DEV Community: Mirage Cloud IA

Keyword Research Without Paid Tools: A Practical Method

Mirage Cloud IA — Tue, 20 Oct 2026 04:08:14 +0000

Paid keyword tools are good, but you do not need one to find out what your customers search for. Most of what you need is already in your inbox, your analytics and the search results themselves. The method below uses free sources and a spreadsheet, and the result is usually more relevant than a list of high-volume phrases that nobody who might buy from you ever types.

Start with relevance, not volume
A phrase searched a thousand times a month by people who will never buy is worth less than one searched thirty times a month by people ready to book you. For a small business the useful keywords are specific, local or niche, and tied to a problem you solve. Keep that filter in mind throughout.

The method

List the questions your customers ask. Go through emails, calls, messages and sales notes. Write the questions as they were phrased, not as you would phrase them. This gives you the vocabulary people actually use.
Expand with the search box. Type the start of each question into Google and note the autocomplete suggestions. Then look at the People also ask box and the related searches at the bottom of the page, and add the useful ones to your list.
Mine your own data. If your site is verified in Google Search Console, open the performance report and look at the queries that already bring impressions. Phrases where you appear on page two, roughly positions 8 to 20, are the quickest wins, because a better page for that query can move you up.
Check seasonality and trends. Google Trends is free and shows whether interest in a phrase is rising, falling or seasonal, and it lets you compare two phrases directly.
Look at rough volume ranges. Google's Keyword Planner is free with an advertising account and shows ranges rather than exact numbers when you are not running ads. Ranges are enough to tell a phrase with a handful of searches from one with thousands.
Read the results page. For each candidate, search it and look at who ranks. If the first page is full of large brands, government sites and national publications, the phrase is hard. If it shows forums, thin pages or out-of-date articles, there is room.
Group and map. Cluster phrases that mean the same thing and give each cluster one page: one main keyword per page, with the variations used naturally in the text.
Prioritise. Score each cluster from one to three on relevance to what you sell, closeness to a purchase, how beatable the results are and how much effort the page needs. Write the highest totals first.

Match the intent
Every search has a purpose. Some people want to learn, some want to compare, some want to buy and some want to find a place. A page that answers a buying question with a long explanation, or an information question with a sales pitch, disappoints the visitor and performs poorly. Look at the format that already ranks, whether a guide, a list, a service page or a map, and match it.

Searching in French
If your customers are French, research in French. Direct translations of English phrases often miss the terms people really use, and accents, plurals and synonyms change the results. Use the autocomplete and related searches in French, and test two or three variants of the same idea, such as the different words for quote, estimate and price.

Keep the list working
Put everything in a spreadsheet with columns for the phrase, the intent, the page it belongs to, the priority and the date you last reviewed it. Check Search Console monthly to see which pages are gaining impressions and which phrases you are close to ranking for. Update older pages instead of writing duplicates, and never target the same phrase with two pages, because they compete with each other.

What AI can and cannot do here
An assistant is good at clustering a long list of phrases, suggesting variations, drafting outlines and writing first versions of titles and descriptions. It cannot tell you real search volumes, and if you ask for them it will produce plausible numbers that are made up. Take volumes and rankings from your tools and use the assistant for the thinking and the drafting. Mirage Cloud includes a marketing agent covering SEO, content and digital presence, which fits that second role.

How to Cut No-Shows and Late Cancellations

Mirage Cloud IA — Mon, 19 Oct 2026 04:20:47 +0000

An empty appointment slot is a cost with no income against it: the rent, the equipment and the person are all paid for, and the hour cannot be sold again. For businesses that run on bookings, such as clinics, salons, trades, consultants and tutors, missed and late-cancelled appointments are often the largest leak in the week, and they rarely appear in any report.

Put a number on it
For one month, count the appointments that were missed or cancelled with less notice than you need to refill the slot, then multiply by the average revenue per appointment. That figure is what the problem costs you, and it tells you how much effort a fix is worth.

Why people miss appointments
Mostly they forget, or something comes up and cancelling feels awkward or takes effort. A smaller group booked several providers and kept the one that suited them best. Knowing which reason dominates in your business shapes the fix: forgetting is solved with reminders, awkward cancellation with an easier process, and low commitment with a deposit or a firm policy.

What works
➞ A confirmation at booking with the date, time, place, what to bring and how to change the appointment.
➞ Reminders at 48 hours and 24 hours, and a short message on the morning of the appointment for the highest-value slots. Each one should let the person confirm or reschedule in one tap.
➞ An easy way to reschedule. A link in every message. People who can move an appointment in ten seconds are less likely to simply vanish.
➞ A clear cancellation policy, stated at booking and repeated in the reminders: how much notice you need and what happens otherwise.
➞ A deposit or card guarantee for long or expensive slots, or for people with a history of missing.
➞ A waiting list you can message when a slot opens, so a late cancellation does not always mean an empty chair.
➞ Shorter booking windows where you can. The further ahead someone books, the more likely it is that something changes.

Deposits: acompte or arrhes
In France the wording matters. Under the Consumer Code, when a consumer pays money in advance it is treated as arrhes unless the contract clearly says otherwise. Arrhes let either side withdraw: the consumer loses what they paid, and a professional who withdraws must return double. An acompte, by contrast, is a first instalment that commits both sides. If you want a deposit that actually discourages no-shows, say acompte in your terms, your quote and your receipt, and use the same word everywhere. A deposit held for more than three months can also start earning legal interest for the consumer.

Keep the policy fair and consistent
Apply it the same way to everyone, with room for genuine emergencies. A fee charged to one client and waived for another becomes an argument. Keep any fee proportionate, because a court can reduce a penalty it finds manifestly excessive. Show the policy at booking, in the confirmation and in the reminders, and consider waiving the fee the first time with a polite note about the rule.

Reminders and data protection
Appointment reminders are service messages that support the booking, and they generally do not need the same consent as marketing. Keep them to the appointment and do not add promotions. Make sure your booking tool or messaging provider has a data processing agreement and does not use phone numbers for its own purposes.

Track what happens
Record the no-show rate by day, time, service and booking channel. A pattern, such as Monday mornings or online bookings, tells you where to apply deposits or extra reminders instead of burdening everyone. Review it monthly and adjust.

The calls that never become appointments
The other half of the problem is demand you never see: callers who ring while you are with a client and go elsewhere. A phone receptionist that answers, takes an accurate message and passes urgent requests to you closes that gap. Mirage Cloud's receptionist agent, Xavier, does this for businesses that cannot answer every call, and it can be tested in the browser before a number is attached. Reminders and a policy fix the appointments you have. Answering the phone fixes the ones you do not.

A Customer Onboarding Sequence for Small Businesses

Mirage Cloud IA — Sat, 17 Oct 2026 04:19:06 +0000

Businesses spend heavily to win a customer and then leave the first month to chance. The signature or the first payment feels like the finish line, but for the customer it is the start of a period in which they decide whether they made the right choice. A short, planned sequence over the first thirty days costs little and removes most of the avoidable early cancellations.

Define what success looks like
Pick the moment at which the customer first gets the value they bought. For a bookkeeping client it might be the first month's accounts delivered. For a gym, the third visit. For a training course, completing the first module. Call it activation and work backwards from it: everything in the first weeks exists to get the customer there quickly. The time between purchase and activation is the number to shorten.

The sequence
➜ Day 0, welcome. Confirm what they bought, what happens next and who to contact. One message, not five. If there is a form to fill in, send it now and make it the only task.
➜ Days 1 to 2, the first win. Give them something small they can use or see: a first deliverable, a setup call, a quick-start page. Early progress matters more than completeness.
➜ Day 3, check the basics. A short message asking whether anything is unclear or blocked. Make replying effortless.
➜ Day 7, a personal contact. A short call or a personal email from a named person. Ask what is going well and what is not. It is the most valuable touchpoint in the sequence and the one most often skipped.
➜ Day 14, show progress. Summarise what has been done and what comes next, so the customer can see the value without hunting for it.
➜ Day 30, review. Ask how it is going and offer to adjust, and only then ask for a testimonial or a referral if they are happy.

Gather information once
Ask for what you need in a single form, pre-fill what you already know and never ask the same question twice. Every repeated request tells the customer that your left hand does not know what your right is doing. If sales promised something specific, write it in a short handover note so that delivery knows about it.

Messages that work
Keep them short and specific. A welcome message can say who they will hear from, what happens this week and what you need from them, in four or five lines. A day 3 check can be as short as: How is the first week going? If anything is unclear, reply here and I will sort it out today. Short messages from a named person get answered. Long messages from a company address get filed.

What to measure
➜ Time from purchase to activation.
➜ Share of new customers who complete the first step within a
week.
➜ Cancellations in the first month, and the reason given.
➜ Replies to the day 7 contact, and what they say.

Also call five customers who left in the first two months and ask what happened. The patterns you hear in five calls will tell you more than a dashboard.

What to automate and what not to
Automate the confirmations, the form, the reminders and the scheduling. Keep a person on the day 7 contact and anywhere the customer expresses doubt. Customers can tell the difference between a sequence that is automated and one that is impersonal, and only the second hurts.

Adapting it for members
The same structure works for associations and clubs with new members: a welcome, a first meeting or activity within the first fortnight, a personal contact from someone on the committee and a check-in at one month. The aim is the same, to get members to a first activity quickly, and a small association can manage it with a shared checklist and a few volunteers.

Where tools help
The sequence is a series of repeatable messages and reminders, which suits automation. Mirage Cloud connects to Brevo, Google Calendar and Yousign, which covers the email, the scheduling and the signature, and it includes agents that can draft the messages and the handover note. The first-week call stays with a person.

A 60-Minute Month-End Routine for Small Businesses

Mirage Cloud IA — Fri, 16 Oct 2026 04:49:27 +0000

Most small business owners meet their books in two bursts: a scramble before a VAT deadline and a bigger one at the year end. Between them the numbers drift, receipts go missing and a mistake made in February is still sitting there in December, much harder to trace. A short routine on the same day each month stops that, and it takes about an hour once the habit settles.

Why monthly
➠ Errors are cheapest to fix while you still remember the
transaction.
➠ VAT and payroll declarations are built on the same data, so

clean months make clean returns.
➠ You see problems with cash and margins in time to react.
➠ Your accountant spends less time untangling things and
charges accordingly.

The routine
1.Bank reconciliation, ten minutes. Match every bank transaction to an invoice, a receipt or a known category. Put anything you cannot explain on an exceptions list instead of guessing.

2.Sales, ten minutes. Check that every job delivered this month has been invoiced, that credit notes are recorded and that nothing was invoiced twice.

3.Purchases and expenses, ten minutes. Make sure every supplier invoice and receipt is captured and attached, and that expense claims are in. Chase what is missing now, while suppliers still remember.

4.Money owed to you, ten minutes. Review unpaid invoices by age, send reminders for anything past due and flag disputes. This is where late payment penalties come from, so apply your rules consistently.

5.Money you owe and what is coming, ten minutes. List what falls due next month, including VAT, payroll contributions, loan repayments, insurance and tax instalments, and check that the cash is there.

6.Five numbers and three lines, ten minutes. Record revenue, gross margin, cash in the bank, overdue invoices and committed payments for next month. Write three lines on what changed and why. Send the pack to your accountant or your business partners.

Keep it light
➠ Pick a fixed day, such as the third working day of the

month, and put it in the calendar.
➠ Use the same checklist every time. A routine you improvise
is a routine you skip.
➠ Lock the month once it is closed, so figures do not change
underneath you.
➠ Keep personal spending out of the business account. Mixed
accounts multiply the time this takes.
➠ Note unusual items in a sentence at the time. Memory fades
faster than you expect.

What a good month-end gives you
After three or four cycles you have a series of comparable months, which makes patterns visible: margins slipping, a customer paying a little later each time, a cost creeping up. The annual accounts show you the same things, but only months after they stopped being easy to fix.

French specifics
VAT and payroll declarations have fixed dates that depend on your regime and your size, so build them into the checklist. Since 1 September 2026 French businesses must also be able to receive structured electronic invoices, with the obligation to issue them following in phases, so check that incoming invoices reach the platform you use and are captured as part of step three.

Where tools help
The routine is quick when bank, invoices and accounts sit in one place and slow when they do not. Mirage Cloud connects to Qonto and Pennylane and includes a finance agent scoped to cash position, margins and budgets, which suits steps one, four and six. Connected tools save time on collection, not on judgement. You still decide what an odd transaction is.

How to Write AI Prompts That Work for Everyday Business Tasks

Mirage Cloud IA — Thu, 15 Oct 2026 04:27:23 +0000

Most advice about prompts treats them as a technical skill with secret formulas. In practice a prompt is a brief, and the people who get good results are the ones who would also write a good brief for a new colleague: say what the job is, give the background, show what good looks like and say how you will check it. A capable assistant with no context fills the gaps with guesses, and the guesses are where the poor results come from.

A six-part structure

Context. Who you are, what the business does, who the output is for and which country and date the rules should come from.
Task. One clear job, stated with a verb: write a reply, summarise, compare, list, check.
Inputs. The material to work from, with a clear marker for where it starts and ends, so it is not confused with your instructions.
Constraints. Length, tone, what to include and what to leave out.
Format. Email, table, numbered list, two paragraphs. Say which.
Check. Ask it to list its assumptions and to flag anything it could not establish from the material.

You will not need all six every time. For a quick rewrite two lines are enough. For anything with a number, a date or a legal point in it, use all six.
Worked example: a follow-up email
Weak prompt: Write a follow-up email for a quote.
Better prompt: I run a three-person landscaping company in Lyon. Write a follow-up email to a homeowner who received a quote for a garden redesign five days ago and has not replied. The quote was 6,400 euros and includes a free site visit before work starts. Tone: friendly and brief, no pressure. Under 90 words. End with one question that can be answered yes or no. Do not invent discounts or deadlines.
The second version gives the assistant a business, a reader, a set of facts, a tone, a length and a limit on invention. The result needs a light edit rather than a rewrite.
Worked example: a contract
Prompt: Below is a supplier agreement. List the obligations that fall on us, with the clause number for each. Then list every date, deadline and notice period. Then list penalties and termination conditions. If something is unclear, say so instead of guessing. Text: [paste].
Notice what this asks for: extraction with references, not an opinion. You can check every line against the document, which is the point. Then take the list to someone qualified before you rely on it.
Worked example: commentary on figures
Prompt: Here are this month's figures: [paste]. Write three sentences for the owner explaining what changed against last month. Use only the numbers provided. If a reason is not in the data, say it cannot be determined. Do not guess causes.
The last three sentences are what stop a general assistant from inventing explanations.
Habits that make results better
Name the jurisdiction and the date. For rules, say French law and the current year. It reduces answers drawn from another country or an older version.
Show an example of your own good work. One paragraph written the way you want it is worth a page of description.
Work in two or three turns. Ask for a plan or options first, choose one, then ask for the full text.
Ask for alternatives. Three subject lines are more useful than one.
Keep a prompt library. When a prompt works, save it with its date and what it was for.
Ask for sources, then check them. Verify any figure, date or reference before using it.
Mistakes to avoid
Bundling five jobs into one request.
Burying the key fact in the middle of a long paste.
Setting contradictory constraints, such as very short and very detailed.
Pasting passwords, personal data you have no basis to process or other people's confidential material.
Expecting the assistant to remember last week's conversation. State the context again or use a tool that keeps it.
Treating confident output as correct output.
When the tool already knows your business
Much of the effort in a prompt goes on re-explaining context. Tools that are connected to your accounting, banking or calendar remove that step, and tools built around specific roles carry their scope with them. Mirage Cloud works that way, with agents scoped by function and connections to tools such as Qonto, Pennylane and Google Calendar. The briefing discipline above still helps, but you spend less of it on background and more on the task.

What AI Can and Cannot Do for Your Bookkeeping

Mirage Cloud IA — Wed, 14 Oct 2026 05:42:57 +0000

Bookkeeping is a natural target for AI. It is repetitive, structured and full of rules, and most small business owners would happily hand it over. Some of it can be handed over today. Some of it cannot, and the difference matters because errors in accounts do not announce themselves: they surface at the year end, during a VAT review or when the accountant asks why a number looks odd.
What AI does well
➠ Suggesting categories for bank transactions, based on the

description and your past choices.
➠ Matching payments to invoices, including part payments and
slightly different amounts.
➠ Reading data from invoices and receipts, such as the
supplier, date, totals and VAT.
➠ Flagging oddities: duplicate payments, unusual amounts, a
supplier invoice with no matching order.
➠ Drafting payment reminders and supplier queries.
➠ Explaining a report in plain language, which helps the
people who dislike reading spreadsheets.
➠ Preparing a clean list of questions for the accountant.

These are tasks where a mistake is visible, easy to correct and low in consequence, and where the tool saves real time.

Where it is unreliable
Tax treatment of unusual transactions is the main one. A model will state a VAT rate, a depreciation period or a deduction rule with complete confidence, and it may be applying a rule from another country or from a year ago. Anything that changed recently is especially exposed. Judgement calls are the other: whether an expense is personal or business, how to treat a deposit, what to do with a refund that spans two years. These depend on facts and on your accountant's view, not on pattern matching.

Consistency is a quieter problem. The same supplier may be categorised one way in March and another way in October unless you fix a rule, and inconsistent categories make the figures harder to read and to defend.

The accountant's role does not go away
In France the title expert-comptable is protected, and the accountant carries professional responsibility for the figures they prepare or certify. An AI tool cannot take that on. What it can do is reduce the cost of the mess you hand over: well-categorised transactions, matched invoices and a short list of exceptions mean fewer hours at the accountant's rate. Ask your accountant how they want data delivered, which categories they use and which items they always want to see.

Guardrails that keep it safe
Review before posting. For the first few months a person approves every categorisation. After that, review a sample each month, say one in ten.
Build a rules list from your accountant's guidance and apply it consistently, covering things like fuel, meals or software subscriptions.

Keep the source documents attached to every entry. An audit needs the invoice, not a model's reading of it.
Reconcile the bank every month and investigate differences straight away.
Never accept a suggested VAT rate or tax treatment without checking it.
Use read-only connections to bank and accounting tools where you can.
Lock each month once it has been reviewed, so figures do not change silently.

A weekly routine
Twenty minutes is enough. Import the week's transactions, let the tool suggest categories and matches, approve or correct them and look at the list of exceptions. Once a month, send the accountant a summary of what changed and any open questions. Done weekly, the work stays small.

Why e-invoicing helps
Since 1 September 2026 French businesses must be able to receive structured electronic invoices, and the obligation to issue them follows in phases, from large and mid-size companies now to small ones in September 2027. Structured data is easier to read than a PDF, which makes extraction more reliable and cuts re-keying. It does not remove the need to check, but it does remove one source of error.

Protecting the data
Bank data and supplier invoices are commercial information and, for sole traders, personal data. Check that any tool you connect has a data processing agreement, lists its sub-processors and lets you delete what it has stored. Where it relies on an external model provider, find out whether your content is used for training.

Where this leaves you
Mirage Cloud connects to Qonto and Pennylane and includes an accounting agent scoped to French accounting, VAT and tax questions. Its own acceptable use policy says output must be verified before use, which is the right way to treat it: a fast assistant that prepares the work and explains it, with a person and an accountant making the decisions.

AI Meeting Notetakers in Europe: What the Law Expects

Mirage Cloud IA — Tue, 13 Oct 2026 06:11:57 +0000

An AI notetaker takes one click to switch on and produces something genuinely useful: a transcript, a summary and a list of actions, without anyone typing. It also records people's voices, which are personal data, and creates a written record of what was said that can travel further than anyone intended.

Most of the legal questions can be answered with a short announcement and a few settings.

Recording a meeting is processing personal data
Under GDPR a voice recording and its transcript are personal data. That means you must tell participants about the recording, rely on a lawful basis, collect no more than you need and keep it no longer than you need.

French criminal law adds a separate rule: recording words spoken in private or confidential circumstances without the consent of the people speaking is an offence under Article 226-1 of the Penal Code. An ordinary business meeting that has been announced is a different situation, but that rule is the reason never to record people secretly.

Information is not the same as consent
The GDPR does not require the consent of every participant before an AI tool processes a meeting. What it does require is information, given before the recording starts: that the meeting is being recorded and transcribed, why, who will have access, how long it will be kept and how to object. The lawful basis can then be consent, or legitimate interests where you can show the recording is needed and that no lighter option, such as taking notes, would do.

With employees, consent is a weak basis, because the relationship of subordination makes it hard to say it was freely given. Employers usually rely on legitimate interests and document the reasoning. Before introducing a system that records or monitors staff, employees must be informed in advance and, where there is a staff committee, it is consulted.
An announcement that works
Put a notice in the invitation and repeat it at the start of the meeting. Something like this works: This meeting will be recorded and transcribed by an AI tool so that we can produce minutes. The transcript goes to attendees and is deleted after 30 days. Tell us now if you would rather we did not record.
If someone objects, stop the recording and take notes by hand. External participants deserve the notice most, since they have not agreed to your internal policies.
Settings that matter
Do not let the tool join every calendar event automatically. Turn it on for the meetings where it is needed.

Keep the transcript and drop the audio where you can, or set a short retention period for the audio.
Do not share transcripts automatically with everyone invited. Send them to the people who need them.

Limit who can search past transcripts.
Switch off any feature that scores participants, measures sentiment or infers emotions. Since February 2025 the AI Act has banned systems that infer emotions in the workplace, and a meeting tool that does so crosses that line.

Do not reuse recordings to evaluate people's performance, which is a different purpose from producing minutes.
Meetings you should not record
Disciplinary and dismissal meetings, conversations about someone's health, meetings covered by legal professional privilege and discussions with employee representatives need a human note-taker, not a bot. The harm from a misplaced transcript in those situations is high and the benefit is low.

What to check in the vendor
Read the data processing agreement and the sub-processor list. Ask where recordings and transcripts are processed and stored, whether the vendor uses your content to train its models, how deletion works and how quickly, and whether you can export or erase everything at the end. A tool that cannot answer in writing is not ready for business meetings.

A lighter alternative
If the aim is minutes, you may not need a recording at all. Taking your own notes, pasting them into an assistant and asking it to turn them into minutes with decisions and actions gives you most of the benefit and avoids voice data entirely. Mirage Cloud generates documents from descriptions and includes a generalist agent for this kind of task. Whichever tool you use, check the draft against what was actually decided before you circulate it.

AI Chatbot on Your Website: A Compliance Checklist for Europe

Mirage Cloud IA — Mon, 12 Oct 2026 05:34:03 +0000

A chatbot on a website feels like a small addition, a widget you paste in before lunch. Legally, it is several things at once: a new way of collecting personal data, a new source of statements made in your name, and, since August 2026, a system that triggers transparency obligations under the AI Act.

None of this is particularly difficult to handle, but everything is easier to address before launch than after the first complaint.

Tell People They Are Talking to a Machine

Article 50 of the AI Act, applicable since 2 August 2026, requires people interacting with an AI system to be informed that they are talking to AI, unless this is obvious from the circumstances.

A widget with a friendly name and a human photo does not necessarily make this obvious.

What you should do:

Identify the chatbot as AI in the widget label.
Include a clear disclosure in the first message.
For voice agents, include the disclosure in the greeting.
Make it easy for users to request human assistance.

For example:

I am an AI assistant for this business. A person can step in at any time.
Treat the Chat as Personal Data

Conversations contain names, email addresses, order numbers and, because people type what they like, sometimes health or financial details you never asked for.

That makes chat logs personal data, and the usual data protection obligations follow.

A lawful basis: For customer support, this may be contractual necessity or legitimate interests rather than consent, depending on the circumstances.
A privacy notice: Explain the chatbot's purposes, the vendor, relevant AI model providers, recipients, and any international data transfers.
A data processing agreement: Put the appropriate agreement in place with the vendor and review its subprocessor list.
A retention rule: Define how long conversations are stored and implement a deletion routine.
A process for data requests: Make sure applicable access and deletion requests can be handled for chatbot logs.
Data minimisation by design: Do not ask for more information than the task requires, and tell users not to enter unnecessary sensitive details.
Cookies and Scripts

Many chat widgets load third-party scripts that set cookies or access identifiers on a visitor's device.

In France, the CNIL generally expects consent before non-essential trackers are used, subject to applicable exceptions. Refusing must be as easy as accepting.

If the widget loads before the visitor has made the required choice, you may have a compliance problem, whatever the vendor says.

What you should do:

Load the widget only after the required consent has been obtained.
Choose a solution that does not use non-essential trackers where appropriate.
Verify what scripts and cookies the widget actually loads.
Remember that the CNIL recommends renewing consent choices at appropriate intervals, commonly every six months.
Decide What the Bot Is Allowed to Say

Whatever the chatbot tells customers about prices, delivery, refunds or conditions may be understood as coming from your business.

Limit it to a knowledge base you control, make it show or link its sources for factual answers, and have it decline topics outside its scope instead of improvising.

Set clear boundaries:

Use approved, up-to-date business information.
Prevent the chatbot from inventing prices, discounts, or refund promises.
Make it acknowledge uncertainty when it cannot verify an answer.
Restrict legal, medical, and financial advice where appropriate.
Escalate quickly when a customer is upset, asks to make a complaint, mentions a deadline, or repeats a question without getting a satisfactory answer.

The chatbot should know when to stop answering and involve a person.

Make the Handoff Real

A visible route to a human, with a stated response time, is both good service and a safeguard.

Put a contact option in the widget.
Explain when users can expect a response.
Make sure someone is actually monitoring incoming requests during the hours you promise.
Test the handoff before launch.

A human support option is only useful if it works when customers need it.

Check Accessibility

If your site sells to consumers in the EU, the European Accessibility Act may cover the e-commerce service, subject to its scope and applicable exemptions, including the microenterprise exemption.

The chat widget is part of that service.

Before launch, check that it:

Works with a keyboard.
Supports screen readers.
Has sufficient colour contrast.
Does not trap keyboard focus.
Provides accessible labels and controls.

Test the actual widget rather than trusting a vendor's claim of compliance.

Test Before You Go Live

Write twenty real questions from your own inbox and run them through the bot, including French-language questions where relevant, poor spelling, and off-topic requests.

Then try to break it.

Ask it to ignore its instructions.
Ask it to reveal its configuration.
Ask it to promise an unauthorised refund.
Test whether it invents answers when information is missing.
Check whether it escalates questions it cannot safely answer.

Fix what fails, and run the test set again after every significant change to the knowledge base or chatbot configuration.

During the first month, review transcripts every week, using appropriate access controls, and keep a list of questions the chatbot could not answer.

The Launch Checklist

Before going live, confirm that:

AI disclosure appears in the widget label and the first message.
The privacy notice is updated, the data processing agreement is in place where required, and subprocessors have been reviewed.
A retention period is defined and deletion is scheduled.
Required consent is obtained before non-essential tracking scripts load.
The knowledge base is limited to approved information and sources.
Human handoff is visible and monitored.
Accessibility has been tested.
The twenty-question test has passed and weekly reviews are scheduled.
The Same List Applies to Voice AI

The checklist works for phone agents too.

Mirage Cloud's receptionist agent, Xavier, answers calls for businesses, and the points above apply directly:

The AI disclosure belongs in the greeting where required.
Call recordings and transcripts containing personal data need an appropriate retention and deletion policy.
Transferring a caller to a person should be straightforward.
The provider's data processing arrangements and subprocessors should be reviewed.

Mirage Cloud publishes a full subprocessor annex that identifies its voice provider, so businesses can review the data-processing arrangements using documentation rather than relying solely on marketing claims.

The principle is the same for both chat and voice: put the safeguards in place before launch, not after the first complaint.

Donation Tax Receipts for French Associations in 2026

Mirage Cloud IA — Sat, 10 Oct 2026 05:21:25 +0000

Donations are a lifeline for many small associations, and the tax reduction is often what tips a donor from thinking about giving to actually giving. France's scheme is among the more generous in Europe, but it only works if the association is eligible, issues the right receipt and describes the benefit accurately. It also changed recently in a way that most published guides have not caught up with.

The 2026 change most guides miss
The higher 75% rate applies to gifts to organisations that provide free meals, housing or care to people in difficulty, and to organisations that help victims of domestic violence.
It used to apply to the first 1,000 euros given in a year.

Article 28 of the Finance Act for 2026, the law of 19 February 2026, raised that ceiling to 2,000 euros for gifts made from 14 October 2025, and tax authority guidance published in July 2026 confirms it. Above 2,000 euros, the rest of the gift moves to the 66% rate.

The difference is easy to see on a 2,000 euro gift to a food aid association: 1,500 euros back at the new rate, against 1,410 under the old one, which gave 750 euros on the first thousand and 660 on the second.

How the reduction works for individuals
For most eligible associations a donor gets a tax reduction of 66% of the gift, up to 20% of their taxable income, and any excess can be carried forward over the following five years. The 75% rate is reserved for the organisations described above and is capped as explained.
Companies and patronage
Businesses that give to an eligible association can claim a reduction in corporate tax equal to 60% of the gift, within a ceiling of 20,000 euros or 0.5% of turnover excluding VAT, whichever is higher. Any excess can be carried forward over five years. Above 2 million euros in gifts the rate falls to 40%, which rarely matters to a small business.
Patronage is not sponsorship. Patronage is a gift made without a counterpart of real value to the donor, and it earns the reduction. Sponsorship is a commercial arrangement in which the business buys visibility, and it is treated as an advertising expense instead, with no donation reduction. Mixing the two on one receipt is a classic error.
Who can issue receipts
Only associations that qualify as organisations of general interest can issue receipts that open the reduction. That means a non-profit purpose, disinterested management, activities that are not for the benefit of a restricted circle of people, and a field of activity such as charitable, educational, scientific, social, humanitarian, sporting, family or cultural work, or the promotion of national heritage. If you are not sure you qualify, you can ask the tax authority for a formal position, called a rescrit, and rely on the answer. Issuing receipts without qualifying can bring penalties, including a fine calculated on the amounts wrongly shown.
What the receipt must show
Use the official receipt model, form 11580. It records the association's identity and the legal basis of its eligibility, the donor's identity and address, the amount, the date and the form of payment, among other mentions. Keep copies. Receipts can be issued electronically, and many associations send them in one batch each year, before donors file their return, with a separate receipt on request.
What does not qualify
Membership dues that give members real benefits in return, such as services or a magazine. Check before issuing receipts for dues.
Payments for an event ticket, a purchase at a charity sale or a service. A gift with a real counterpart is not a pure gift.
Gifts where the benefit to the donor is more than symbolic. Counterparts must stay small relative to the gift.
Time given by volunteers. Time is not a donation, although a volunteer who waives reimbursement of expenses can sometimes treat the waived sum as a gift, subject to conditions.
Keeping it clean
Keep a numbered register of receipts, so you can reconcile what you issued with your accounts, and keep the supporting documents for the period during which the tax authority can audit, which is several years. A prompt thank-you that mentions the receipt costs nothing and is both good practice and good fundraising.
Where AI helps
Receipts, thank-you letters and the annual summary mailing are repeatable work. Mirage Cloud includes an agent scoped to French association governance and connects to Brevo for email, which suits the mailing side. The association's eligibility and the wording of each receipt need a human check, since a wrong receipt exposes the association, not the software.

How to Run an Association's Annual General Meeting

Mirage Cloud IA — Fri, 09 Oct 2026 08:36:33 +0000

The law of 1901 says very little about how an association must hold its meetings. Almost everything is left to the statutes, which is freedom for the founders and a trap for everyone who comes after. Votes get challenged, board elections get annulled and banks refuse to recognise new signatories because a meeting was convened badly or minuted carelessly, and a little discipline in the run-up prevents most of it.

Start with the statutes
Read them before anything else, because they answer most of the questions people ask. Who can convene a meeting and how, how much notice is required and by what means, what counts as a quorum, what majority each type of decision needs, whether members can vote by proxy and how many proxies one person can hold, whether remote or written voting is allowed, and who chairs. If the statutes are silent on something you need, do not improvise on the day. Propose a change at an extraordinary meeting or put the rule in your internal regulations.

A timeline that works
About eight weeks before: choose the date, book the room or the video link, update the membership list and confirm who has the right to vote, since some statutes require fees to be paid.
Four weeks before: send the convocation by a means the statutes allow, with the date, time, place or link, the agenda, a proxy form and the documents members need in order to vote: the activity report, the accounts, the budget and the draft resolutions.

One week before: collect proxies, count them and prepare the attendance sheet.

On the day: check the quorum, appoint a chair and a secretary, adopt the agenda, present the reports, vote on resolutions one by one, hold any elections and record the result of each vote.

Within days: finish the minutes, have the chair and the secretary sign them and circulate them to members.

Within three months: declare any change of officers, address or statutes through the association registry, and tell your bank.

Resolutions to prepare
Typically you will approve the activity report and the accounts, decide what to do with the result, adopt the budget, set membership fees, elect or renew board members and, where the statutes provide for it, give the board discharge. Put each decision on the agenda as a separate item with the text of the resolution.
A vote on something that was not on the agenda is the classic ground for a challenge.

Voting rules
Most statutes give one vote per member, with proxies capped. Know how abstentions and blank votes count, because it changes whether a motion passes.
If the statutes allow it, a secret ballot can be requested for elections.
For hybrid or remote meetings, check that the statutes allow remote participation and electronic voting, that you can identify who is voting and that secrecy is kept wherever it is required. If the statutes are silent, amend them before relying on remote voting.
The minutes
Good minutes are short and complete: the date and place, who convened the meeting, who attended and who was represented, the quorum, the agenda, a summary of the discussion on each item, the text of each resolution, the votes for, against and abstaining, and any elections. Sign them, number the pages and keep them in a minute book. Your bank, your funders and the registry will ask for them, and so will whoever takes over from you in five years.

Mistakes that cause trouble
A convocation sent late, incomplete or by a means the statutes do not allow.
The wrong body convening the meeting.
Votes on items that were not on the agenda.
No attendance sheet, so nobody can prove the quorum.
Minutes written weeks later, or never signed.
Board changes never declared, leaving the registry and the bank with former officers.

Statutes amended without the quorum or majority they require.
Where AI helps
The paperwork is repetitive and follows fixed formats, which suits an assistant: convocations, agendas, proxy forms, draft resolutions and minutes from your notes. Mirage Cloud includes an agent scoped to French association governance and connects to Gmail and Brevo for sending, which helps with both the drafting and the mailing. Treat what it produces as a draft. A defective convocation can undermine every decision taken at the meeting, so someone who knows your statutes should read it first.

Using AI to Screen Candidates in France: What Is Allowed in 2026

Mirage Cloud IA — Thu, 08 Oct 2026 04:54:14 +0000

A lot of recruiters heard that the EU AI Act's rules for hiring tools were pushed back to December 2027 and concluded that nothing needs doing until then. That reading is half right. The heavy obligations for high-risk systems have moved, but several rules that touch AI in recruitment apply today, and the CNIL has put recruitment among its inspection priorities for 2026.
What is actually delayed
Under the AI Act, systems used to recruit or select people, filter applications or evaluate candidates are classified as high risk. The Digital Omnibus, in force since 27 July 2026, moved the date from which the high-risk obligations for these uses apply from 2 August 2026 to 2 December 2027. That covers the documentation, risk management, logging and human oversight duties, and the duties of the businesses that use these tools, known as deployers.
What applies today
GDPR Article 22. People have the right not to be subject to a decision based solely on automated processing that significantly affects them, and rejecting a candidate qualifies. A human who simply clicks approve does not fix this. In its 2023 SCHUFA ruling the Court of Justice held that when an automated score plays a determining role in a decision, the scoring itself counts as the automated decision. If a tool cuts a thousand applications to a shortlist of ten, the shortlist is the decision.
Discrimination law. An algorithm that disadvantages people on a protected ground produces unlawful discrimination, and the employer is the one who answers for it, whatever the vendor's brochure says.
The ban on emotion recognition at work. Since 2 February 2025 the AI Act has prohibited systems that infer people's emotions in the workplace, except for medical or safety reasons. Video interview tools that claim to read confidence, stress or enthusiasm from faces or voices are the obvious example. If a product you use has a feature like that, switch it off or drop the product.
Transparency. Since 2 August 2026, people must be told when they are dealing with an AI system, such as a candidate chatbot, unless it is obvious.
Data protection basics. Information for candidates, retention limits as described in the companion article on recruiting, and, where the processing amounts to systematic evaluation of people, an impact assessment is often required.
What you will need by December 2027
As a deployer of a high-risk system you will have to use it according to the provider's instructions, give oversight to people with the competence and authority to override it, keep the logs the system generates, and inform workers' representatives and affected workers before putting it into use at the workplace. Candidates must also be told when a high-risk system is being used to help decide about them. Providers carry the heavier duties of documentation, testing and conformity, which is why the vendor conversation matters.
Questions to put to a vendor
Is the system classified as high risk, and what documentation can you provide today?
What data was it trained on, and how was bias tested?
Does it infer emotions, personality or similar traits at any point?
Where is candidate data processed, which sub-processors are involved and is there a data processing agreement?
Is candidate data used to train your models?
Can a recruiter see why a candidate was ranked where they were, and override the result?
Can decisions be logged and exported?
Making the human review real
A human review only counts if it can change the outcome. Give reviewers the authority and the time to read applications the tool marked down, sample rejected candidates regularly and record the cases where the person disagreed with the tool. If nobody ever overrides it, what you have is an automated decision with an extra click.
Low-risk uses
Drafting adverts, scheduling interviews, writing up your own interview notes with candidates informed, and a candidate FAQ assistant that discloses it is an AI all help without ranking anyone.
Where this leaves you
For a small business the sensible position is conservative: use AI to prepare and organise recruitment, keep a person responsible for the choice, and get the documentation habit in place before 2027. Using an assistant to draft an advert or an interview guide sits on the preparation side of that line, and ranking or rejecting people sits on the other. Mirage Cloud's HR agent, scoped to recruitment and French labour law, is useful on the first kind of task. For the second, apply this checklist to any tool you consider.

Recruiting in France: What You Can Say, Ask and Keep

Mirage Cloud IA — Thu, 08 Oct 2026 04:10:13 +0000

Recruitment feels informal, a few messages and a couple of interviews, but French law touches every stage of it. The advert, the questions you ask, the notes you take and the way you store the application file all have rules attached, and most of them are limits: what you may not say, what you may not ask and how long you may keep what you collect.
Writing the job ad
An advert must not discriminate. The list of protected grounds in the Labour Code is long and covers, among others, origin, sex, family situation, age, health, disability, religion, political opinion, union activity, appearance and name. In practice, wording that hints at a preferred profile carries risk, whether that is a young and dynamic team, a native speaker when the language is not a real requirement, or a gendered job title with no F/H indication.

Describe the job and its real requirements. Ask for the qualifications and experience the work needs rather than the profile you picture, state the contract type, location, working hours and applicable collective agreement, and make sure everything you say about the role is accurate, because it can be quoted back to you later.

Pay information: what is coming
Directive (EU) 2023/970 on pay transparency had to be transposed into national law by 7 June 2026. France missed the date. The government presented a bill to the Council of Ministers on 10 September 2026, and it is not yet law. As drafted it would require employers to give the pay or the pay range in the job advert, or before the interview when there is no advert, forbid asking candidates about their pay history and forbid clauses that stop employees from discussing their pay.
Until it is adopted those rules do not bind you in France, but the preparation is the same either way: decide a pay range for each role before you advertise. If you cannot say what a role pays, you are not ready to recruit for it. Check the bill's progress before relying on any detail of it.

Interview questions
The Labour Code limits what you can ask. Questions must serve one purpose, which is to assess the person's ability to do the job or their professional skills, and they must have a direct and necessary link with the post. That rules out pregnancy and plans for children, marital status, religion, political or union views and origins, and health beyond what the occupational physician checks later.
Write the questions in advance and ask every candidate the same core set. It gives you a fairer comparison and a record that you stayed on relevant ground. Background checks follow the same logic: a criminal record extract can be requested only where a specific text requires it for the role, and looking at a candidate's public profile is acceptable when it relates to their professional life rather than their private one.
Candidate data
Every CV is personal data, and the CNIL's recruitment guide of January 2023 is the practical reference. Collect only what you need to assess the application. Tell candidates what you do with their data, who sees it and how long you keep it, when you first collect it. When you find someone through a professional network and contact them first, the same duty applies at that first contact, and at the latest within a month.
On retention, the CNIL's position is that data on unsuccessful candidates should not be kept for more than two years after your last contact with them, and only if they have been told. Keeping an application in a talent pool beyond the recruitment it was sent for is a separate choice that candidates should agree to. Then do the deleting: a recruitment mailbox that has never been cleared becomes a store of unlawful data after two years.
Candidates can also ask for a copy of what you hold about them, which includes your interview notes. Write notes you would be comfortable showing the person they describe.
A short checklist
Review job adverts for discriminatory wording and gendered titles.
Set a pay range for every role now.
Prepare a standard question set linked to the job.
Add a privacy notice to your application page or address.
Set a two year deletion rule and put the date in your calendar.
Record who has access to candidate files.
Where AI helps
Drafting adverts, structuring interview guides and turning role requirements into a scoring grid are low-risk uses of an assistant. Mirage Cloud includes an HR agent scoped to recruitment and French labour law, with a PayFit integration for the payroll side. Using AI to rank or reject candidates is a different matter with its own rules, covered in the companion article on screening candidates with AI.