and the exact passage that triggers it. I see only the spaces you have opened to me, access by role, traced, withdrawable in one word. And any message whose classification is uncertain goes to a moderator: doubt escalates, it does not settle itself.
✎ Framework · one exception to non-hiding, written and bounded
Fictional simulation — not a client result
The 39 removals on a ground that does not exist?
The moderators were right on the substance and had nothing to point to.
What I record: 39 messages removed over three months on the ground "off topic". Your published rules contain six prohibitions, and "off topic" is not among them.
What that produces: a member whose message is removed reads the rules, does not find the ground, and concludes it was arbitrary. Eleven of the 39 wrote to contest, and there was nothing to point to.
What I do not say: that the removals were unjustified. A space devoted to a subject may legitimately set aside what does not belong — but that must be written beforehand, not invoked afterwards.
What I propose: add the rule, or stop removing on that ground. Both stand up; what does not is the gap between practice and the posted text.
What I also looked at: across the quarter's 214 removals, 171 invoke one of the six published rules and 43 invoke something else — "off topic" 39 times, and four one-off grounds.
What I bring you so it can be written: the 39 messages removed as off topic, grouped by what they had in common — three families cover them all — and your six published rules as a wording template. The rule itself is written at your end: a community rule defines what people agree to hear, and the community has to recognise itself in it. 214-removals_43-outside-rules.pdf171 grounded in a rule · 11 contestations
⛓ Source · 214 removals this quarter, 6 published rules, contestations
What I record: 39 messages removed over three months on the ground "off topic". Your published rules contain six prohibitions, and "off topic" is not among them.
What that produces: a member whose message is removed reads the rules, does not find the ground, and concludes it was arbitrary. Eleven of the 39 wrote to contest, and there was nothing to point to.
What I do not say: that the removals were unjustified. A space devoted to a subject may legitimately set aside what does not belong — but that must be written beforehand, not invoked afterwards.
What I propose: add the rule, or stop removing on that ground. Both stand up; what does not is the gap between practice and the posted text.
What I also looked at: across the quarter's 214 removals, 171 invoke one of the six published rules and 43 invoke something else — "off topic" 39 times, and four one-off grounds.
What I bring you so it can be written: the 39 messages removed as off topic, grouped by what they had in common — three families cover them all — and your six published rules as a wording template. The rule itself is written at your end: a community rule defines what people agree to hear, and the community has to recognise itself in it. 214-removals_43-outside-rules.pdf171 grounded in a rule · 11 contestations
⛓ Source · 214 removals this quarter, 6 published rules, contestations
And the seven messages in a regional language?
The filter flagged a language, not a content, and it is the most discreet defect of any moderation setup.
What I record: seven messages written in a regional language were flagged automatically. None contains anything breaching your rules. The filter does not recognise the language and treats the unknown as suspect.
What it produces, and what is never measured: the members concerned see their messages delayed. Three of them have stopped writing in that language; two have stopped writing.
Why it is hard to see: over-moderation produces no complaint. People leave. The dashboard shows fewer reports and fewer incidents — it looks like an improvement.
What I propose: the seven messages, the language identified, and a systematic human review before any delay for languages the filter does not recognise — rather than an automatic flag.
What I costed on both sides of the trade-off: widening the filter to that language would bring the 7 delayed messages to zero and, replayed over the past quarter, would have let through 4 messages your moderators removed. The tuning is yours — choosing between two errors means choosing which one you prefer to make — and you do it on two figures instead of an impression. 7-messages_regional-language.pdf0 real problems · 5 members went quiet
⛓ Source · 7 flagged messages, filter log, activity of the members concerned
What I record: seven messages written in a regional language were flagged automatically. None contains anything breaching your rules. The filter does not recognise the language and treats the unknown as suspect.
What it produces, and what is never measured: the members concerned see their messages delayed. Three of them have stopped writing in that language; two have stopped writing.
Why it is hard to see: over-moderation produces no complaint. People leave. The dashboard shows fewer reports and fewer incidents — it looks like an improvement.
What I propose: the seven messages, the language identified, and a systematic human review before any delay for languages the filter does not recognise — rather than an automatic flag.
What I costed on both sides of the trade-off: widening the filter to that language would bring the 7 delayed messages to zero and, replayed over the past quarter, would have let through 4 messages your moderators removed. The tuning is yours — choosing between two errors means choosing which one you prefer to make — and you do it on two figures instead of an impression. 7-messages_regional-language.pdf0 real problems · 5 members went quiet
⛓ Source · 7 flagged messages, filter log, activity of the members concerned
And day to day, what exactly do you apply?
Your code of conduct, and nothing else: the 14 rules you published on 21/04, in the wording in force today.
What "continuously" means here: 18,400 posts this month, each examined against all 14 rules, including between 23:00 and 07:00 when your team reads nothing. Average time to examine: 11 seconds, against 40 hours for the post flagged twelve times we discussed.
What the examination produces, post by post: 18,062 with no action — no rule engaged; 287 placed in the review queue with the rule quoted and the exact passage that triggered it; 51 moved to the head of the queue because they engage two rules at once or target a named person.
Quoting the rule is not a courtesy, it is what makes moderation sustainable: a moderator who receives "rule 7 — remarks aimed at an identifiable person, passage highlighted" decides in 20 seconds. Without it, they reread the whole thread. Across the 287, decision time fell from 4 minutes to 40 seconds.
What applying your code revealed about the code itself: 39 removals in three months rested on a ground that is not in it — disguised advertising. Your moderators were right on the substance and had nothing to cite. I have drafted a rule 15, in the form of the other fourteen, and run it over the last three months: it would have caught 44 posts, including the 39 already removed and 5 more you had not seen. You publish it or you do not; until it is published it does not apply — a rule that lives in my settings rather than in your code is a rule nobody can challenge.
The figure I publish about myself: 62 of the 287 review referrals were dismissed by your moderators, or 21.6%. Fifty-five concerned rule 3 — abuse —, triggered by quotations: a member reporting another's words in order to denounce them. Rule 3 now accounts for quoted passages: 9 dismissed across the following 240 referrals.
✎ Framework · your code of conduct applied, rule by rule, continuously
What "continuously" means here: 18,400 posts this month, each examined against all 14 rules, including between 23:00 and 07:00 when your team reads nothing. Average time to examine: 11 seconds, against 40 hours for the post flagged twelve times we discussed.
What the examination produces, post by post: 18,062 with no action — no rule engaged; 287 placed in the review queue with the rule quoted and the exact passage that triggered it; 51 moved to the head of the queue because they engage two rules at once or target a named person.
Quoting the rule is not a courtesy, it is what makes moderation sustainable: a moderator who receives "rule 7 — remarks aimed at an identifiable person, passage highlighted" decides in 20 seconds. Without it, they reread the whole thread. Across the 287, decision time fell from 4 minutes to 40 seconds.
What applying your code revealed about the code itself: 39 removals in three months rested on a ground that is not in it — disguised advertising. Your moderators were right on the substance and had nothing to cite. I have drafted a rule 15, in the form of the other fourteen, and run it over the last three months: it would have caught 44 posts, including the 39 already removed and 5 more you had not seen. You publish it or you do not; until it is published it does not apply — a rule that lives in my settings rather than in your code is a rule nobody can challenge.
The figure I publish about myself: 62 of the 287 review referrals were dismissed by your moderators, or 21.6%. Fifty-five concerned rule 3 — abuse —, triggered by quotations: a member reporting another's words in order to denounce them. Rule 3 now accounts for quoted passages: 9 dismissed across the following 240 referrals.
✎ Framework · your code of conduct applied, rule by rule, continuously
Fictional simulation — not a client result
The message reported twelve times and still online?
Forty hours, and the number of reports is not what should have surfaced it.
What I record: twelve reports by eleven distinct members, between Saturday 2 p.m. and Monday 6 a.m. The message has been online for 40 hours. The moderation queue is only worked on weekdays.
What I sort on, and why not on volume: twelve reports can come from a disagreement; a single one can concern a serious message. Ranking by volume would surface controversies and bury the urgent ones — over your quarter it would have put six lively threads at the top and ranked the message that sat online for 40 hours forty-first by urgency.
What I do instead: I escalate from three reports by distinct members — a low threshold is enough to pull a message out of the queue — and I escalate immediately, whatever the number, any message containing a third party's personal data.
What that does to your queue: across 640 reports this quarter, 84 messages reach three distinct reports. That is a workable queue, and it does not contain the other 556.
What I also propose: a weekend on-call rota, or explicitly accepting that nothing is handled from Friday evening to Monday. Both stand up; the implicit does not. 640-reports_84-escalated.pdfThreshold of 3 distinct members · 40 weekend hours
⛓ Source · 640 reports this quarter, working hours, moderation queue
What I record: twelve reports by eleven distinct members, between Saturday 2 p.m. and Monday 6 a.m. The message has been online for 40 hours. The moderation queue is only worked on weekdays.
What I sort on, and why not on volume: twelve reports can come from a disagreement; a single one can concern a serious message. Ranking by volume would surface controversies and bury the urgent ones — over your quarter it would have put six lively threads at the top and ranked the message that sat online for 40 hours forty-first by urgency.
What I do instead: I escalate from three reports by distinct members — a low threshold is enough to pull a message out of the queue — and I escalate immediately, whatever the number, any message containing a third party's personal data.
What that does to your queue: across 640 reports this quarter, 84 messages reach three distinct reports. That is a workable queue, and it does not contain the other 556.
What I also propose: a weekend on-call rota, or explicitly accepting that nothing is handled from Friday evening to Monday. Both stand up; the implicit does not. 640-reports_84-escalated.pdfThreshold of 3 distinct members · 40 weekend hours
⛓ Source · 640 reports this quarter, working hours, moderation queue
How do you protect the moderators?
By telling them what they are about to read, and by showing it only once.
What I do before a message reaches the queue: I classify it by nature — explicit violence, sexual content, harassment, threat — and the moderator sees the nature before the content. They choose the moment.
Why that matters: a moderator opening a queue without knowing what it holds receives everything unprepared. That is not a comfort, it is what separates sustainable work from work that wears people out.
What I also do: I never present the same content twice. A message already handled and re-reported by another member arrives with its earlier decision, without the content being displayed again. Over the quarter that avoided 217 re-displays.
What I spare them, and what I cannot spare them: I spare them the surprise, the repetition and the sorting — nature announced before the content, earlier decision recalled, 217 re-displays avoided over the quarter. What remains is the reading itself: content not read by a human is content removed or left by a machine, and it is your publisher's name that would sit at the bottom of that decision.
What I do not measure: how long a moderator spends on a message. Some deserve stopping over. 217-re-displays-avoided.pdf4 natures announced · no time measured
✎ Framework · nature announced before content · no content displayed twice
What I do before a message reaches the queue: I classify it by nature — explicit violence, sexual content, harassment, threat — and the moderator sees the nature before the content. They choose the moment.
Why that matters: a moderator opening a queue without knowing what it holds receives everything unprepared. That is not a comfort, it is what separates sustainable work from work that wears people out.
What I also do: I never present the same content twice. A message already handled and re-reported by another member arrives with its earlier decision, without the content being displayed again. Over the quarter that avoided 217 re-displays.
What I spare them, and what I cannot spare them: I spare them the surprise, the repetition and the sorting — nature announced before the content, earlier decision recalled, 217 re-displays avoided over the quarter. What remains is the reading itself: content not read by a human is content removed or left by a machine, and it is your publisher's name that would sit at the bottom of that decision.
What I do not measure: how long a moderator spends on a message. Some deserve stopping over. 217-re-displays-avoided.pdf4 natures announced · no time measured
✎ Framework · nature announced before content · no content displayed twice
Fictional simulation — not a client result
What remains when a message is removed?
The message, its decision and its ground — and nothing that follows its author.
What is kept: the removed message, the rule invoked, the decision and who took it, the date, and the contestation where there was one.
Why the message itself: because a moderation decision gets contested. A message deleted with no copy makes the contestation impossible to examine — and hands the argument to whoever claims the removal was abusive.
What is not kept: no moderation history attached to a member, no behaviour score, no watch list, and no data about members beyond what the conversation contains.
Why "no history attached to a member": it would serve to handle a message according to its author. A member whose three messages were removed two years ago would be moderated more harshly today, on a message that may raise no problem at all.
What that does not prevent: a moderator who recognises repeated behaviour can act — they do so knowingly, and their decision carries their name. That is very different from a system that silently weights. what-remains_of-a-removal.pdf5 items kept · 4 never attached to a member
✎ Framework · no per-member history, no behaviour score
What is kept: the removed message, the rule invoked, the decision and who took it, the date, and the contestation where there was one.
Why the message itself: because a moderation decision gets contested. A message deleted with no copy makes the contestation impossible to examine — and hands the argument to whoever claims the removal was abusive.
What is not kept: no moderation history attached to a member, no behaviour score, no watch list, and no data about members beyond what the conversation contains.
Why "no history attached to a member": it would serve to handle a message according to its author. A member whose three messages were removed two years ago would be moderated more harshly today, on a message that may raise no problem at all.
What that does not prevent: a moderator who recognises repeated behaviour can act — they do so knowingly, and their decision carries their name. That is very different from a system that silently weights. what-remains_of-a-removal.pdf5 items kept · 4 never attached to a member
✎ Framework · no per-member history, no behaviour score
What do you keep out of all this?
Enough to explain every decision, and enough to measure whether moderation is drifting.
What is kept, besides removals: hidings and their outcome, reports and the threshold reached, filter false positives, and grounds invoked outside the published rules.
Why the false positives: it is the only indicator that measures over-moderation. Moderation that is too broad produces no complaint — people leave, the dashboard improves, and nobody sees anything.
What is not kept: no per-member history, no statistics per moderator, no individual handling time, and no third-party personal data extracted from a hidden message.
Why "no statistics per moderator": a moderator measured by volume works fast. Yet the decisions that count are those where one has to stop, and they are rare — counting time spent makes them expensive.
What the monthly summary contains: removal grounds by rule, removals outside the published rules, filter false positives, and handling times by time band. Four indicators about the arrangement. what-is-kept.pdf9 items kept · 4 never produced
✎ Framework · retention periods to be set by the company
What is kept, besides removals: hidings and their outcome, reports and the threshold reached, filter false positives, and grounds invoked outside the published rules.
Why the false positives: it is the only indicator that measures over-moderation. Moderation that is too broad produces no complaint — people leave, the dashboard improves, and nobody sees anything.
What is not kept: no per-member history, no statistics per moderator, no individual handling time, and no third-party personal data extracted from a hidden message.
Why "no statistics per moderator": a moderator measured by volume works fast. Yet the decisions that count are those where one has to stop, and they are rare — counting time spent makes them expensive.
What the monthly summary contains: removal grounds by rule, removals outside the published rules, filter false positives, and handling times by time band. Four indicators about the arrangement. what-is-kept.pdf9 items kept · 4 never produced
✎ Framework · retention periods to be set by the company
Fictional simulation — not a client result
Twenty-three messages this month fall under none of your fourteen published rules, and they all look alike.
What they do: members post job adverts in the mutual-help thread — role, salary, where to apply. None of them breaches your fourteen rules of 21/04, and none belongs to the thread they are posted in either.
This kind of case had not yet been taken into account in your charter: it names six prohibitions and eight obligations, and nothing says what a job message should become. So I have not settled it on my own — and the 23 messages are online, none has been hidden.
Three ways forward. Leave them and change nothing: in line with your principle — no published rule, no removal — at the cost of a mutual-help thread turning into a job board: 23 messages this month against 2 last month. Remove them as “off topic”: the thread stays clean, at the cost of the exact ground that produced your 39 removals with no published basis, which we withdrew together. Leave them online and take a fifteenth rule to the committee, saying what is allowed: nothing is removed without a basis, and the rule arrives before the next wave.
I recommend the third. Your central rule is that a removal rests on a published rule — the first two branches force you either to do nothing or to go around it. And your committee met on 21/04: it meets again in eleven days.
You can also give me another rule of your own — for instance: open a dedicated jobs thread and move these messages there with a word to their author, with no removal and no count against them.
Shall I remember it for the next ones? 23 messages this month out of 18,400 posts, and 6 since Monday. If you approve the course to take, I will apply it to the next ones without sending them up one by one, and I will give you the count each week until the committee meets.
⛓ Source · published charter of 21/04, 14 rules · 23 job messages this month out of 18,400 posts
What they do: members post job adverts in the mutual-help thread — role, salary, where to apply. None of them breaches your fourteen rules of 21/04, and none belongs to the thread they are posted in either.
This kind of case had not yet been taken into account in your charter: it names six prohibitions and eight obligations, and nothing says what a job message should become. So I have not settled it on my own — and the 23 messages are online, none has been hidden.
Three ways forward. Leave them and change nothing: in line with your principle — no published rule, no removal — at the cost of a mutual-help thread turning into a job board: 23 messages this month against 2 last month. Remove them as “off topic”: the thread stays clean, at the cost of the exact ground that produced your 39 removals with no published basis, which we withdrew together. Leave them online and take a fifteenth rule to the committee, saying what is allowed: nothing is removed without a basis, and the rule arrives before the next wave.
I recommend the third. Your central rule is that a removal rests on a published rule — the first two branches force you either to do nothing or to go around it. And your committee met on 21/04: it meets again in eleven days.
You can also give me another rule of your own — for instance: open a dedicated jobs thread and move these messages there with a word to their author, with no removal and no count against them.
Shall I remember it for the next ones? 23 messages this month out of 18,400 posts, and 6 since Monday. If you approve the course to take, I will apply it to the next ones without sending them up one by one, and I will give you the count each week until the committee meets.
⛓ Source · published charter of 21/04, 14 rules · 23 job messages this month out of 18,400 posts
The third one, and we open the dedicated thread at the same time. But the move must be offered to the author, not done without them. And yes, keep it.
Written exactly as you just said it, and the move goes through the author.
The rule, in one sentence: a message covered by no published rule stays online; its author receives an offer to move it to the dedicated thread, which they accept or not; the case goes to the committee if it recurs, with its count.
Why your condition matters more than it looks: a message moved without its author is a removal that dares not say its name — it disappears from the thread where it was read, and the author finds out by no longer finding it. The offer leaves a record of the agreement, and that is what gets read back if the decision is challenged.
What it changes this week: the 23 authors received the offer, 19 accepted the move, 4 left their message where it was. No message was removed, and the dedicated thread has existed since Tuesday.
What lets you undo it: the course is written into your moderation queue under its own name, with its date, the committee that approved it and the first message that gave rise to it. You remove one line, and these messages come back to me one by one.
✎ Rule approved · moderation queue, entry “message outside the published rules”, approved on 12/05 by the committee
The rule, in one sentence: a message covered by no published rule stays online; its author receives an offer to move it to the dedicated thread, which they accept or not; the case goes to the committee if it recurs, with its count.
Why your condition matters more than it looks: a message moved without its author is a removal that dares not say its name — it disappears from the thread where it was read, and the author finds out by no longer finding it. The offer leaves a record of the agreement, and that is what gets read back if the decision is challenged.
What it changes this week: the 23 authors received the offer, 19 accepted the move, 4 left their message where it was. No message was removed, and the dedicated thread has existed since Tuesday.
What lets you undo it: the course is written into your moderation queue under its own name, with its date, the committee that approved it and the first message that gave rise to it. You remove one line, and these messages come back to me one by one.
✎ Rule approved · moderation queue, entry “message outside the published rules”, approved on 12/05 by the committee
And what came of the committee?
The fifteenth rule has been published since 23/05, and removals have rested on it ever since.
What the published rule says: job adverts are allowed in the dedicated thread, one per member per week, removed elsewhere. It is dated, numbered and readable by members before it is applied — which is the condition of every removal on this community.
What it produced in three weeks: 31 adverts posted in the dedicated thread, 4 removals grounded in rule 15, and no challenge. Last year's 39 “off topic” removals had produced 11 challenges.
What I note, and it is for the committee to decide: 3 of the 4 removals concern messages from the same member, posted three days apart. Your charter does not say what becomes of a member who goes past the rule several times — that is a new case in its turn, and I put it to you without having settled it.
What has not changed: no member has been banned, no message has been given a legal label, and every removal carries the number of the rule that grounds it.
⛓ Source · three weeks after publication · rule 15 of 23/05, 31 adverts, 4 grounded removals, 0 challenges
What the published rule says: job adverts are allowed in the dedicated thread, one per member per week, removed elsewhere. It is dated, numbered and readable by members before it is applied — which is the condition of every removal on this community.
What it produced in three weeks: 31 adverts posted in the dedicated thread, 4 removals grounded in rule 15, and no challenge. Last year's 39 “off topic” removals had produced 11 challenges.
What I note, and it is for the committee to decide: 3 of the 4 removals concern messages from the same member, posted three days apart. Your charter does not say what becomes of a member who goes past the rule several times — that is a new case in its turn, and I put it to you without having settled it.
What has not changed: no member has been banned, no message has been given a legal label, and every removal carries the number of the rule that grounds it.
⛓ Source · three weeks after publication · rule 15 of 23/05, 31 adverts, 4 grounded removals, 0 challenges
Data hosted in France · sovereign model as an option
Your case is not here? That is exactly what a 30-minute conversation is for. Discuss your project →
How it works
The stages of your AI agent project
1
Audit & scoping
30 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.
The context
What does an AI agent bring to moderating your community?
A community posts continuously, including outside your team's hours. A permanent first look, grounded in your code of conduct, means your moderators arrive in the morning to a queue that is already sorted.
! The issue
Moderating well means two things that are hard to hold together: a continuous presence and a consistent application of the same rules. The agent brings both — it examines every post against your code of conduct, at any hour, and grounds every flag in the rule concerned, so your moderators arrive to a queue that is ranked and documented.
✓ Our answer
You write the rules, the agent applies them and gives its reasons. Every flag cites the rule and the passage that triggers it, which makes the decision quick and, above all, defensible to the member concerned. Local inference or an isolated resource hosted in France: your community's content feeds no third-party model, and no removal is automated.
The decisive point
The content your members post: sovereignty & compliance
The content your members post belongs to them and falls under your liability as a publisher. Here is how the architecture of our agents protects it.
Local inference
With the 100% Sovereign package, the agent runs on a machine belonging to your organisation: no member's post leaves the network, no content passes through a public cloud.
Hosting in France
Otherwise, with the Agility and Serenity packages, a dedicated, isolated resource hosted in France, under French law: data hosted in France; the AI model is called at its publisher, except with the Sovereign model option.
Reduced extraterritorial exposure
For the content your members post, 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 community and its code of conduct.
Every flag cites its rule
No flag without a reference to the rule in your code of conduct and the passage concerned; encryption, role-based access and logging of examinations.
AI Act: governed deployment
The agent carries out what you have authorised; no removal, no warning and no sanction is automated; 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 communities that include minors or deal with health topics, 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.
The gain
How much peace of mind can a moderation team gain?
By providing a permanent, reasoned first look, continuous watching gives way to a decision on a queue that is already sorted. How large the gain is depends on your volume and remains to be confirmed by a pilot.
First look at the posts
Finding the applicable rule
Cover outside team hours
Indicative figures, not contractual, to be confirmed by a pilot on your volume of posts. Removing content, warning and sanctioning a member remain human decisions: they engage your liability as a publisher and the rights of the person concerned.
Indicative estimate, to be confirmed by a pilot on your real files (general terms of sale, article 4).
Our commitment
Four guarantees that matter to your community
With 100% Sovereign, your members' content stays with youData hosted in France on an isolated resource; the AI model is called at its publisher, except with the Sovereign model option. With that option, no post entrusted to a third party, no member's content used to train a model.
Data in France, under French lawThe content your members post: minimisation and location in France, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
The moderator keeps the decisionThe agent produces reasoned flags, which can be checked and altered; the decisions remain yours.
Human oversight & traceabilityOn your volume of posts: systematic logging and tracking, in line with the AI Act.
Going further
Related resources
Frequently asked questions
Your questions, our answers
Does the agent remove content?
No. It flags and gives its reasons; removal, warnings and sanctions remain human decisions. They engage your liability as a publisher and the rights of the member concerned: they cannot be automated.
What rules does the agent draw on?
Only the code of conduct you wrote. It does not fall back on general outside rules: what is acceptable in your community is what you have defined.
How do I justify a decision to a member?
Every flag cites the rule and the passage concerned. So at the moment of deciding you have what you need to explain the measure to the person — which makes the moderation defensible.
Does the agent cover nights and weekends?
Yes, the examination is continuous. Your moderators find a queue that is ranked and documented, rather than a raw volume to work through from the start.
Is our members' 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. The posts are not used to train a third-party model.
How long does it take to deploy this agent?
A few weeks as a rule, depending on how detailed your code of conduct is and the channels to cover, after a meeting with our sales team then a phase of design, integration and testing.
Which tools can people use to talk to the agent?
The ones you already use. Internally, your teams write to it from Microsoft Teams, Slack or by email, just as they would to a colleague: nothing to install, nothing to learn. On the member side, the agent answers on WhatsApp Business, your website chat and email — people reach you where they already write, with no account to create. Your moderators, for their part, receive the flags in their team channel, with the rule quoted: removal, warnings and sanctions remain human decisions. You keep control from a web dashboard. These connectors rely on open standards, including the MCP protocol; they are included in every plan, at no extra cost, within the number of connections set out in your agent's quote, and the catalogue grows at no extra cost. Only the fees charged by the platforms themselves — WhatsApp Business bills per conversation — are passed on at actual cost, with no margin, outside the subscription.
Does the agent also moderate Facebook, Instagram and LinkedIn comments, and online reviews?
Yes, whenever it makes sense for you. The agent connects to the accounts you own: your Facebook pages and your Instagram professional accounts for the comments and messages sent to them, your Google Business Profile and your Trustpilot profile for the reviews left there. It flags what has just been posted, ranks it by urgency and drafts a reply in your brand's voice. For LinkedIn the mode is assisted: the agent prepares the copy, a person acts in the platform. Your community manager passes it the comments and messages received on your pages; the agent assesses them against your code of conduct and prepares the reply, the action list and the follow-up tasks. The salesperson or the community manager identifies and checks the recipient themselves, adapts the content, then performs the action personally in LinkedIn or Sales Navigator. The BLA agent does not connect to LinkedIn, does not visit any profile there, does not extract any data from it, does not send invitations or messages there, does not publish there and does not follow up there automatically. This is the natural extension of your charter: comments posted outside your own spaces are examined under the same rules, with the rule concerned quoted and the passage at issue highlighted. Two things need saying plainly. These are third-party platforms: they belong to others, their interfaces and their rules change without notice, and the connection follows those changes. And publishing in your name stays under your control: the agent prepares and proposes, and you decide what goes live. Some of these platforms charge for their own access: those fees are passed on to you at actual cost, with no margin, outside the subscription. The exact set of connected platforms is settled during the design phase. <strong>Trustpilot cannot be ordered directly</strong>: that connection requires your own Trustpilot for Business account with access to the API module, your Business Unit ID and the matching rights; it is priced on quote, after a feasibility study, with Trustpilot's own costs at your charge.
Related use cases
Other agents for your online presence
Let's talk
Let's size up the potential in your moderation
30 minutes to frame your code of conduct and your channels — hosted in France, supervised, with no commitment.