Agent for replying to customer reviews: brand tone and moderation
A review left unanswered for days, a generic reply that reassures no one, a negative review published without anyone seeing it in time: replying to reviews takes steadiness and tact. Your agent reads the reviews received, drafts a reply in your brand's tone and flags those that call for moderation before publication. Hosted in France — on local inference or an isolated resource — your exchanges with your customers stay with you.
Updated on
Reply proposed: acknowledgement of the inconvenience, thanks for the feedback, an invitation to contact the department concerned to discuss it — with no generic apology and no figures promised.
I have published nothing: a negative review deserves your reading before it goes out.
✎ Action · draft ready to read over — you approve before publication
All are placed as drafts in the moderation queue — none is published without your go-ahead, not even the simplest.
⛓ Source · this week's reviews + the brand's reply guidelines
A Blue Lemon Agent review reply agent reads the reviews received across your platforms, drafts a reply in your brand's tone and flags those that call for moderation — negative reviews, sensitive remarks, criticism of a department. Every draft is placed in a queue, never published on its own. It runs on local inference or is hosted in France: your exchanges with your customers are never exposed to a foreign service, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Live in one to two weeks, Your teams write to it from Microsoft Teams, Slack or their email, and your customers 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 your level includes. The reviews followed are those on your Google Business Profile and on Trustpilot, along with the comments left on your Facebook pages and your Instagram professional accounts. These platforms belong to third parties and their interfaces change without notice; publishing in your name stays under your control, and any access 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.
These figures describe our offer, not results measured at a client: how large the gain is on your volume of reviews received is confirmed by a pilot.
What does an AI agent bring to your online customer relations?
Every review deserves a personalised reply, consistent with the brand, and negative reviews call for tact. But entrusting your customer exchanges to a consumer tool means losing control of them.
! The issue
Customer reviews arrive continuously, across several platforms, and each deserves a reply that is neither generic nor late. Yet most consumer AI tools amount to entrusting reviews, customer exchanges and reply guidelines to a third party, often hosted outside Europe and subject to the Cloud Act.
✓ Our answer
AI is only of interest to customer relations if it is sovereign and faithful to your tone. Local inference or an isolated resource hosted in France, systematic human oversight, no reply published without approval: time gained on drafting is never paid for in a dented reputation. The aim is not to replace your judgement, but to prepare each reply so that all you have to do is read it over.
Your reviews and your customer exchanges: sovereignty & compliance
A customer review sometimes contains personal or sensitive information. Here is how the architecture of our agents protects it, review by review.
Local inference
The agent can run on a machine belonging to the company: no review leaves the network, nothing passes through a public cloud.
Hosting in France
Otherwise, a dedicated and isolated resource, hosted in France under French law — your exchanges: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
For your customer exchanges, 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 immunityeven when hosted in Europe by an American provider.
A dedicated moderation queue
Every sensitive review is set apart in a queue to approve, with no mixing with replies already published.
Brand tone preserved
Your reply guidelines are not used to train a third-party model; encryption, role-based access and logging.
AI Act: governed deployment
The agent is strictly in support; no reply is published 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.
See the agent at work
4 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.
· Seven reviews in three weeks describe the same incident — a wait at delivery, always the same slot. That is not an image problem, it is an operations problem.
· A one-star review mentions a verifiable fact: an order placed on 14/07 and delivered on the 29th. Your system confirms both dates.
· Eleven positive reviews arrived in four days, all written in a very similar form. I conclude nothing; I show you.
· Forty-one reviews have never received a reply, nine of them negative and more than six months old. morning-watch_4-flags.pdf4 flags · 1 operations problem
⛓ Source · review platforms opened to the agent, order log, reply history
What I record: seven reviews in three weeks mention a wait at delivery. Six of seven concern the Friday afternoon slot, and five the same delivery point.
What I cross-checked: your delivery data. The Friday afternoon slot shows an average delay of 47 minutes since 12 June, against 8 minutes on the other slots.
What that means: the reviews do not describe a perception, they describe a measurable fact you already hold in your own data. Seven customers took the time to write what your dashboard was already showing.
What I propose: the flag goes to operations, not to communications. Replying to the seven reviews without fixing the slot would produce seven courteous replies and an eighth review.
What I have written for the seven, and what it is waiting on: the reply is drafted, one per review, and it names the Friday slot and the date it will be fixed. It is waiting on that date, not on a proofread: a templated reply about a real, unresolved problem shows, and it costs more than silence. The day operations give the date, all seven go out in a minute. 7-reviews_1-slot.pdf47 min average delay · flagged to operations
⛓ Source · 7 reviews, delivery data since 12/06, slots
Routing follows the nature of the problem: a review describing an operational fact goes to the department concerned, not to communications — that is the most important rule on this page; an isolated negative review to the site manager, within 24 hours; a review mentioning a verifiable fact with the verification already done; an abnormally homogeneous run of reviews to management, with no characterisation.
With a chase: 24 h on an unanswered negative review, 7 days on the rest. Then a monthly summary: by reason for dissatisfaction and by site, never by an employee named in a review.
What this read has already given you: one operational cause pinned down — 47 minutes of average delay on Friday afternoons since 12 June, against 8 on the other slots — and 41 unanswered reviews picked back up, 9 of them negative and over six months old, each with its reply written to your tone of voice.
From tomorrow: fix the slot and that is seven reviews that will not be written, and an eighth one too. Publishing stays your voice, and I hand it back to you in minutes: every reply rests on the fact already verified in your own data, all that is left is to read and approve. The day you want me to publish myself — thanking positive reviews, for instance — you give me the mandate: written, bounded to the cases you set, dated, withdrawn on a word. Your reviews and your operations data never leave your walls, opened platform by platform and logged.
The next step is ready: the 41 replies are queued, the 9 old ones first. Give the green light to reread, and the Friday delay goes to operations this morning.
✎ Framework · no publication, no review written, no removal requested
What the review says: order placed on 14/07, delivered on the 29th. Fifteen days for a product advertised at five.
What I checked: both your dates. Both are accurate. The order was held for nine days on a stock shortage that was never notified to the customer.
What the reply contains: acknowledgement of the delay, its real cause — a shortage, not negligence —, what has changed since, and an offer of direct contact. It contains no generic apology, no "we are sorry you felt that way", and no commercial gesture.
Why no commercial gesture in a public reply: because it is public. A gesture announced under a review becomes a price, and the next person will ask for it by quoting it.
Who the reply is written for: not the author — they have often moved on. It is written for the people who will read the review in six months, and who will judge the company on how it replies far more than on the incident.
I do not publish it. reply_one-star-review.pdf2 dates verified · real cause · no gesture
⛓ Source · review of 30/07, order of 14/07, stock log
What I do on every negative review: I separate facts verifiable in your data — a date, an amount, a reference — from what is a matter of appreciation — the welcome, perceived quality, value for money.
Across the quarter's 34 negative reviews: 19 contain at least one verifiable fact. Of those 19, 14 are accurate, 3 are inaccurate, and 2 concern an order I cannot find in your records.
What I prepare for the 3 inaccurate ones: the correction, resting on the line of your data that establishes it — date, amount or reference, with its source. A removal request is not the right tool: on an inaccurate review it rarely succeeds, sometimes gets noticed, and always gets retold — and the inaccuracy usually concerns a detail, not the experience.
What I propose instead: a reply that corrects the fact without contradicting the feeling. "Our records show delivery on the 22nd rather than the 29th; the delay is still too long, and here is why." Correcting a fact and acknowledging the problem are not contradictory — it is the only reply that stands up.
The 2 untraceable orders are flagged separately, with no characterisation. 34-reviews_19-verifiable-facts.pdf14 accurate · 3 inaccurate · 2 untraceable
⛓ Source · 34 negative reviews this quarter, order data
What I record: eleven positive reviews in four days, against an average of two a week over the previous six months. Nine of eleven run between 22 and 31 words. Eight use the same structure: an adjective, a named service, a recommendation.
What I do not conclude: that they are fake. A legitimate review-request campaign produces exactly that profile — same period, same form, same questions put to customers. So does a commercial operation.
What I ask you: did you launch anything around 2 August? If so, it is explained and there is nothing to do. If not, it is information management should have.
What I guarantee you, whatever the answer: no review written by me, none solicited in exchange for a benefit, no competitor's review reported as suspect. All three are done, they get noticed, and they cost far more than they return — your reputation is worth more than eleven extra reviews.
What I provide: the eleven, their timestamps, their length, and the curve of the previous six months to place the anomaly. 11-reviews_homogeneous-profile.pdf2/week usual · 11 in 4 days · no conclusion
⛓ Source · 11 reviews from 02 to 05/08, 6-month history
What I did: sorted the quarter's 214 reviews by what they describe, not by their rating.
· 81 concern the product itself — that is what you already measure.
· 64 concern the lead time.
· 41 concern a precise moment in the journey: ordering, delivery, returns.
· 28 concern an exchange with a person.
The group of 41 is the one you are missing: 29 of the 41 name the same step — the handover from order to preparation, where the customer receives nothing for an average of three days.
What that is worth: twenty-nine customers described a journey defect free of charge, in precise words. No study costs less.
What I hand back instead of a score: the 29 reviews, the step they name, and the three days of silence they describe. An overall rating would make those 29 disappear into an average that rises or falls without saying why — a breakdown by step, on the other hand, can be handed to whoever can act on it. 214-reviews_by-subject.pdf41 on the journey · 29 on one step
⛓ Source · 214 reviews this quarter, sorted by subject
Across the quarter's 1,240 reviews, 96 came out for priority review, under three grounds, each named on the review sheet:
· Strongly negative tone — 54 reviews. A low score is not enough: what brings them out is a one- or two-star review describing a verifiable, dated fact. A poor score with no fact does not call for urgency, it calls for a reply.
· Sensitive matter — 19 reviews. Health, someone's safety, a third party's personal details published in the text, remarks liable to target an individual. Those 19 go to the head of the queue and to a named reviewer, not into the common pile.
· A service put in question — 23 reviews. The review names a service, a site, a procedure. Those count twice over: they call for a reply, and they often describe an operational fact — seven of them concerned the same delivery wait.
The time detection won back: the one-star review flagged twelve times had waited 40 hours. Today, a review in priority review is on the reviewer's desk in 3 minutes, with the ground it came out on, the passage highlighted, and the operational data that confirms or contradicts it.
What moderation does not do, and this is the heart of the page: no review is hidden, no removal is requested, no reply is published. Detection speeds up human review, it does not replace it.
The figure that does not suit me: 18 of the 96 referrals were needless, or 18.7% — all negative tone triggered by irony: "flawless service, I loved the wait". Detection now reads the fact described before the tone used: 3 needless referrals across the following 71. And the opposite error is the one that counts: among the 1,144 reviews not referred, your team caught 2 that should have been, both on sensitive matter.
✎ Framework · detection and moderation — priority review, never publication
What I record: 28 reviews this quarter mention an exchange with a person, 17 of which name or identifiably describe them. 12 are positive, 5 negative.
What I do: I pass the review to the site manager, with the name if the review carries it — because a manager must be able to deal with a real incident.
What I do count, and it is the figure you can act on: mentions by reason and by time of day. Of the 28, 19 fall in the 12-2 pm slot and 11 carry the same complaint about waiting at the counter. What is not counted: mentions per employee, a ranking, an attachment to a person's file.
Why, and this is not a formality: an employee accumulating negative mentions is almost always the one handling the difficult cases. A count would designate them as the problem when they are the place where the problem arrives.
The 5 negative reviews naming somebody: I pass them one at a time, with the order context. Three concern a refusal that complied with your own rules.
What I propose: that reviews naming a person appear in no summary, and reach only the manager, as they come. 17-reviews_naming-a-person.pdf12 positive · 5 negative · no count
✎ Framework · no count per employee, no attachment to a personnel file
What is kept: the review, the subject it describes, the facts verified and their result, the reply prepared and whether it was published, and the operational flag where there was one.
What is not kept: no count per employee, no profile of the review's author, no reputation score, and no list of reviews "to be removed".
Why "no reputation score": an overall rating rises and falls without saying why, and it makes the twenty-nine reviews describing the same journey step disappear. What gets fixed is not a rating.
And "no removal list": such a list rarely exists without eventually being used, and its very existence is a problem the day somebody finds it.
What the monthly summary contains: reasons for dissatisfaction by subject, reviews describing a verified operational fact, reply times, and unanswered reviews. Four indicators about the service. what-is-kept.pdf5 items kept · 4 never produced
✎ Framework · retention periods to be set by the company
Your case is not here? That is exactly what a 15-minute conversation is for. Book the free audit →
What does the agent actually do?
One task, carried out on your customer reviews. All the uses work in support, subject to your approval.
Reading across platforms
Brings into a single queue the reviews on your Google Business Profile, those on Trustpilot and the comments left on your Facebook pages and Instagram accounts.
A reply in your brand's tone
Drafts a personalised reply, consistent with your reply guidelines, never generic.
Detection & moderation
Flags the reviews to moderate (negative tone, sensitive remarks, criticism of a department) for priority reading.
Customer support
For questions and requests outside public reviews, a customer support agent answers from your documentation.
On quote View the agent page →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.
Social media posts
For your regular posts, a dedicated assistant prepares templates, calendar and scheduling.
Social media assistant (posts) from 522 € excl. VAT / month Social media posts →Monitoring & reporting
Tracking trends and brand mentions, with reports ready to talk through to steer your reputation.
Campaign watch & reporting from 539 € excl. VAT / month Monitoring & reporting →In 15 minutes we identify the most relevant agent — without oversizing the project.
How much time can a team win back?
By preparing each reply and prioritising sensitive reviews, drafting time and the delay before replying both come down, reinvested in the customer relationship. How large the gain is depends on your volume and remains to be confirmed by a pilot.
The stages of your AI agent project
Audit & scoping
15 minutes to target the use case with the best return.
Quote or direct sign-up
A catalogue offer is bought online; a specific need gets a costed quote.
Design
We design the agent and its guardrails.
Integration & testing
We connect your tools to the agent, which is itself hosted in France.
Rollout
Going live and training your team.
Operation
Continuous supervision and improvement.
One package, one agent
An agent for replying to customer reviews (brand tone, moderation), installed and operated for you.
Setup + controlled subscription
- Installation, configuration and training for your teams
- Operation, human oversight, updates and support
- Sovereign hosting in France, a dedicated and isolated resource
All inclusive, no setup fee
- Setup included (installation, configuration, training)
- Operation, human oversight, updates and support
- Sovereign hosting in France, managed end to end
On site, you own it
- Hardware installed on your premises (you own it)
- French / European AI models run locally
- Secure remote maintenance (Pro support included)
Four guarantees that matter to your customer relations
Related resources
Your questions, our answers
Does the agent publish replies by itself?
How does the agent handle a negative review?
Are our reviews and customer exchanges protected?
Does it connect to our existing review platforms?
How long does it take to deploy this agent?
Does the agent reply to positive reviews too?
Which tools can people use to talk to the agent?
Does the agent follow our Google and Trustpilot reviews, and the comments on our social accounts?
Other agents for customer relations
Let's size up the potential in your customer reviews
15 minutes to identify your priority platforms — hosted in France, supervised, with no commitment.