AI agent for accessible communication: plain language, FALC project, translation and voice synthesis
Part of the public remains cut off from public information: people who do not speak French, people with a cognitive disability, people uncomfortable with the written word. This agent carries four distinct capabilities — a plain-language version, a draft transcription following the FALC easy-read rules and submitted for human validation, multilingual translation, and voice synthesis. Hosted in France — on local inference or an isolated resource — the data stays under control. The public officer reviews and approves before publication.
Updated on
Version ready for review — you approve before publication.
⛓ Source · your website content and the easy-read reference framework
Translation and audio to be reviewed by your department before publication.
✎ Action · drafts ready for review — the public officer decides
For a local authority, a central government department or a public body, a Blue Lemon Agent agent prepares four things that nothing allows to be confused: a plain-language version, a draft transcription following the FALC easy-read rules submitted for human validation, multilingual translation and voice synthesis. It runs on local inference or is hosted in France: no data is exposed to a foreign service, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Every output is reviewed and approved by your departments before publication. The aim: equal treatment and continuity of public service, leaving nobody out. Your public-sector staff write to it from Microsoft Teams, Slack or their email, and your users reach it on WhatsApp Business, the website chat or email — with no account to create and nothing to install. Reaching the administration from the tool people already have means less non-take-up of rights and equal access to the service. 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.
Making public information accessible to everyone — an obligation and a matter of equality
Part of the public remains cut off: people who do not speak French, people with a cognitive or visual disability, people uncomfortable with administrative writing. Yet public information has to be understood by everyone, with no breach of equality.
! The issue
Departments want to make their content accessible — simplified versions, translations, audio — but the work is slow, manual and rarely kept up to date with every regulatory change. And most consumer AI tools amount to entrusting content and the public's data to a third party, often hosted outside Europe and subject to the Cloud Act.
✓ Our answer
AI is only of interest to a public service if it is sovereign and confidential by design. Local inference or an isolated resource hosted in France, systematic human oversight, approval reserved to the public officer before any publication: the accessibility gained is never paid for in lost sovereignty. The aim is not to replace the public officer, but to give them back time to serve every member of the public.
Sovereignty & compliance: the non-negotiable condition for public data
A public service handles information belonging to everyone, including the most vulnerable. Here is how the architecture of our agents protects it, content by content.
Local inference
The agent can run on a machine belonging to the authority: no content 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
The data: architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
One isolated resource per entity
No pooling of data: an environment strictly dedicated to your authority or institution.
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 automatic publication; 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.
· Three standard letters exceed the reading level you set. They have been sent 1,240 times since January.
· One page changed in French on 12/06; its four translations did not follow. They have since been announcing a deadline that no longer exists.
· The audio of one page was generated from a version predating 03/05. It says something different from the text right above it.
· The site's accessibility statement has not been updated since 2023, although the site has changed twice since. morning-watch_4-flags.pdf4 flags · volumes concerned
⛓ Source · standard letters, page versions, audio files, published statement
· Notification of refusal of support — 612 sent. Sentences of 34 words on average, four subordinate clauses in the first.
· Request for further documents — 428 sent. The document requested appears in the third paragraph.
· Appointment summons — 200 sent. The date comes after six lines of regulatory preamble.
What I can do: a reworded version of all three, structure unchanged.
What I do not touch: the statements of appeal routes and deadlines. They are not simplified — their wording is what makes them enforceable, and a clearer letter whose appeal wording has been rewritten is a weaker letter.
All three versions are ready, showing what changed and what did not, line by line. 3-standard-letters_reworded.pdf1,240 sent · what changes and what stays
⛓ Source · 3 standard letters, sending volumes, readability threshold you set
Routing follows your organisation: the letters to the department that issues them, not to communications — the service department answers for what they say; the out-of-step translations to the editorial lead; the audio to whoever produced it; the statement to the accessibility officer and the chief executive, because it is published and dated.
With a chase: 48 h on the translations and the audio — they say something false today; 7 days on the rest. Then a monthly summary to the accessibility officer: by material, never by author.
What this morning has already produced: the three most-sent letters in the council — 1,240 mailings since January — rewritten, with what was changed and what was not, line by line; four translations and one audio file realigned on the French version of 12/06; the accessibility statement brought back to its date.
What that gives the service from tomorrow: a refusal notice that is understood on first reading, a requested document that appears on the first line — that many calls to the front desk that no longer need to happen, and officer time given back to the counter, and deadlines met because the resident sends the right document first time. The same letter says the same thing to everyone, in every language and on every medium: that is equal treatment, and it can be evidenced. The appeal routes and time limits stay word for word as you settled them — their exact wording is what makes them enforceable, and I carry them across intact into every version. Nothing leaves the service: I reach what you open, role by role, every read is logged, and you close it again on a word.
The next step is ready: approve the three letters and I will rank your other templates by mailing volume and reading level. You will have the next twenty today.
✎ Proposal · watch and chases to be configured — you set the thresholds
Unchanged, word for word: the legal ground for the refusal, the reference to the resolution, and the statement of appeal routes and deadlines. Those three blocks are copied, not reworded.
Reworded: everything else. The opening sentence goes from 34 words to 11. The ground for refusal, which appeared in the fourth paragraph, now comes second — it is what the resident is looking for, and it is what they could not find.
One point I flag without correcting it: the current letter says "your application could not be granted". That wording is gentle and it is ambiguous — some residents read it as a postponement, judging by the six follow-ups received by the department.
The proposed version says "your application is refused". That is harder, and it is what the decision says. refusal-letter_before-after.pdf3 blocks unchanged · 1 wording flagged
⛓ Source · standard letter in force, resolution cited, follow-ups received
What is measured: six residents came back to the department believing their application was still pending. Out of 612 letters, six is few — and they are six people who waited for nothing, two of whom let the appeal deadline pass.
What the gentleness costs: not average comprehension, but comprehension by the most vulnerable. That is precisely the audience simplification is for.
Three wordings are offered, from the most direct to the most cushioned, each with what it leaves ambiguous.
What is yours, and nothing else: the choice between the three. The tone of a refusal letter shapes the authority's relationship with its residents, and that is not decided on a readability criterion — but it is decided in five minutes, with the three wordings and their ambiguities in front of you.
And an observation: across your eleven standard letters, seven use an indirect wording to announce an unfavourable decision. The list is attached, with no recommendation. standard-letters_7-indirect-wordings.pdf11 letters · 7 indirect wordings
✎ Support · 3 wordings offered — the tone stays the authority's decision
Preparing a draft transcription under the easy-to-read (FALC) rules: I take your standard award letter — 640 words, 4 cross-references to legal articles, 3 sentences over 40 words — and I prepare a draft: one idea per sentence, an everyday word in place of each administrative term, the procedure in numbered steps, dates written out in full, and every legal term kept and then explained alongside rather than replaced. The draft runs to 310 words and covers the same 11 pieces of information: I set them out side by side, one by one, so that nothing is lost unseen.
What that draft is not yet, and the word matters: the FALC rules require a transcription and then a review involving the people concerned. Until that review has taken place, the text remains a draft subject to human validation, and we do not call it FALC. Your department runs the session; I supply the draft, the list of 11 pieces of information and the 6 points where I hesitate between two wordings. Of the 14 drafts already reviewed, 9 came back changed, and 21 of the changes were on a word I had taken for everyday.
What that gives back: preparing a transcription by hand takes 3 hours; reading over a draft prepared under the rules takes 40 minutes, validation session included. Across your 60 documents a year, more than 138 hours given back, rounded down.
Voicing, and when it happens: speech synthesis is produced only from an already validated version — voice, pace and language chosen with you — and the file is listened to and approved by your department before it goes online. That is exactly what was missing on the page whose audio predates 3 May: the voiced version had outlived the update to the text and said something other than the sentence displayed just above it. From now on, any change to the text invalidates its voiced version and takes it down, and the list of recordings to redo reaches you the same day: 4 pages are in that state this morning.
⛓ Sourced · 640 words down to 310 for 11 pieces of information, 9 of 14 drafts changed in review, 4 recordings to redo
What changed: a single sentence, about the assessment deadline. The rest of the page is identical across all five versions.
What that produces: a resident reading English, Spanish, Arabic or Turkish expects two months and gets an answer in six weeks — which, this time, is in the right direction. Had the deadline been lengthened, it would be the reverse.
What is missing before publishing, and it is a single act: the human review. Information published in a language binds the authority in that language, and an unreviewed translation is unapproved information.
All four corrected versions are ready — one sentence each — awaiting review.
What I can add: the gap has lasted 57 days, and nothing in your publishing chain would have flagged it. That is the real subject. page_5-versions-out-of-step.pdf1 sentence · 4 languages · 57 days
⛓ Source · page versions in 5 languages, change history
What that covers:
· 4 pages where only a date or a deadline diverges.
· 5 pages where a whole paragraph is missing from the translations.
· 2 pages where the French version was withdrawn and the translations were not — they are still online.
The last two are the most serious: the authority is publishing, in four languages, a page it judged it had to withdraw in French.
What I can do, on your approval: on every change to a French page, flag its translated versions to the editorial lead, with the changed sentence isolated. No translation is published without review — but nobody can say any longer that they did not know. translated-pages_11-out-of-step.pdf43 pages · 11 out of step · 2 withdrawn in French only
✎ Proposal · flagged on every change — publication stays reviewed
What it announces: a compliance rate, a list of inaccessible content, and an audit date — all three relating to the 2023 version of the site.
What has changed since: two redesigns, a new search engine, and 214 new pages.
What I can do: record the technically observable gaps on the current site — contrast, image alternatives, heading structure, form labels — and tell you which are measurable and which are not.
What stays with the auditor, and why it has to be them: the compliance rate and the statement. An accessibility audit binds the authority — a figure produced by a tool is not an audit figure, and it stands up neither before a resident nor before the administration. What I hand them is the technical record already made across the 214 new pages: that many hours they will not have to spend.
What the record gives you: enough to prepare the audit, not enough to avoid it. technical-gaps_record.pdf4 families of gaps · measurable and not
⛓ Source · statement published in 2023, current site, redesign history
What a tool observes: insufficient contrast, an image with no alternative, a skipped heading level, a field with no label. I recorded 1,847 occurrences across 214 pages.
What no tool observes: whether an image's alternative describes what matters; whether the reading order makes sense; whether a complex table stays comprehensible aloud; whether a long form remains usable for someone navigating by keyboard.
The first point is the most misleading: 412 images do have an alternative, so no tool flags them. I read 40 at random: 17 say "image" or repeat the file name. Technically compliant, practically useless.
What I propose, and it is honest: the full technical record, plus the 40 alternatives read, stating that it is a sample. The rest belongs to an audit, and to testing with the people concerned — which I do not replace.
✎ Support · technical record + sample read — the audit remains to be run
Your case is not here? That is exactly what a 15-minute conversation is for. Book the free audit →
Four distinct capabilities, and what each one covers
Plain language, the FALC project, translation and voice synthesis do not carry the same cost, the same risk or the same responsibility. Each is therefore named on its own. All of them work in support, subject to approval by your departments.
Preparation of a plain-language version
Rewriting a page, a letter or a form in short sentences, one idea per sentence, acronyms spelled out, no jargon. Your department reviews and approves before publication.
Preparation of a draft transcription under the FALC easy-read rules, submitted for human validation
The agent prepares a DRAFT, and nothing more. The method requires a transcription and then a review involving the people concerned: without that step the text is not FALC, and we do not call it that.
Multilingual translation
Translating your information and your exchanges with people who do not speak French, including live at the counter. The languages, volumes and formats are settled during the audit and written into the quotation.
Voice synthesis (text-to-speech)
Producing the spoken version of a piece of content already approved: voice, pace and language chosen with you. Your department listens to the file and approves it before publication.
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.
Help with online procedures
To support the filling in of online services, a dedicated agent takes it on — using your translated and easy-read content.
Support with online procedures from 780 € incl. VAT / month Help with procedures →Multichannel information
To answer everyday questions 24/7, by phone or online, a dedicated agent takes over.
Multichannel administrative information from 700 € incl. VAT / month Multichannel information →Multilingual support agent (2–3 languages)
Multiple languages, channels.
Multilingual support agent (2–3 languages) from 654 € incl. VAT / month Discover the agent →Multilingual international support agent
Several languages and channels.
Multilingual international support agent from 664 € incl. VAT / month Discover the agent →One-stop-shop adviser across public bodies
Answering on the procedures of several public bodies, preparing the support and handing the person a summary of their documents: that agent staffs the counter, this one makes its content accessible.
One-stop-shop adviser across public bodies from 765 € incl. VAT / month Discover the agent →In 15 minutes we identify the agent that will give your staff the most time back — without oversizing the project.
How much time can a service win back?
By preparing plain-language versions, FALC drafts, translations and spoken content, a department shifts its work towards review — and the time won goes back into the quality of the service delivered to everyone.
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
One agent carrying the four capabilities — plain language, FALC project, translation, voice synthesis — installed and operated for you. The languages, volumes, formats and limits are settled during the audit and written into the quotation.
Four guarantees that matter to a public service
Your questions, our answers
Can the FALC draft be published as it stands?
How many languages are covered, and for what volumes?
Does this make my website accessible in the sense of the RGAA?
Is the public's data protected?
Do we have to change our tools or our existing website?
Does the agent state that it is an artificial intelligence?
How long does it take to deploy an agent?
Which tools can users use to reach the agent?
Other solutions serving the public
Let us estimate the accessibility potential in your departments
15 minutes to identify the use case most useful to your public — hosted in France, supervised, with no commitment.