Case study · MicroStars
MicroStars: an AI assistant that drafts customer service replies, and never sends anything on its own.
Before, to answer a parent, an employee opened five tools, looked for the right file and wrote the reply from scratch. Today an assistant built on Claude, Anthropic's AI, gathers the file and prepares a draft. The employee reads it, edits if needed, and sends. The AI never speaks to parents directly.
- Who
- A network of 44 micro-nurseries
- Who is involved
- The customer service team answering parents
- Existing tools
- 5 tools already in place, none replaced
- What to watch
- Children's data: health, custody, invoices
− 61%
parent wait time before a first reply, during business hours, compared with last year
47%
of the requests the assistant is allowed to handle arrive with a reply already drafted
84%
of those drafts are used by the team, 59% of them sent as is
The client
MicroStars runs a network of 44 micro-nurseries. A small customer service team answers parents every day: why has my invoice changed? has my direct debit gone through? where does my CAF application stand? Every reply matters to the relationship with the family, and touches information about children.
The problem
For a simple billing question, the employee went through the customer inbox, the client file, billing, direct debits and the CAF benefits tracker in turn, then identified the family from an email address or a phone number. Every manual cross-check was a chance for a discrepancy. Some requests were left unsorted. The whole thing rested on about thirty automations built through successive iterations, the oldest of which ran to 116 steps, and whose maintenance called for a clean rebuild.
The solution
First, a single record per family that gathers everything, without touching the existing tools. Then an AI assistant that reads the question, finds the right family and prepares a draft reply. It checks its own work, then drops the draft where the team already works.
- We started by tidying up, not with the AI. We inventoried everything already in place without breaking anything, then created a single record per family, fed by the existing tools and hosted in France.
- We set who sees what: 8 access profiles, and by default, nobody sees anything.
- The assistant only ever receives simple answers (is the direct debit mandate active? yes or no), never the documents themselves.
- Sort: it reads the question. If the subject is sensitive, it leaves it to a human.
- Find: it finds the right family. If in doubt, it hands over.
- Draft: it writes a draft in the style of the team's real replies.
- Self-check: a second AI and fixed rules verify the draft before it is shown.
One way to picture it
It is like an order picker: it fetches the right parts from the warehouse and prepares the parcel. But it is the team lead who checks and approves the shipment. And some parcels (children's health, custody rights, disputes) never go through the picker: they go straight to the lead.
Results
- Parents wait 61% less time for their first reply (during business hours), and their request is resolved 55% faster, compared with the same period last year.
- 47% of the requests the assistant is allowed to handle arrive with a draft ready, that is 24% of all requests. The remaining 38% are not suited to it: sensitive subjects, automated messages, bank notifications.
- 84% of drafts are used by the team, 59% of them sent without changing a word.
- 100% of replies are reviewed and sent by a person. The AI never writes to parents.
- Less than 10 US cents of AI cost per request handled.
What changed
The team no longer starts from a blank page: for nearly one handleable request in two, the reply is already waiting, with the right file in view. Parents get an answer sooner, and their problem is resolved sooner.
Sensitive information circulates less: the assistant works on yes or no answers, never on the documents. And the single record per family becomes the foundation for what MicroStars does next: CAF benefits, phone calls, new types of requests, with the same safeguards.
Safeguards, at every step
- On arrival
- Termination, complaint, unpaid invoice, health, custody: straight to a human.
- In the data
- The assistant can read, never modify. It has no access to custody files or to documents.
- Before showing
- An unknown amount, an IBAN, a missing first name: the draft is discarded.
- At the end
- A person reviews, adapts and sends. Nothing goes out on its own.
“We were not trying to answer more cheaply, but to answer better: the time saved on hunting for information goes back into the relationship with families, on the phone in particular.”
Translated from French. Original quote: « Nous n'avons pas cherché à répondre moins cher, mais à répondre mieux : le temps gagné sur la recherche d'information est rendu à la relation avec les familles, notamment au téléphone. »
“We have to keep a human in the loop: the assistant augments the team without replacing it, our advisers keep the judgment and the signature at every step.”
Translated from French. Original quote: « On est obligés d'avoir un humain dans la boucle : l'assistant augmente l'équipe sans s'y substituer, nos conseillers gardent à chaque étape le jugement et la signature. »
“A scope pruned and prioritized from the start is what separates a project that ships from one that never will.”
Translated from French. Original quote: « Un cadrage élagué et priorisé dès le départ, c'est ce qui sépare un projet qui aboutit d'un projet qui n'aboutira jamais. »
Under the hood
- The AI
- Anthropic's Claude: one model that drafts (Sonnet 5.5), another that reviews (Opus 5.5)
- The data
- A secured database, 8 access profiles, sensitive documents encrypted
- Hosting
- In France, updated every hour on weekdays
- Reliability
- 600 automated tests, trials on real requests before going into service
- Learning
- Replies approved by the team become examples from the next day
Less than 10 US cents per request. The AI bill goes directly to MicroStars: the tool belongs to them.
Publication
- Published
- October 2026
- Reference
- GW · CS-2026-006
- Project team
- Growth Wave · delivery lead: Raphaël Masson · data: Guillaume Piet · development: Thomas Alpa · data protection: Eden Henninot
- MicroStars · Louis Aubé, Chief Administrative and Financial Officer
Case study published by Growth Wave OÜ, with the agreement of MicroStars. Usage rates taken from the assistant's reports over the first weeks in service; wait times (median values) measured on parent requests from September 23 to October 12, 2026, compared with the same period in 2025; cost taken from the Anthropic account. Figures validated by MicroStars. No family data appears in this document.
Are your teams spending their days hunting for information?
Tell us about your CRM, your team and where you're losing time. We reply within 24h.