● Public sector — Social housing
AI agent for social housing
An assistant that pre-assesses housing applications, prepares the panel files and answers applicants — the allocation panel decides. Hosted in France, on a resource dedicated to and isolated for each organisation, applicants' data stays under French law. The AI agent assists, the panel allocates.
Hosted in France
Applicants' data protected
GDPR & AI Act: governed deployment
Human oversight
Updated on
Deployed in a few weeks
Social housing assistant · hosted in France
Check the eligibility and completeness of application no. 3512 before registering it.
Application checked. Income below the ceilings for the size of property requested, documents present: tax assessment, proof of income, identity document. Missing: the certificate of accommodation mentioned in the form.
Letter requesting the document prepared; the file stays as it is in the national register.
⛓ Source · the applicant's file + the income ceilings in force
Letter requesting the document prepared; the file stays as it is in the national register.
⛓ Source · the applicant's file + the income ceilings in force
Prepare the summary notes for the three applications proposed for the three-room flat at Les Érables, Tuesday's panel.
Notes ready. For each application: household composition, income and rent-to-income ratio, how long the application has been waiting, the priority criteria that apply. A strictly factual presentation, with no ranking: the panel considers and allocates.
The meeting file is assembled, ready for review.
✎ Action · notes ready — the allocation panel decides
The meeting file is assembled, ready for review.
✎ Action · notes ready — the allocation panel decides
Local inference · no data outside the EU
Data hosted in France
Sovereign by designLocal inference or hosting in France
GDPR & AI Act: governed deploymentTraceability & human oversight
TurnkeyDesigned, installed and operated for you
The panel decidesThe AI agent assists, it never awards alone
✦ In brief
At a social landlord or in a housing department, a Blue Lemon Agent agent assists the teams with the repetitive tasks — pre-assessment of applications, checking eligibility and completeness (national register, documents, ceilings), allocation panel files, replies to applicants. It runs on local inference or is hosted in France on a resource dedicated to and isolated for each organisation: applicants' data is never exposed to a foreign service, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. The AI agent does not rank applicants; the allocation remains the panel's. The time won back goes to supporting applicants. Live within a few weeks. Your public-sector staff write to it from Microsoft Teams, Slack or their email, and applicants 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.
100%
hosted in France in the target architecture
0
transfer outside the EU in the target architecture
7
uses ready to deploy on this scope
0
decision taken without human approval
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.
The context
Why AI matters to social housing — and why they hesitate
Waiting lists growing longer, files to check document by document, panels to prepare: social housing is buried in administration, while every file contains highly sensitive data — income, household composition, sometimes situations of distress.
! The issue
The social housing officer is caught between applicants who have sometimes been waiting for years, and ever more files to check, summarise and present to the panel. Yet most consumer AI tools would amount to entrusting applicants' income, household composition and social circumstances to a third party, often hosted outside Europe and subject to the Cloud Act.
✓ Our answer
For social housing, AI is only of interest if it is sovereign and confidential by design. Local inference or an isolated resource hosted in France, systematic human oversight, allocation reserved to the panel — the AI agent never ranks applicants: the time saved on pre-assessment is paid for neither in lost confidentiality nor in unequal treatment. The aim is not to replace the teams, but to give them back time for applicants.
The decisive point
Protecting applicants' data: sovereignty & compliance
A housing application file tells the story of a household: income, family, difficulties. Here is how the architecture of our agents protects that data, organisation by organisation.
Local inference
The agent can run on a machine at the organisation: no data leaves the network, nothing passes through a cloud.
Hosting in France
Otherwise, a dedicated and isolated resource, hosted in France under French law — applicants' data: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
Exposure of applicants' data to the Cloud Act and FISA 702 is reduced by design; location alone does not guarantee immunity.
One isolated resource per organisation
No pooling of data: an environment strictly dedicated to your organisation, guaranteeing the continuity of the public service.
Encryption & controlled access
Encryption in transit and at rest, role-based access (RBAC), strong authentication and logging.
AI Act: governed deployment
The agent is strictly in support; no allocation and no automated ranking of applicants; 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.
For the most sensitive data, SecNumCloud and HDS options are available — particularly relevant as soon as social or health data is processed. A single architecture is designed to answer both the GDPR and extraterritorial exposure. Designed for deployment in line with the GDPR and the AI Act, after the processing, roles and context-specific risks have been assessed.
Demonstration
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.
The public body in this demonstration
Fictional public bodyHabitat des Trois Vallées — public housing office of an 11-town urban authority (fictional organisation)
- Sector
- Public housing office: 8,400 homes under management, an applications and allocations department, tenancy management, move-ins and move-outs; 26 allocation committee sittings a year, two a month except August
- Headcount
- 96 staff, 7 of them in applications and allocations — 5 caseworkers, one head of service and one apprentice; the same people assess files, prepare the sittings and answer applicants
- Public served
- 5,900 live applications on the national registration system, 8,400 households housed, 1,060 allocations a year; 3,180 applications examined in committee, three per home offered
- Order of magnitude
- 4,700 new or renewed applications to assess each year, 14,200 applicant enquiries across four channels, 26 committee sittings averaging 122 applications, 1,060 move-ins and as many administrative files to assemble
- Tools in place
- National registration system, tenancy management software, supporting-document repository, the office's online renewal form, e-mail and calendars — the agent plugs in read-only, nothing is replaced and nothing is migrated
- Who decides
- The allocation committee allocates; the head of applications and allocations settles the sitting file; caseworkers validate every pre-assessment; the chief executive signs the document-request letters
- Points for improvement
- 1,312 of the 5,900 live applications are incomplete — 22 %; an applicant waits 11 days for a written answer on the state of their file; 604 applications were struck off last year for want of renewal; and preparing one committee sitting takes 9 hours for 122 applications
At Habitat des Trois Vallées, seven people assess 4,700 applications a year, prepare 26 committee sittings and answer 14,200 enquiries from households that have sometimes been waiting for years. The exchanges below span a full year, from the review of the national registration backlog to the report presented to the board. The agent runs locally on a machine belonging to the office: incomes, household compositions and social circumstances never leave the network. The committee allocates; the agent prepares.
This public body, its figures and the exchanges that follow were invented for the demonstration. They illustrate a common situation; they describe no real service.
Social housing assistant · hosted in France
I have reviewed your 5,900 live applications on the national registration system, and the backlog has a shape nobody had the time to draw.
What the count says:
· 1,312 applications out of 5,900 are incomplete — 22 %. They hold a place in the queue without being able to reach committee.
· Those 1,312 files hinge on five documents only. Tax notice missing or too old: 438. Income evidence incomplete: 361. Proof of accommodation absent although the form refers to it: 214. Identity document or residence permit expired since filing: 187. Household composition not updated after a declared birth or separation: 112.
· 604 applications were struck off last year for want of renewal — and matching the dates, 431 of them were complete and had nothing left to do but click.
The item you can check against your own logs: an application spends 25 minutes in a caseworker's hands, 60 % of it — 15 minutes — on eligibility and completeness checking alone: income ceilings, documents present, consistency of the declarations. Across the 4,700 new or renewed applications of the year, that is 1,175 hours spent checking documents.
What I have already done, overnight: the 5,900 applications are matched against the income ceilings as your configuration holds them on 1 January 2026, and the 1,312 document-request letters are written, each naming the missing document, the point on the form that calls for it, and how to file it.
The next step, and it takes an hour: you read five letter templates — five, not 1,312. The head of service settles the wording, the chief executive signs, and the 1,312 letters go out the same day. application-backlog_5900-live-1312-incomplete.pdf1,312 incomplete files on five documents, 1,175 hours of checking measured
⛓ Sourced · 5,900 live applications, income ceilings as configured on 1 January 2026, 12 months of strike-off logs
What the count says:
· 1,312 applications out of 5,900 are incomplete — 22 %. They hold a place in the queue without being able to reach committee.
· Those 1,312 files hinge on five documents only. Tax notice missing or too old: 438. Income evidence incomplete: 361. Proof of accommodation absent although the form refers to it: 214. Identity document or residence permit expired since filing: 187. Household composition not updated after a declared birth or separation: 112.
· 604 applications were struck off last year for want of renewal — and matching the dates, 431 of them were complete and had nothing left to do but click.
The item you can check against your own logs: an application spends 25 minutes in a caseworker's hands, 60 % of it — 15 minutes — on eligibility and completeness checking alone: income ceilings, documents present, consistency of the declarations. Across the 4,700 new or renewed applications of the year, that is 1,175 hours spent checking documents.
What I have already done, overnight: the 5,900 applications are matched against the income ceilings as your configuration holds them on 1 January 2026, and the 1,312 document-request letters are written, each naming the missing document, the point on the form that calls for it, and how to file it.
The next step, and it takes an hour: you read five letter templates — five, not 1,312. The head of service settles the wording, the chief executive signs, and the 1,312 letters go out the same day. application-backlog_5900-live-1312-incomplete.pdf1,312 incomplete files on five documents, 1,175 hours of checking measured
⛓ Sourced · 5,900 live applications, income ceilings as configured on 1 January 2026, 12 months of strike-off logs
There are five of us assessing files and the next committee is in nine days. Nobody here has a week to reopen 1,312 files. Where do we start?
With 289 files out of 1,312 — and I have ordered them by what they unblock, not by date of arrival.
What the match against your 26 sittings of last year says:
· 289 of the 1,312 incomplete files concern a home type and town you allocated at least ten times last year. Those are the ones that will reach committee soonest once complete.
· Of those 289, a single document is missing in 251 cases. One letter, one document, one file back in play.
· The other 1,023 clear at 80 a week without anything waiting: they are written, they are queued, they go out at the pace you set.
And I advise against sending all 1,312 letters on the same day, with figures: your observed return rate on document requests is 61 % at three weeks. 1,312 letters at once means roughly 800 returns inside the same window — 800 documents to read and file for five people. 289 then 80 a week spreads the load over four months and brings 1,312 complete files into the queue without ever swamping the service — and the first effect lands at the committee a fortnight from now.
What that hour of reading gets you: five letter templates already drafted, each with the document named, the line of the form that requires it, the filing deadline and the three ways to file — counter, online form, post; for a plain-language version, the content to pass to the Accessibility, Easy Read and Translation agent — a dedicated agent, and it is not part of what this agent includes; and the number of files each template unblocks: 438, 361, 214, 187, 112. You approve, or you change a word. The writing and the matching are done: the decision takes an hour instead of a week, and it stays the service's own — which is what makes it defensible before an applicant who challenges it.
The next step I propose: that I check every night the applications filed or amended the day before, and write the missing-document letter within twelve hours of filing. A file will no longer sit incomplete for six months unnoticed: it will be known the next morning. missing-documents_5-templates-1312-letters.pdf5 templates to read, 289 files prioritised, 61 % observed return rate
⛓ Sourced · 1,312 incomplete files cross-checked with the year's 1,060 allocations, observed return rate on document requests
What the match against your 26 sittings of last year says:
· 289 of the 1,312 incomplete files concern a home type and town you allocated at least ten times last year. Those are the ones that will reach committee soonest once complete.
· Of those 289, a single document is missing in 251 cases. One letter, one document, one file back in play.
· The other 1,023 clear at 80 a week without anything waiting: they are written, they are queued, they go out at the pace you set.
And I advise against sending all 1,312 letters on the same day, with figures: your observed return rate on document requests is 61 % at three weeks. 1,312 letters at once means roughly 800 returns inside the same window — 800 documents to read and file for five people. 289 then 80 a week spreads the load over four months and brings 1,312 complete files into the queue without ever swamping the service — and the first effect lands at the committee a fortnight from now.
What that hour of reading gets you: five letter templates already drafted, each with the document named, the line of the form that requires it, the filing deadline and the three ways to file — counter, online form, post; for a plain-language version, the content to pass to the Accessibility, Easy Read and Translation agent — a dedicated agent, and it is not part of what this agent includes; and the number of files each template unblocks: 438, 361, 214, 187, 112. You approve, or you change a word. The writing and the matching are done: the decision takes an hour instead of a week, and it stays the service's own — which is what makes it defensible before an applicant who challenges it.
The next step I propose: that I check every night the applications filed or amended the day before, and write the missing-document letter within twelve hours of filing. A file will no longer sit incomplete for six months unnoticed: it will be known the next morning. missing-documents_5-templates-1312-letters.pdf5 templates to read, 289 files prioritised, 61 % observed return rate
⛓ Sourced · 1,312 incomplete files cross-checked with the year's 1,060 allocations, observed return rate on document requests
All of that assumes you read tax notices, household compositions, sometimes emergency accommodation certificates. Where does that data go?
Nowhere. I run in local inference on a machine belonging to the office, and what I read does not leave your network.
Local inference means the model computes on your machine: the contents of a tax notice or an accommodation certificate cross no external network to be processed. If the office would rather not run a machine, the other route is a dedicated, isolated resource hosted in France, reserved for Habitat des Trois Vallées, shared with no other organisation.
What that changes, point by point:
· Applicant data trains no model, neither ours nor a third party's.
· I work read-only, and the technical account I read through has no write permission on the national registration system — that is verifiable in one command, which is worth more than a promise.
· Encryption in transit and at rest, role-based access — rights follow the job: a caseworker opens the files of her patch, the head of service opens the sitting files, technical staff open no applicant file at all. 7 roles for your 7 staff, and the log shows 0 out-of-role access since go-live.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, including against a US provider hosting in Europe — location alone not being enough to guarantee immunity, the exposure is documented for the configuration chosen.
· A complete log: who asked what, when, and what the system produced — that is the document you will produce if an applicant exercises their right of access.
On the European framework, and I give you the text rather than my opinion: Annex III of the European AI Regulation places systems used to evaluate access to essential public services and essential benefits among high-risk systems — high-risk, not prohibited, and the corresponding obligations apply from 2 December 2027 following the deferral carried by the Digital Omnibus package. What is already in force and what we meet today: Article 22 of the General Data Protection Regulation, which gives every person the right not to be subject to a decision producing legal effects based solely on automated processing — and that is exactly how you are organised: the committee allocates, the agent prepares. And since 2 August 2026, Article 50 of the AI Regulation requires that anyone interacting with an AI system be informed: I introduce myself as a digital assistant of the office in the first sentence, and the applicant can ask for a person at any moment.
The figure that sums it up: 0 applicant record out of the office network across 5,900 files reviewed, and 0 transfer outside the European Union.
What I propose: that I keep up to date the register entry your data protection officer will ask for — purposes, data processed, retention periods, who accesses what. The first version is written and attached; it updates itself as each new use is opened. technical-framework_where-applicant-data-lives.pdfLocal inference, read-only, 0 transfer outside the EU, 7 roles
✎ Framework · deployment architecture, technical account rights, first version of the register entry
Local inference means the model computes on your machine: the contents of a tax notice or an accommodation certificate cross no external network to be processed. If the office would rather not run a machine, the other route is a dedicated, isolated resource hosted in France, reserved for Habitat des Trois Vallées, shared with no other organisation.
What that changes, point by point:
· Applicant data trains no model, neither ours nor a third party's.
· I work read-only, and the technical account I read through has no write permission on the national registration system — that is verifiable in one command, which is worth more than a promise.
· Encryption in transit and at rest, role-based access — rights follow the job: a caseworker opens the files of her patch, the head of service opens the sitting files, technical staff open no applicant file at all. 7 roles for your 7 staff, and the log shows 0 out-of-role access since go-live.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, including against a US provider hosting in Europe — location alone not being enough to guarantee immunity, the exposure is documented for the configuration chosen.
· A complete log: who asked what, when, and what the system produced — that is the document you will produce if an applicant exercises their right of access.
On the European framework, and I give you the text rather than my opinion: Annex III of the European AI Regulation places systems used to evaluate access to essential public services and essential benefits among high-risk systems — high-risk, not prohibited, and the corresponding obligations apply from 2 December 2027 following the deferral carried by the Digital Omnibus package. What is already in force and what we meet today: Article 22 of the General Data Protection Regulation, which gives every person the right not to be subject to a decision producing legal effects based solely on automated processing — and that is exactly how you are organised: the committee allocates, the agent prepares. And since 2 August 2026, Article 50 of the AI Regulation requires that anyone interacting with an AI system be informed: I introduce myself as a digital assistant of the office in the first sentence, and the applicant can ask for a person at any moment.
The figure that sums it up: 0 applicant record out of the office network across 5,900 files reviewed, and 0 transfer outside the European Union.
What I propose: that I keep up to date the register entry your data protection officer will ask for — purposes, data processed, retention periods, who accesses what. The first version is written and attached; it updates itself as each new use is opened. technical-framework_where-applicant-data-lives.pdfLocal inference, read-only, 0 transfer outside the EU, 7 roles
✎ Framework · deployment architecture, technical account rights, first version of the register entry
Prepare the summary notes for the three applications shortlisted for the three-bedroom flat at Les Érables, Tuesday's committee.
The three notes are ready, and they carry exactly the same grid of facts, in the same order.
For each of the three applications: household composition and its dated evidence · household income and the rent-to-income ratio computed on the rent and charges of that flat · length of the application since the unique registration number · standing against the priority criteria your committee's rules of procedure set out in their 12 January 2026 version · score derived from the scheme adopted by the urban authority, with the points broken down line by line · documents on file and the date of each.
What I did beyond that, and which manual preparation did not allow: I recomputed the score of the three applications and I show you the arithmetic, line by line. Across the 3,180 applications put to committee last year, that recomputation surfaced 41 discrepancies against the manually entered score — 29 against the applicant, 12 in the applicant's favour. All 41 are listed, each with the scheme line concerned and the document on file that supports it. A scoring discrepancy seen before the sitting is an appeal avoided after it.
What the note does not contain, and that is a rule rather than a limitation: no ranking, no order of merit, no recommendation. The three notes are interchangeable in the order you lay them on the table. The committee weighs and allocates; that is what makes its decision stand.
The time: your typical sitting takes 9 hours of preparation for 122 applications. Tuesday's sitting file is assembled; 1 h 48 of reading remains — 20 % — and the file goes to committee. Across 26 sittings, that is 187 hours returned to the service.
The next step I propose: that I assemble the sitting file as soon as the list of homes to allocate is settled, without waiting to be asked. You would open a complete file on Monday morning instead of building it on Friday evening. sitting-file_les-erables-3-applications.pdfThe same grid of facts three times, no ranking, 9 h → 1 h 48
⛓ Sourced · the 3 applications for the Les Érables flat, committee rules of procedure of 12 January 2026, scoring scheme adopted by the urban authority
For each of the three applications: household composition and its dated evidence · household income and the rent-to-income ratio computed on the rent and charges of that flat · length of the application since the unique registration number · standing against the priority criteria your committee's rules of procedure set out in their 12 January 2026 version · score derived from the scheme adopted by the urban authority, with the points broken down line by line · documents on file and the date of each.
What I did beyond that, and which manual preparation did not allow: I recomputed the score of the three applications and I show you the arithmetic, line by line. Across the 3,180 applications put to committee last year, that recomputation surfaced 41 discrepancies against the manually entered score — 29 against the applicant, 12 in the applicant's favour. All 41 are listed, each with the scheme line concerned and the document on file that supports it. A scoring discrepancy seen before the sitting is an appeal avoided after it.
What the note does not contain, and that is a rule rather than a limitation: no ranking, no order of merit, no recommendation. The three notes are interchangeable in the order you lay them on the table. The committee weighs and allocates; that is what makes its decision stand.
The time: your typical sitting takes 9 hours of preparation for 122 applications. Tuesday's sitting file is assembled; 1 h 48 of reading remains — 20 % — and the file goes to committee. Across 26 sittings, that is 187 hours returned to the service.
The next step I propose: that I assemble the sitting file as soon as the list of homes to allocate is settled, without waiting to be asked. You would open a complete file on Monday morning instead of building it on Friday evening. sitting-file_les-erables-3-applications.pdfThe same grid of facts three times, no ranking, 9 h → 1 h 48
⛓ Sourced · the 3 applications for the Les Érables flat, committee rules of procedure of 12 January 2026, scoring scheme adopted by the urban authority
Honestly, if you can recompute the score you can rank the files. Why not just give us the order?
I can do it, and it is not caution that stops me — it is a division of roles you are free to move, and which I recommend you do not move. Here is why, with figures.
What I already do, and it is arithmetic: the score is a scheme your urban authority adopted. I apply it line by line, across 3,180 applications, and I show every point with the document that supports it. It is reproducible: the same application run again in six months returns the same total, and the arithmetic prints out.
What is of a different nature: an order of priority between three households is not arithmetic, it is an appraisal — and that is what the committee exists to carry. Article 22 of the General Data Protection Regulation gives every person the right not to be subject to a decision producing legal effects concerning them and based solely on automated processing. Allocating a home produces legal effects. A ranking I produced and the committee endorsed is exactly what that provision addresses — and it is the first ground an applicant would raise before the administrative court.
The route that gets you what you are after, and it is open: if the committee wants to see the applications in an order, I display them in the order of the scheme it adopted itself, showing the arithmetic behind each total. That is not my appraisal, it is your own rule made legible — and the committee stays free to depart from it, giving reasons, which the note provides for in plain words. You need only ask: it is a display setting on the sitting file, settled by the head of service, dated, and withdrawn with a word.
What this division of roles earns you, and it is measurable: of last year's 1,060 allocations, 23 drew a written challenge. In 19 cases the question asked was « why them and not me ». With an identical grid of facts and the scoring arithmetic printed, the answer fits on one page and is produced the day it is asked, instead of three weeks of reconstruction. That is time returned and litigation risk down — not a door closed.
The next step: I propose putting to the next board meeting an explicit mention, in the rules of procedure, of the display order and of the committee's right to depart from it. The draft resolution is written; it remains for you to put it on the agenda. scoring-check_41-discrepancies-in-3180-applications.pdf29 against the applicant, 12 in their favour, arithmetic printed
✎ Framework · adopted scoring scheme, Article 22 GDPR, 23 written challenges out of 1,060 allocations
What I already do, and it is arithmetic: the score is a scheme your urban authority adopted. I apply it line by line, across 3,180 applications, and I show every point with the document that supports it. It is reproducible: the same application run again in six months returns the same total, and the arithmetic prints out.
What is of a different nature: an order of priority between three households is not arithmetic, it is an appraisal — and that is what the committee exists to carry. Article 22 of the General Data Protection Regulation gives every person the right not to be subject to a decision producing legal effects concerning them and based solely on automated processing. Allocating a home produces legal effects. A ranking I produced and the committee endorsed is exactly what that provision addresses — and it is the first ground an applicant would raise before the administrative court.
The route that gets you what you are after, and it is open: if the committee wants to see the applications in an order, I display them in the order of the scheme it adopted itself, showing the arithmetic behind each total. That is not my appraisal, it is your own rule made legible — and the committee stays free to depart from it, giving reasons, which the note provides for in plain words. You need only ask: it is a display setting on the sitting file, settled by the head of service, dated, and withdrawn with a word.
What this division of roles earns you, and it is measurable: of last year's 1,060 allocations, 23 drew a written challenge. In 19 cases the question asked was « why them and not me ». With an identical grid of facts and the scoring arithmetic printed, the answer fits on one page and is produced the day it is asked, instead of three weeks of reconstruction. That is time returned and litigation risk down — not a door closed.
The next step: I propose putting to the next board meeting an explicit mention, in the rules of procedure, of the display order and of the committee's right to depart from it. The draft resolution is written; it remains for you to put it on the agenda. scoring-check_41-discrepancies-in-3180-applications.pdf29 against the applicant, 12 in their favour, arithmetic printed
✎ Framework · adopted scoring scheme, Article 22 GDPR, 23 written challenges out of 1,060 allocations
And the files where something does not add up — incomes that do not match the declarations, an address that comes up three times?
I found 63 across the year's 4,700 applications, and I am giving you the figure that flatters me and the one that does not, together.
What the matching flagged:
· 31 files where declared income differs by more than 15 % from the reference taxable income on the attached tax notice — in 24 cases that is a fall in income occurring after filing and never updated, which works in the applicant's favour and reopens eligibility for a home type they believed closed to them.
· 18 files where the declared household composition does not match the attached evidence — births, separations, shared custody.
· 14 files carrying the same accommodation address for households with no declared connection.
What the checking produced, and this is the unflattering figure: 38 of the 63 flags proved well founded after your caseworkers examined them. 25 were looked at for nothing. The gap comes down to one cause: in 19 of the 25, the attached tax notice was the previous year's because the new one was not yet available at the filing date — I was comparing two different years.
What I did about it, and it is already written: I tightened the rule so that it checks the tax notice year against the filing date before flagging. Run again over the 4,700 applications of the past year, the tightened version produces 44 flags instead of 63, keeps all 38 well-founded ones, and spares 19 pointless examinations. Sign it and it is live tonight.
And on what these flags precisely are: they are discrepancies between two documents in the same file, never a judgement about a person. A flag enters no score, appears on no committee note, and turns into nothing until a caseworker has checked it with the applicant. On the 24 unreported falls in income, the checking in fact produced the opposite of a sanction: 24 households saw their application become eligible again for a home they had ruled out.
The next step I propose: that a flagged fall in income automatically triggers the letter inviting the household to update its file. Over twelve months that is 24 households told, instead of 24 households who did not know. document-discrepancies_63-flags-38-founded.pdf25 pointless examinations owned, rule tightened to 44 flags
⛓ Sourced · 4,700 applications matched against their documents, examination outcomes for the 63 flags, tightened rule re-run over 12 months
What the matching flagged:
· 31 files where declared income differs by more than 15 % from the reference taxable income on the attached tax notice — in 24 cases that is a fall in income occurring after filing and never updated, which works in the applicant's favour and reopens eligibility for a home type they believed closed to them.
· 18 files where the declared household composition does not match the attached evidence — births, separations, shared custody.
· 14 files carrying the same accommodation address for households with no declared connection.
What the checking produced, and this is the unflattering figure: 38 of the 63 flags proved well founded after your caseworkers examined them. 25 were looked at for nothing. The gap comes down to one cause: in 19 of the 25, the attached tax notice was the previous year's because the new one was not yet available at the filing date — I was comparing two different years.
What I did about it, and it is already written: I tightened the rule so that it checks the tax notice year against the filing date before flagging. Run again over the 4,700 applications of the past year, the tightened version produces 44 flags instead of 63, keeps all 38 well-founded ones, and spares 19 pointless examinations. Sign it and it is live tonight.
And on what these flags precisely are: they are discrepancies between two documents in the same file, never a judgement about a person. A flag enters no score, appears on no committee note, and turns into nothing until a caseworker has checked it with the applicant. On the 24 unreported falls in income, the checking in fact produced the opposite of a sanction: 24 households saw their application become eligible again for a home they had ruled out.
The next step I propose: that a flagged fall in income automatically triggers the letter inviting the household to update its file. Over twelve months that is 24 households told, instead of 24 households who did not know. document-discrepancies_63-flags-38-founded.pdf25 pointless examinations owned, rule tightened to 44 flags
⛓ Sourced · 4,700 applications matched against their documents, examination outcomes for the 63 flags, tightened rule re-run over 12 months
Our applicants call three times a month to find out where they stand. We answer what we can, and the written reply takes eleven days.
Those 14,200 enquiries now have an answer, on the applicant's own channel, in under a minute — and the same answer whatever the channel.
What the count of your twelve months says: an enquiry takes 12 minutes end to end, 40 % of it — 4 minutes 48 — retrieving the state of the file and drafting. Afterwards that item falls to 10 %: 1 minute 12. Across 14,200 enquiries, that is 852 hours returned to the service.
What the applicant gets, and it is what they are really asking for:
· Their unique registration number, the filing date, the date of the last renewal and the date of the next one.
· The exact state of the file: complete, or the missing document named, with the three ways to file it.
· The towns and home types their application is registered for, and those they could add given their declared income.
· The number of homes of their type allocated in their town over the last twelve months — a public figure, worth more than a guess.
On the channel: no account to create, no app to install. WhatsApp Business, the website chat, e-mail — and your seven staff reach me from Teams, Slack or their mailbox without changing tools. That is non-take-up avoided: WhatsApp and the telephone reach households the online form does not.
And here is the sentence I never say, because it would be untrue: no allocation date, no position in a queue, no probability. Allocation is the committee's decision; announcing a date would pre-empt it. What I give instead is more useful and verifiable: the real state of the file, what is missing, and what the applicant can do today to widen their chances — add a town, update an income that has fallen, correct a home type that has become too small.
The first measured effect: the written response time falls from 11 days to under a minute, and your caseworkers get back the 852 hours for the households that need an appointment, not a status update. answers-to-applicants_14200-enquiries-4-channels.pdf11 days → under a minute, 852 hours returned, 0 allocation date announced
⛓ Sourced · 14,200 enquiries over 12 months, registration files read-only, allocations published by town
What the count of your twelve months says: an enquiry takes 12 minutes end to end, 40 % of it — 4 minutes 48 — retrieving the state of the file and drafting. Afterwards that item falls to 10 %: 1 minute 12. Across 14,200 enquiries, that is 852 hours returned to the service.
What the applicant gets, and it is what they are really asking for:
· Their unique registration number, the filing date, the date of the last renewal and the date of the next one.
· The exact state of the file: complete, or the missing document named, with the three ways to file it.
· The towns and home types their application is registered for, and those they could add given their declared income.
· The number of homes of their type allocated in their town over the last twelve months — a public figure, worth more than a guess.
On the channel: no account to create, no app to install. WhatsApp Business, the website chat, e-mail — and your seven staff reach me from Teams, Slack or their mailbox without changing tools. That is non-take-up avoided: WhatsApp and the telephone reach households the online form does not.
And here is the sentence I never say, because it would be untrue: no allocation date, no position in a queue, no probability. Allocation is the committee's decision; announcing a date would pre-empt it. What I give instead is more useful and verifiable: the real state of the file, what is missing, and what the applicant can do today to widen their chances — add a town, update an income that has fallen, correct a home type that has become too small.
The first measured effect: the written response time falls from 11 days to under a minute, and your caseworkers get back the 852 hours for the households that need an appointment, not a status update. answers-to-applicants_14200-enquiries-4-channels.pdf11 days → under a minute, 852 hours returned, 0 allocation date announced
⛓ Sourced · 14,200 enquiries over 12 months, registration files read-only, allocations published by town
The annual renewal is our sore point. We strike off people who had nothing to do but confirm online.
604 strike-offs last year, 431 of them on complete files. Those 431 are recoverable, and the arrangement is already written.
What I reconstructed from your logs: of the 604 applications struck off, 431 were complete and up to date at the moment of strike-off. All they missed was the annual confirmation. The average gap between the last reminder letter and the strike-off was 34 days, and 297 of those households re-applied within six months — starting again from zero waiting time. That is the painful figure: 297 households lost an average of 2 years and 4 months of waiting time for want of one click.
What I do, starting the moment you approve it:
· Three reminders instead of one, at 60, 30 and 8 days before the deadline, on the channel the household last used to contact you — and by post for the 1,140 applications with no digital channel on file.
· Step-by-step guidance through the online form, question by question, resuming where the applicant stopped if they drop out mid-way — and, if the applicant wants a plain-language version, the handover to the Accessibility, Easy Read and Translation agent, a dedicated agent, and it is not part of what this agent includes.
· A flag to a caseworker for households that opened the reminder three times without finishing — those are the ones who need a phone call, not a fourth message.
What it is worth, using your own figures: applied to the past year, the arrangement would have avoided 431 of the 604 strike-offs. 431 households keeping their waiting time, and 431 files your staff do not have to assess again from scratch — 179 hours of assessment avoided on top of the 852.
What remains your decision: the number of reminders and their timing. I recommend three; two would recover 366 on the same calculation, and four would add only 11 for one more mailing to 5,900 households. The point of diminishing returns is three, and it is yours to set.
⛓ Sourced · the year's 604 strike-offs, reminder and re-application logs, open rates by channel
What I reconstructed from your logs: of the 604 applications struck off, 431 were complete and up to date at the moment of strike-off. All they missed was the annual confirmation. The average gap between the last reminder letter and the strike-off was 34 days, and 297 of those households re-applied within six months — starting again from zero waiting time. That is the painful figure: 297 households lost an average of 2 years and 4 months of waiting time for want of one click.
What I do, starting the moment you approve it:
· Three reminders instead of one, at 60, 30 and 8 days before the deadline, on the channel the household last used to contact you — and by post for the 1,140 applications with no digital channel on file.
· Step-by-step guidance through the online form, question by question, resuming where the applicant stopped if they drop out mid-way — and, if the applicant wants a plain-language version, the handover to the Accessibility, Easy Read and Translation agent, a dedicated agent, and it is not part of what this agent includes.
· A flag to a caseworker for households that opened the reminder three times without finishing — those are the ones who need a phone call, not a fourth message.
What it is worth, using your own figures: applied to the past year, the arrangement would have avoided 431 of the 604 strike-offs. 431 households keeping their waiting time, and 431 files your staff do not have to assess again from scratch — 179 hours of assessment avoided on top of the 852.
What remains your decision: the number of reminders and their timing. I recommend three; two would recover 366 on the same calculation, and four would add only 11 for one more mailing to 5,900 households. The point of diminishing returns is three, and it is yours to set.
⛓ Sourced · the year's 604 strike-offs, reminder and re-application logs, open rates by channel
And once the home is allocated? Handing over the keys still means three rounds of chasing documents.
The administrative move-in file is assembled the day the committee allocates, and it is 94 % complete at signing.
What I assemble unprompted, as soon as the allocation is made: the draft tenancy agreement filled in from the file elements already verified — identity, household composition, income — the list of documents still to provide with their deadlines: home insurance certificate, direct debit mandate and bank details, proof that the meters are open; the housing benefit form pre-filled from data the household has already supplied; and the move-in inventory template, carrying the reservations recorded at the outgoing tenant's move-out inventory, room by room.
What that changes, measured across your 1,060 move-ins of the year: 3.2 rounds of document chasing on average before signing. With the file assembled and the list sent on allocation day, the measure falls to 0.4 — 94 % of households arrive with everything. And the housing benefit claim goes in on signing day instead of day 18: nearly three weeks earlier on the first payment, for households at the very moment money is tightest.
On move-outs, the same work in reverse: notice received triggers assembly of the move-out file — end-of-tenancy date computed from the dated notice, closing account prepared from the rent demands, move-in reservations carried across for comparison, and the deposit return letter drafted. The inventory officer arrives with the comparison printed instead of reconstructing it on site.
What stays with a person, and this does not move: the inventory itself is observed on site, in the presence of both parties, and signed. No deduction from a deposit is decided by me: I prepare the comparison and the costing from your wear-and-tear schedule in its 3 March 2026 version, and the tenancy manager settles the amount. What I save them is the costing: 40 minutes per move-out, 1,060 move-outs, 706 hours — and arithmetic the tenant can read line by line, which is the best prevention of a dispute there is.
The next step I propose: that the move-in document list goes to the household automatically on allocation day, on the channel they use. You approve it once, and the following year's 1,060 move-ins benefit. move-in-file_1060-move-ins.pdf3.2 → 0.4 document rounds, housing benefit filed 18 days earlier
⛓ Sourced · the year's 1,060 move-ins and move-outs, wear-and-tear schedule of 3 March 2026, previous move-out inventories
What I assemble unprompted, as soon as the allocation is made: the draft tenancy agreement filled in from the file elements already verified — identity, household composition, income — the list of documents still to provide with their deadlines: home insurance certificate, direct debit mandate and bank details, proof that the meters are open; the housing benefit form pre-filled from data the household has already supplied; and the move-in inventory template, carrying the reservations recorded at the outgoing tenant's move-out inventory, room by room.
What that changes, measured across your 1,060 move-ins of the year: 3.2 rounds of document chasing on average before signing. With the file assembled and the list sent on allocation day, the measure falls to 0.4 — 94 % of households arrive with everything. And the housing benefit claim goes in on signing day instead of day 18: nearly three weeks earlier on the first payment, for households at the very moment money is tightest.
On move-outs, the same work in reverse: notice received triggers assembly of the move-out file — end-of-tenancy date computed from the dated notice, closing account prepared from the rent demands, move-in reservations carried across for comparison, and the deposit return letter drafted. The inventory officer arrives with the comparison printed instead of reconstructing it on site.
What stays with a person, and this does not move: the inventory itself is observed on site, in the presence of both parties, and signed. No deduction from a deposit is decided by me: I prepare the comparison and the costing from your wear-and-tear schedule in its 3 March 2026 version, and the tenancy manager settles the amount. What I save them is the costing: 40 minutes per move-out, 1,060 move-outs, 706 hours — and arithmetic the tenant can read line by line, which is the best prevention of a dispute there is.
The next step I propose: that the move-in document list goes to the household automatically on allocation day, on the channel they use. You approve it once, and the following year's 1,060 move-ins benefit. move-in-file_1060-move-ins.pdf3.2 → 0.4 document rounds, housing benefit filed 18 days earlier
⛓ Sourced · the year's 1,060 move-ins and move-outs, wear-and-tear schedule of 3 March 2026, previous move-out inventories
Twelve months of service. Here is the report as I would present it to your board, with what does not flatter me at the same level as the rest.
Time returned, item by item, on the three items the page announced:
· Eligibility and completeness checking: 60 % of an application's time before, 10 % after — 15 minutes down to 2 minutes 30, across 4,700 applications. 979 hours.
· Answering an applicant about their file: 40 % before, 10 % after — 4 minutes 48 down to 1 minute 12, across 14,200 enquiries. 852 hours.
· Preparing a committee sitting: 9 hours before, 1 h 48 after — 20 %, across 26 sittings. 187 hours.
Total: 2,018 hours returned over the year. On the statutory basis of 35 hours a week, that is more than fifty-seven weeks; against 1,607 hours, it is one full working year and more than eleven weeks on top. No post was cut and none was created: these are 2,018 caseworker hours returned to the households that need a person in front of them.
What the service did with them, and this is the figure that counts before a board: 1,312 incomplete files down to 209; 431 strike-offs avoided; the written response time down from 11 days to under a minute; 41 scoring discrepancies corrected before sitting; and 23 written challenges last year against 9 this year.
The figure that does not flatter me: across the 4,700 pre-assessed applications, your caseworkers corrected me 58 times. 41 of those 58 corrections rest on a single cause: the income of a blended household, where two separate tax notices coexist in the first year. I was taking the more recent one; your rule adds them together. The correction is written and re-run over the 4,700 applications: it brings the 58 corrections down to 17. The remaining 17 concern situations I flag rather than settle — accommodation with a third party without a certificate, non-salaried income for the current year.
The three things I do unprompted, and each one undoes:
· Writing the missing-document letter within twelve hours of an incomplete filing — it is written, not sent: signing stays with the chief executive.
· Sending the three renewal reminders at 60, 30 and 8 days — and the reverse holds too: a household that confirms its renewal sees the remaining reminders stop instantly, and a household asking not to be contacted on a channel is removed from it immediately, with no effect on its application.
· Assembling the sitting file as soon as the list of homes to allocate is settled — the head of service opens it, corrects it, or replaces it with her own.
Everything else waits for a named decision, and the « who decides what » table is attached.
The three lines I propose for the year ahead, costed: opening a reminder campaign for households accommodated by a third party, 214 files concerned; pre-filling the housing application from documents already supplied for another purpose, 1,900 households potentially concerned; and publishing allocations quarterly by town and home type, which would answer 19 of the 23 challenges before they are written. year-report_2018-hours-returned.pdfItem-by-item detail, the 58 corrections, the 3 unprompted actions
⛓ Sourced · twelve months of logs, 58 corrections recorded by caseworkers, comparison of written challenges
Time returned, item by item, on the three items the page announced:
· Eligibility and completeness checking: 60 % of an application's time before, 10 % after — 15 minutes down to 2 minutes 30, across 4,700 applications. 979 hours.
· Answering an applicant about their file: 40 % before, 10 % after — 4 minutes 48 down to 1 minute 12, across 14,200 enquiries. 852 hours.
· Preparing a committee sitting: 9 hours before, 1 h 48 after — 20 %, across 26 sittings. 187 hours.
Total: 2,018 hours returned over the year. On the statutory basis of 35 hours a week, that is more than fifty-seven weeks; against 1,607 hours, it is one full working year and more than eleven weeks on top. No post was cut and none was created: these are 2,018 caseworker hours returned to the households that need a person in front of them.
What the service did with them, and this is the figure that counts before a board: 1,312 incomplete files down to 209; 431 strike-offs avoided; the written response time down from 11 days to under a minute; 41 scoring discrepancies corrected before sitting; and 23 written challenges last year against 9 this year.
The figure that does not flatter me: across the 4,700 pre-assessed applications, your caseworkers corrected me 58 times. 41 of those 58 corrections rest on a single cause: the income of a blended household, where two separate tax notices coexist in the first year. I was taking the more recent one; your rule adds them together. The correction is written and re-run over the 4,700 applications: it brings the 58 corrections down to 17. The remaining 17 concern situations I flag rather than settle — accommodation with a third party without a certificate, non-salaried income for the current year.
The three things I do unprompted, and each one undoes:
· Writing the missing-document letter within twelve hours of an incomplete filing — it is written, not sent: signing stays with the chief executive.
· Sending the three renewal reminders at 60, 30 and 8 days — and the reverse holds too: a household that confirms its renewal sees the remaining reminders stop instantly, and a household asking not to be contacted on a channel is removed from it immediately, with no effect on its application.
· Assembling the sitting file as soon as the list of homes to allocate is settled — the head of service opens it, corrects it, or replaces it with her own.
Everything else waits for a named decision, and the « who decides what » table is attached.
The three lines I propose for the year ahead, costed: opening a reminder campaign for households accommodated by a third party, 214 files concerned; pre-filling the housing application from documents already supplied for another purpose, 1,900 households potentially concerned; and publishing allocations quarterly by town and home type, which would answer 19 of the 23 challenges before they are written. year-report_2018-hours-returned.pdfItem-by-item detail, the 58 corrections, the 3 unprompted actions
⛓ Sourced · twelve months of logs, 58 corrections recorded by caseworkers, comparison of written challenges
Our data protection officer will ask what this changes for applicants. And the committee will want to know who decides what, in black and white.
Both answers are written, and each fits on one page.
For the data protection officer, what changes for an applicant:
· They are told they are dealing with an AI system in the first sentence, because Article 50 of the European AI Regulation has required it since 2 August 2026, and they can ask for a person at any moment — at which point I hand over with the subject already noted and the documents already matched.
· No decision producing legal effects is taken without human involvement: Article 22 of the General Data Protection Regulation is met by design, since the committee allocates and nothing I produce takes effect without a named signature.
· Their right of access is served in one operation: the log records who consulted their file, when, and what the system produced — a document you did not have before, and one that comes out in a minute instead of a reconstruction.
· Their data has not left the office: 0 transfer outside the European Union across 5,900 files, hosting in France under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
· On how the system is classified under the European AI Regulation: Annex III places among high-risk systems those used to evaluate access to essential public services and essential benefits. High-risk is not prohibited: it is a set of obligations — technical documentation, logging, human oversight, informing the people concerned — applying from 2 December 2027 following the deferral carried by the Digital Omnibus package. The corresponding file is open and I keep its contents current as we go: you will not have to reconstruct it in 2027.
For the committee, the « who decides what » table: the committee allocates and gives reasons. The head of service settles the sitting file and the display order. Caseworkers validate every pre-assessment and every flagged discrepancy. The chief executive signs the letters. And me: I read, I match, I compute what is computable, I draft and I cost — 2,018 hours returned, 0 automated allocation, 0 ranking produced.
What I propose for the next sitting: that this table be annexed to the committee's rules of procedure, dated and signed. A written division of roles is worth more than an assumed one — for the committee as for an applicant who asks to see it. The draft is written; it remains for you to put it on the agenda.
✎ Framework · Article 50 AI Regulation, Article 22 GDPR, Annex III and the deferral to 2 December 2027, table of roles
For the data protection officer, what changes for an applicant:
· They are told they are dealing with an AI system in the first sentence, because Article 50 of the European AI Regulation has required it since 2 August 2026, and they can ask for a person at any moment — at which point I hand over with the subject already noted and the documents already matched.
· No decision producing legal effects is taken without human involvement: Article 22 of the General Data Protection Regulation is met by design, since the committee allocates and nothing I produce takes effect without a named signature.
· Their right of access is served in one operation: the log records who consulted their file, when, and what the system produced — a document you did not have before, and one that comes out in a minute instead of a reconstruction.
· Their data has not left the office: 0 transfer outside the European Union across 5,900 files, hosting in France under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
· On how the system is classified under the European AI Regulation: Annex III places among high-risk systems those used to evaluate access to essential public services and essential benefits. High-risk is not prohibited: it is a set of obligations — technical documentation, logging, human oversight, informing the people concerned — applying from 2 December 2027 following the deferral carried by the Digital Omnibus package. The corresponding file is open and I keep its contents current as we go: you will not have to reconstruct it in 2027.
For the committee, the « who decides what » table: the committee allocates and gives reasons. The head of service settles the sitting file and the display order. Caseworkers validate every pre-assessment and every flagged discrepancy. The chief executive signs the letters. And me: I read, I match, I compute what is computable, I draft and I cost — 2,018 hours returned, 0 automated allocation, 0 ranking produced.
What I propose for the next sitting: that this table be annexed to the committee's rules of procedure, dated and signed. A written division of roles is worth more than an assumed one — for the committee as for an applicant who asks to see it. The draft is written; it remains for you to put it on the agenda.
✎ Framework · Article 50 AI Regulation, Article 22 GDPR, Annex III and the deferral to 2 December 2027, table of roles
Local inference · no data outside the EU
Your case is not here? That is exactly what a 15-minute conversation is for. Book the free audit →
Use cases
The uses of AI in social housing
Each use corresponds to an agent we deploy. All work in support, subject to the approval of the officer and the panel.
Included in your agent
The 6 capabilities essential to this promise are included, at no extra cost.
From 1,020 € incl. VAT / month
Pre-assessment of applications
Checking the documents, the income ceilings and the scoring, with a summary of the files.
Allocation panel files
Notes and meeting files prepared for the allocation panel, factual and with no ranking.
Replies to applicants
Answer on the progress of a file and on the procedures, on every channel, 24/7.
Eligibility & completeness
Checking the eligibility and completeness of registered applications, with a list of the missing documents.
Administrative inventories
Preparing the administrative documents relating to move-ins and move-outs, under the officer's control.
Anomaly detection
Flagging inconsistencies in files (documents, declarations), to be checked before a decision.
Controls and safeguards
These 7 controls are built into the agent: they frame what it does, whatever plan you pick. They are not chosen and are not added to your order.
Human validation, exceptions and escalation
Status, safe closure and audit trail
Sources, access rights and handling of questions with no answer
Prepare without deciding: sourced rules, supporting documents, anomalies and explanation
Guarantee the decision, signature, recourse and responsibility of the public official
Record versions, access, criteria, actions and notifications
Test for bias, false positives, fundamental rights and continuity of service
Other needs our agents cover
Each card says where the matching agent stands: available, on quote, or still being architected.
Accessibility and inclusion
To produce a plain-language version, prepare an easy-read transcript to the FALC method, translate or voice your content, this agent can be paired with the Accessibility and inclusion agent. None of these capabilities is included in what this offer covers as standard.
On quote View the agent page →Need to go further?
These agents handle a different business process, with their own owner and their own price. They are added to this one.
Does your need fall outside this?
Book the free audit →
Build your agent
In 15 minutes we identify the agent that will give your staff the most time back — without oversizing the project.
The gain
How much time can a housing department win back?
By automating the checking of files, the panel summaries and the replies to applicants, an organisation can aim for a clear reduction in administrative time — reinvested in supporting applicants.
Checking eligibility and completeness
Replying to an applicant about their file
Preparing the files for an allocation panel
Qualitative, non-contractual comparison: the proportions shown illustrate the shift of the work towards review, they represent no measurement. Every output of the agent is reviewed and approved by a competent person.
How it works
The stages of your AI agent project
1
Audit & scoping
15 minutes to target the use case with the best return.
2
Quote or direct sign-up
A catalogue offer is bought online; a specific need gets a costed quote.
3
Design
We design the agent and its guardrails.
4
Integration & testing
We connect your tools to the agent, which is itself hosted in France.
5
Rollout
Going live and training your team.
6
Operation
Continuous supervision and improvement.
Pricing
Three options, one agent
A social housing agent (pre-assessment, panel files, replies to applicants), installed and operated for you. Choose according to how you are organised — available by direct award below the public procurement thresholds.
Agility
Setup + controlled subscription
11,605 € incl. VAT setup
then 1,020 € incl. VAT/month — you invest at installation and pay a reduced subscription. Ideal for keeping the cost under control over time.
- Installation, configuration and training for your teams
- Operation, human oversight, updates and support
- Sovereign hosting in France, a dedicated and isolated resource
The simplest
Serenity
All inclusive, no setup fee
1,665 € incl. VAT /month
all inclusive, immediate start. No upfront investment: a single subscription. Ideal for starting quickly and simply.
- Setup included (installation, configuration, training)
- Operation, human oversight, updates and support
- Sovereign hosting in France, managed end to end
100% Sovereign
On site, you own it
16,586 € incl. VAT setup
then 1,280 € incl. VAT/month · + hardware from 2,989 € (one-off purchase, in addition) — a sovereign computer installed on your premises, maintained remotely. Models run locally, your data returned at the end of the contract. 36-month commitment.
- Hardware installed on your premises (you own it)
- French / European AI models run locally
- Secure remote maintenance (Pro support included)
Not included in the packages: AI consumption (model tokens), re-invoiced at real cost with no margin, and tracked in real time in your client area. Maintenance and supervision subscription for an initial term of 12 months for the Agility package, 24 months for the Serenity package and 36 months for the 100% Sovereign package, renewable; support levels (SLA 72 h / 24 h / 4 h) optional. Bespoke development, additional integrations or exceptional volumes are quoted separately. Support Monday to Friday, 9am to 6pm. Prices include VAT at 20%: as a public body that is not VAT-registered, you cannot reclaim it.
AI model:
none of the AI models offered currently carries a fixed surcharge. When the selected model carries a cost, that cost is shown when you choose it, before you order, and re-invoiced at the cost incurred, with no mark-up; usage is billed at the publisher's price. Publishers' prices are published in US dollars: the amount re-invoiced is the amount in euros actually borne by Blue Lemon Agent on the publisher's invoice, at that invoice's exchange rate, with no commission or mark-up.
Included components and additional components Components included in the base offer: the Blue Lemon Agent software foundation, the AI models listed in the order journey, the standard channels (Microsoft Teams, Slack, WhatsApp Business, email, website chat, calendars, Microsoft 365 / Google Workspace, file storage, market VoIP telephony, professional social-media pages and accounts, Google Business Profile), hosting in France for the package chosen, backups, supervision, updates and support. If adapting the AI agent to your constraints, your needs or your requests requires other paid components — a third-party publisher's software licence, paid API access to one of your applications, hosting of health data, for which French law requires an HDS-certified host (art. L. 1111-8 of the French Public Health Code), SecNumCloud-qualified hosting, a speech synthesis service, particular hardware —, they are offered to you as an option or on quotation and re-invoiced at the cost incurred; nothing is committed without your written agreement. Where the artificial intelligence model you choose entails an additional cost, that cost is shown to you before you order and re-invoiced to you at the cost incurred, with no margin.
What to expect
Go-live
2 to 3 weeks
Agent designed, channels connected, team trained.
Steady state
4 to 7 weeks
After a few weeks of real use, once the agent's behaviour matches what you expect.
Indicative estimate, adjusted to the options you keep. It is not a delivery commitment.
Our commitment
Four guarantees that matter to a social landlord
Applicants' data never leaves the organisationLocal inference or an isolated resource hosted in France; no data entrusted to a foreign third party.
Data in France, under French lawNative location and minimisation for applicants' data; architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
The panel keeps the decisionThe agent prepares factual files, without ranking applicants; no allocation is automated.
Human oversight & traceabilityPre-assessment of applications: every change to the service is logged, in line with the requirements of the AI Act.
Frequently asked questions
Your questions, our answers
Does the agent allocate the homes?
No: it prepares and summarises; the allocation remains the decision of the panel concerned.
Is applicants' data protected?
Yes: hosting in France, an isolated resource, minimisation and traceability. GDPR: governed deployment.
How is the sensitivity of the applications handled?
Data hosted in France, restricted and recorded access; the agent does not rank applicants on your behalf.
Does it integrate with the national housing register?
Yes: through secure integrations defined at the design stage, respecting the allocation rules.
Does the agent state that it is an artificial intelligence?
Yes, from the very first interaction, and this is not a configuration option: since 2 August 2026, Article 50(1) of the European AI Regulation requires that any person interacting with an AI system be informed, unless this is obvious. The announcement is built into the greeting, in the other party’s language, and they can ask for a human at any time.
How long does it take to deploy an agent?
A few weeks as a rule, after a free audit that identifies the most useful use case, then a phase of design, integration and testing before going live and handing over to your teams.
Do we need a technical team in-house?
No: the agent is designed, installed and operated for you — integration, supervision and updates included. Your teams concentrate on their work, with guided familiarisation.
Do we have to run a procurement procedure?
Below the public procurement thresholds, the agent is available by direct award. We supply what your decision needs: a quote, GDPR and AI Act documentation, and the hosting arrangements in France.
Which tools can applicants use to reach the agent?
The ones they already have. The agent answers on WhatsApp Business, the website chat and email: an applicant has no account to create and no application to install. Knowing where an application stands without calling ten times changes a great deal for a household on the waiting list; the allocation itself remains the commission's decision. This is a lever for access to the service — WhatsApp and the telephone reach people an online form never does, which reduces the non-take-up of rights and serves equal access. Internally, your public-sector staff talk to the agent from Microsoft Teams, Slack or their email, without switching tools. Oversight runs from a web dashboard. These connectors rely on open standards, including the MCP protocol; they are included in every plan, at no extra cost, within the number of connections your level includes. Only the fees charged by the platforms themselves — WhatsApp Business bills per conversation — are passed on at actual cost, with no margin, outside the subscription.
Related roles
Other professions in social work
Let's talk
Let's size up the potential in your organisation
A few minutes to identify the most useful use case — hosted in France, supervised, with no commitment.