AI customer support agent: answering 24/7 without overloading the team
A large share of the requests a customer service receives are repetitive and arrive at any hour: where is my order, how do I return it, how do I reset my access. These are exactly the ones an agent can handle instantly. Hosted in France — on local inference or an isolated resource — your AI agent takes on that mass and gives your teams back the cases that genuinely deserve their expertise. The agent assists, the human decides.
Updated on
If delivery goes beyond tomorrow, I can open a priority follow-up.
⛓ Source · your order system + the carrier
I am preparing the label, to be confirmed.
✎ Action · return label ready — the customer confirms
A Blue Lemon Agent customer support agent is an assistant that answers requests automatically around the clock: frequent questions, order tracking, level 1 issues. It resolves most cases on its own, escalates the rest to a person with the context, across several channels and in your service language. It runs on local inference or is hosted in France: your customers' data is 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.
Reference points describing our offer, not results measured at a client. The scale of the gain is confirmed by a pilot on your own scope.
What an AI agent changes when it comes to automating support
Customers expect an immediate answer, at any hour and on every channel. But answering repetitive requests one by one ties up the team, stretches response times and costs money — especially during peaks, where the data exchanged is also personal data.
! The issue
Your customers expect an immediate answer, at any hour — and most of their questions call for an answer that is already known. An agent available 24/7 answers within the second, while your advisers concentrate on the exchanges that make the difference. What remains is choosing where identities, orders and customer exchanges live: with you, and not with a third party hosted outside Europe and subject to the Cloud Act.
✓ Our answer
A support agent is only of interest if it is sovereign and confidential by design. Local inference or an isolated resource hosted in France, human oversight, systematic escalation of out-of-scope cases: 24/7 availability is never paid for in lost confidentiality. The aim is not to replace your agents, but to give them back the cases that genuinely deserve their expertise.
Your customers' data: sovereignty & compliance
A support agent handles personal data (identity, orders, exchanges). Here is how the architecture of our agents protects it, conversation after conversation.
Local inference
The agent can run on a machine belonging to the company: no customer data leaves the network, nothing passes through a cloud.
Hosting in France
Otherwise, a dedicated and isolated resource, hosted in France under French law — your data: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
Architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
One isolated resource per client
No pooling of data: an environment strictly dedicated to your company.
Encryption & controlled access
Encryption in transit and at rest, role-based access (RBAC), strong authentication and logging.
AI Act: governed deployment
An agent strictly in support; no sensitive decision 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.
- The encryption mechanisms in transit and at rest, their components and key management are those documented for the architecture chosen.
- 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.
· One question has recurred 41 times since Monday, always after Friday's update. It has no answer in any of your help pages.
· Seven requests have waited more than 48 hours, all on the same subject, all waiting on information one department holds.
· A help article describes a screen that no longer exists. It is viewed 60 times a week — the third most-read article.
· Three customers have written twice about the same subject under two different references, and so are getting two answers from two people. morning-watch_4-flags.pdf4 flags · 1 question with no answer
⛓ Source · incoming requests, help base, update log, ticket history
The question, in its most frequent form: "where has the export button gone?" — 41 times since Monday, in 9 wordings.
What I looked for: an answer in your help pages, your release notes and your standard replies. Nothing. Friday's release note mentions "redesign of the reporting interface" and does not say where export moved to.
What I did instead, and it matters: I did not guess. I could have answered "look in the Reports menu", which is plausible and may even be right — but a customer who follows a wrong answer comes back angrier than one who was told we were finding out.
What the 41 received: "this question is going to the product team; you will have an answer within 24 h", and the question did go.
What I propose: one sentence of answer, validated by product, and a line in the release note. The 41 become zero. question-41-times_no-answer.pdf9 wordings · 0 answer in the help base
⛓ Source · 41 requests since 03/08, help pages, release note of 31/07
Routing follows what unblocks: a question with no written answer goes to the product team, with the wordings and the volume — 41 occurrences are settled by one sentence, not by 41 replies; blocked requests to the department holding the information, grouped, not one at a time; an out-of-date article to its author, with its view count; duplicate requests to both people answering, before they send.
With a chase: 4 h on a customer request waiting 48 h, 24 h on a high-volume question, 7 days on the rest. Then a monthly summary: by subject and by help article, never by advisor.
What this morning has already given back: 41 requests reduced to a single sentence to write — 9 wordings, one and the same hole in Friday's release note — 7 requests stuck for over 48 hours unblocked by a single department, and 3 customers who will not get two answers from two people. From tomorrow: the third most-read article in your help centre, 60 views a week, stops describing a screen that no longer exists, and one sentence validated by product takes the 41 to zero. The access is yours: requests, help pages and release notes opened by role, logged, withdrawn with a word — your customers' data stays inside your walls. Every answer comes out of what you have written, and when nothing is written I say so and escalate: "you will have an answer within 24 h" kept, rather than a plausible answer that brings the customer back angrier. What the net catches, and how often it tightens: a request waiting 48 h chased at 4 h, a high-volume question at 24 h, an out-of-date article flagged with its view count — and no request ever closed without an answer. The next step is ready: the reply sentence and the release-note line are drafted, one product sign-off is missing; the 41 customers are identified and can be called back in one go.
✎ Proposal · watch and chases to be configured — you set the thresholds
Over the past month: 1,840 requests. I handled 1,190 end to end, each with the help page cited and its date. 650 went to an advisor.
What I answer: questions whose answer is in your pages, your terms, your release notes or your standard replies. I give the page, and the customer can open it.
What I hand over, and why the 650 split into three: 310 because the answer exists nowhere; 221 because the request concerns a particular case on the account — a refund, a goodwill gesture, an exception; 119 because the customer is unhappy, and unhappiness is handled by a person.
The third category deserves saying: I detect irritation from the wording, and I hand over before answering, not after failing. An irritated customer answered correctly by a robot is still irritated. 1840-requests_650-handed-over.pdf1,190 handled · 650 handed over, in 3 reasons
⛓ Source · 1,840 requests this month, help pages cited
The wording used: "I do not have this information in the documentation available to me. I am passing your question to an advisor; you will have an answer within [time]."
The time frame is not invented: it is the one you set for this type of request. If you have not set one, I announce none — I say an advisor will come back, with no figure. A time frame promised and missed does more damage than one never promised.
Why I do not offer an approximate answer, even announced as provisional: "I believe that…" is read as an answer by a customer who wants to move on, and they will apply it — the note of caution does not survive being copied into an internal ticket.
What I do instead, and it is measurable: I pass the advisor the question as it was asked, what I searched for, where, and the two closest pages with the reason each falls short. The advisor does not start from nothing and does not repeat the same search.
Across the 310 handed over for lack of an answer: the advisor took on average 6 minutes, against 11 on requests arriving without that context. Across 310 requests, that is more than 25 hours handed back to support — rounded down. handover_what-is-passed-on.pdf310 requests · 11 → 6 minutes
⛓ Source · 310 requests handed over, measured handling times
The three cases: an article wrong and little read — annoying; an article missing — annoying, and the customer sees it at once; an article wrong and much read — the customer does not see that it is wrong, believes they did it badly, and tries again.
What I record: the article describes a screen removed in Friday's redesign. 60 views a week, third most-read article.
What it is already producing: 14 of the week's 41 requests cite that article — customers read it, followed it, and write to say it does not work.
What I have written: the banner, dated, ready to place at the top of the article — the screen changed on 31/07, the page is being updated, and here is where to ask in the meantime. Two minutes to publish it, and the 14 requests a week stop.
What is missing to rewrite the article, and it is one single fact: where the button went. That is precisely the question nobody has answered — I put it to the product team this morning, with the 14 requests as evidence. The day I have the answer, the page is rewritten within the hour. out-of-date-article_60-views.pdf60 views/week · 14 requests citing it
⛓ Source · help article, view log, 41 requests this week
What I cross-referenced: for each article, the view count and the number of requests arriving despite it being read — that is, customers who opened it and wrote anyway.
What that gives: 9 articles account for 61% of post-reading requests. They are not the oldest: three are less than six months old.
The three profiles it reveals: 4 articles describe a procedure that has changed; 3 answer a question adjacent to the one the customer has — they are correct and beside the point; 2 are accurate but unreadable, with a seventeen-step procedure.
The second family is the most interesting: those articles are not wrong, they are badly titled. The customer finds them, reads them, and leaves with the answer to a different question.
What I have prepared for the nine: the corrected title for the three badly titled ones, taken word for word from customers' own wording; a four-step re-cut for the two unreadable ones; and, for the four out of date, the dated banner. Nine written proposals, and for each the exact wordings of the customers who wrote after reading it.
What is left to their authors: approval. A help page binds the company to what the product does, and its author answers for it — one review per article, and the 61% falls. 9-articles_61-pc-of-requests.pdf9 articles · 3 profiles · customer wordings
✎ Support · 9 articles · customer wordings supplied for each
What multichannel means concretely: WhatsApp Business, your site chat and email share the same knowledge base, the same standard answers and the same handover. What changes between channels is the form, not the substance: on WhatsApp the answer runs to a few lines and the help-page link is short; by email it carries the request reference and the cited page in full.
Why one base and not three: your three customers who wrote twice did so on two different channels. With three bases they received two answers written in good faith that did not say the same thing. With one, the second request arrives with the first in view.
The service language, and I would rather be plain: the service language is single and you set it — French here. A request arrives in Spanish: I answer in the service language, saying so, and I hand over if the customer cannot read it. I open another language the day you order it, and it is billed per language opened; beyond a few languages, what you need is a dedicated multilingual support agent, not me. What I steer clear of, and here is why: answering in a language nobody on your side proofreads: a wrong answer in Spanish surfaces three weeks later, and it binds you just as much.
The figure that does not flatter me: out of 1,840 requests this month, 96 arrived in a language other than French — 5.2 %, and 71 came from Spain, all on two subjects only, delivery and returns. What I do with it: I hand you the two subjects; two translated help pages would cover 71 of the 96 requests, and the service-language question would then bear on 25 requests a month, not 96.
⛓ Source · 3 channels on one base, 96 non-French requests out of 1,840, 71 on two subjects
Order and case tracking: I read the status in real time and give the exact state, its update time, and the next step with its expected date — never an “in progress” that teaches nothing. If the carrier has posted no status for 48 hours, I say so and give the time of the last reading rather than repeating a stale status. For the month: 640 order-tracking requests, 612 handled end to end, the remaining 28 passed to an adviser because the carrier's status contradicted your system's.
Level 1 resolution: requests settled by a step documented on your side — access reset, delivery address change before dispatch, adding a user, returning an item within the window, issuing an invoice. For the month: 410 level 1 requests, 358 resolved, with the help page cited and the action taken described to the customer so they can repeat it.
Where level 1 stops, and the line is written: as soon as a request touches a refund, a goodwill gesture, an exception to the return window or a complaint, it goes to an adviser with the full history. Level 1 resolution is judged on what it passes on as much as on what it resolves.
The figure that does not suit me: of the 358 level 1 resolutions, 22 produced a second request from the same customer within seven days — 6 %. It was not resolved. The explanation: 17 of the 22 concerned access reset, whose help page describes a screen changed in May. What I did with it: the page is flagged to the product team with its 60 weekly views, and meanwhile I answer with the real path, not with the page. The following month: 8 second requests out of 374, 2 %.
⛓ Source · 640 order-tracking requests, 612 handled, 410 level 1 requests, 358 resolved, 6 % then 2 % repeats
What I record: three customers opened two requests on the same subject a few days apart, under two different references — one from the form, one by email. Two advisors are working on them.
What will happen otherwise: two answers, written in good faith, that will not say exactly the same thing. The customer concludes the company does not know.
What I propose: flag the duplicate to both advisors before sending, with both threads. Not an automatic merge: the two requests may cover two distinct aspects, and merging wrongly would lose one of them.
What links them: the customer account and the subject, not the name or the address.
Over three months: 58 duplicate requests, of which 22 received two diverging answers. That is the figure that counts: the other 36 produced only redundancy, the 22 produced a contradiction. 58-duplicates_22-diverging.pdf58 duplicates · 22 diverging answers
⛓ Source · 3 duplicated requests, 3-month history, conversation threads
What is kept: the request, the answer and the page cited, the handover and its reason, and what the advisor answered afterwards.
What is not kept: no customer profile built from their requests, no satisfaction score attached to a customer, no statistics per advisor, and no analysis of a customer's tone beyond the current request.
The last point matters: I detect irritation in order to hand over immediately — and that detection does not outlive the conversation. A customer irritated on a Tuesday is not "a difficult customer" the following month, and nothing I keep would allow anyone to say so.
What the monthly summary contains: the most frequent subjects, questions with no written answer, articles that produce requests despite being read, and handover reasons. Four indicators about your documentation, none about people. what-is-kept.pdf4 items kept · 4 impossible
✎ 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 →
The uses of a customer support agent
Each use corresponds to an agent we deploy. All of them work in support; out-of-scope cases are escalated to a person.
Answering frequent questions
Delivery, returns, account, opening hours, terms: the agent answers 24/7 from your knowledge base.
Order & case tracking
Looks up the status of an order or a case and gives an exact answer, in real time.
Level 1 resolution
Diagnoses and resolves simple problems: resetting access, configuration, everyday steps.
Escalation to a human agent
Passes complex cases to an adviser with the full history of the conversation.
Multichannel, one service language
Works on WhatsApp Business, your website chat and email, in the service language you define. Further languages are an option at checkout, billed per language opened; beyond a few languages, a dedicated multilingual support agent takes over.
Ticket & CRM updating
Updates the ticket or the CRM and provides reporting: volumes, resolution rate, reasons.
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.
Voice support & switchboard
Takes the load off the switchboard by answering first-line calls, day and night.
Voice agent (AI switchboard) from 856 € excl. VAT / month Voice agent →Appointment booking
Qualifies the requests that need it and automatically schedules an appointment or a call back.
Lead qualification / appointment booking agent from 602 € excl. VAT / month Qualification & appointments →In 15 minutes we identify the most relevant agent — without oversizing the project.
How much time can a customer service win back?
By automating the base of repetitive requests, a customer service can aim for a sharp reduction in cost per contact and a first-response time brought down to immediate — without reinforcing the team during peaks. Gains to be validated on your own volume and request mix.
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.
A support agent, installed and operated for you
A level 1 customer support agent for your website, installed and operated for you. Choose according to your channels and your volume. Prices exclude VAT — annual subscription, the time it takes for the gains to settle in.
Four guarantees that matter to a customer service
Your questions, our answers
Does the AI chatbot replace my support team?
Which channels and languages does it work in?
What happens when the agent does not know the answer?
Is my customer data protected?
Does the agent state that it is an artificial intelligence?
How long before the agent is operational?
Does the agent take sensitive decisions on my behalf?
Which tools can people use to talk to the agent?
Can the agent notify my customers by text message?
Other agents for customer relations
Let us estimate the potential for your customer service
15 minutes to identify the use case with the best return — hosted in France, supervised, with no commitment.