AI for public agencies

A request comes in. The right department gets an already-prepared file.

Shared inboxes, resolutions, replies to residents: an assistant routes, gathers and prepares, your authorised staff approve and sign.

Describe a need

front desk · shared inboxExample

email received · 08:12Request

“Hello, I would like to know the status of my planning application filed in the spring.”

AgentRoutes and gathers

relevant department identified: planning
case found: PC-2026-184
related documents and decisions gathered

RésultatDraft reply prepared

written in plain language, from the case's documents

HumainApproval by the caseworker

nothing goes out without review

Public agencies handle a large volume of requests, letters and documents every day. Some of that work is repetitive and well documented: that is where AI finds its place — not to replace staff, but to free up their time from low-value tasks.

Our approach is no-code, built for French SMEs and public services, with confidentiality as the starting constraint. Every project requires choosing an architecture together with the agency. A local deployment and open models can be considered, depending on the data, the intended use and the constraints chosen.

We work with the staff involved, from defining the use case through to measuring the results. Describe a need from your agency in a first conversation.

Where AI steps in

Inside the path of a request, not beside it

An assistant does not replace a step: it prepares the next one and hands back. Here is where that pays off, and where it does not.

A resident's request

Shared inbox
emails, forms, letters
Agent
Reads the request, sorts it and routes it to the right department
request linked to an existing file
Humain
The front-desk officer checks the routing
Casework
documents, resolutions, procedures
Agent
Gathers the related documents and resolutions
file prepared, outstanding points listed
Humain
The caseworker completes and processes it
Reply
letter to the resident
Agent
Drafts a reply in plain language
a draft, never sent on its own
Humain
The authorised officer approves before sending
Decision
ruling, resolution, signature
No agent here, and that is not a gap: the decision and the signature belong to authorised staff and elected officials.

What it changes, role by role

Senior management and elected officials: Summaries and tracking tables to steer ongoing files, prepare decisions and get a full picture without digging through documents.
Department heads: Heads of registry, planning, schools or HR save time on repetitive tasks and keep control of approvals.
Field staff: An assistant beside every staff member, to find a procedure, prepare a letter or rewrite a text, with no prior technical skill.
Digital officers and IT management: Architecture, security and compliance are documented from the start: data flows, hosting, permissions and approval rules.

Three scenarios

Named the way you name them

Nobody goes looking for “an assistant for department heads”. They go looking for a way to clear a shared inbox without keeping a resident waiting.

Scenario 01

Handling a shared inbox

Requests arrive in an inbox that several staff members check
The agent reads each request and identifies the right departmentthe agent takes over
It links the request to an existing file, when there is one
It drafts a reply for documented cases
The department officer approves, corrects or takes overyou take back the reins

Today: One request in two ends up bounced to another department, and the resident waits in the meantime.
With an agent: A request routed as soon as it arrives, with the file already linked.

Scenario 02

Preparing a committee meeting

The meeting is approaching and the files are scattered
The agent gathers the past resolutions and related documentsthe agent takes over
It summarises the outstanding points in each file
It prepares the meeting file from your template
The department reviews it, completes it and passes it to the elected officialsyou take back the reins

Today: Preparation happens the week before, and the history is dug up by hand.
With an agent: A meeting file built from the sources, with what remains to be decided kept in view.

Scenario 03

Answering a recurring request

The same question comes up several times a month
The agent finds the applicable procedure and textthe agent takes over
It drafts a reply in plain language, with its reference
It flags the cases the procedure does not cover
The authorised officer approves before sendingyou take back the reins

Today: Every staff member rewrites the same reply, in their own words, and the quality varies.
With an agent: A consistent reply from one staff member to another, that cites the text it relies on.

The other tasks we keep seeing

Automations worth exploring with staff and residents.

Resident requests: Help handling and answering residents' letters and requests, with approval from your staff.
Document summarisation: Summarising resolutions, reports and large files, to find the useful information faster.
Administrative drafting: Generating first drafts of reports, notes and standardised documents.
Accessibility & clarity: Rewriting administrative content in plain language, for the agency to approve.
Internal assistants: Assistants connected to your procedures, to equip staff in their daily tasks.
Letters and shared inboxes: Automatic sorting and routing of incoming letters to the right department, with draft replies prepared.
Preparing committee meetings: Gathering past resolutions and related documents, summarising outstanding points and preparing meeting files.
Regulatory watch: Tracking the texts and circulars that apply to your remit, with a summary of their impact for your departments.
Architecture worth exploring: A local deployment and open models can be considered, depending on the data, the intended use and the constraints chosen.

What bounds the assistant

A public-agency assistant answers no one.

Two tables, and they read together: what the assistant is allowed to do before it starts, and what it did once it has finished.

Before — what is allowed

Each department tool gets the rights the task needs, and not one more.

Example permissions granted to an agent, tool by tool.
Shared inbox	Document base	Case management software
allowed : read	allowed : read resolutions and procedures	allowed : read a file's status
allowed : route to a department	allowed : summarise and cite	allowed : log an exchange
denied : reply on its own	denied : amend a document	denied : process, decide

Example. The actual split is decided during framing, tool by tool, and access can be revoked.

After — what it did

Every request handled leaves a traceable record.

08:12:03request received
08:12:05department identified · planning
08:12:09file PC-2026-184 found
08:12:14draft reply prepared
08:12:14sent for approval
awaiting the caseworker

An example. The assistant sent nothing: replying to a resident commits the agency, and the reply is waiting for an authorised member of staff to approve it.

Your first assistant

Where an agency would start

A first scope is chosen small: one inbox, its documented cases, a measurement after thirty days.

first project · shared inboxExample

Handling a single shared inbox

What the agent covers

vérifié : Reading and sorting the requests received
vérifié : Routing to the right department
vérifié : Linking to the existing file
vérifié : Draft reply for documented cases
exclu : Sending to the resident

What it needs from you

a month of real requests
your directory of departments
the procedures for recurring cases

What stays with you

approving every reply
handling the files
the decision and the signature

What we look at after thirty days: the share of requests routed correctly the first time, and the time before the first reply reaches the resident.

estimate · requests per monthYour figures

Requests received per month

requests / month

Average time to route and reply

minutes

Time consumed today

≈ 45 hper month

This total is yours: it only comes from the two figures alongside. Only part of this time can be automated — we look at the volume, how stable the process is, the exceptions and the data available before saying so.

Get this process reviewed

Our method

How a project runs

We move in stages, to check the value before widening the scope.

01
Frame
- The departments' needs, the data involved and the regulatory constraints. You leave with a clear scope and success criteria.
02
Test
- A first assistant on a real use case, tested with the staff involved on a bounded scope, before any wider rollout.
03
Deploy
- Integration into departments' practices, staff training, approval rules and the supervision planned.
04
Measure
- Usage and results tracked with your team, the gap against success criteria measured, the setup adjusted continuously.

Frequently asked questions

What we're asked before getting started

Straight answers on data, regulation and rollout.

Do residents' data stay protected?: Yes, and it is a systematic part of framing. We map the data processed, document every flow and plan for local architectures when sensitivity requires it. Nothing is transferred by default.
Is this use case really worth automating?: Not always, and we say so when it is not. We check the volume, the time it consumes, how stable the process is, how many exceptions it has and what data exists before proposing an agent. A rare task, or one full of special cases, costs more to automate than to do.
Can AI draft official documents?: AI prepares first drafts: letters, minutes, notes. Approval, signature and the decision stay in the hands of authorised staff. The system is set up to flag cases that need to go to a person.
Do we need a large technical setup?: No. Our approach is no-code: the assistants are configured by our team and used by staff through simple interfaces. The choice between local, hybrid or cloud hosting is made during framing, based on your constraints.
Will staff have to change their habits?: The assistants fit into existing tools and current practices. We train the teams and start with a reduced scope, so adoption happens gradually, department by department.
Can we start with a reduced scope?: Yes, and that's what we recommend. A first use case, such as handling a shared inbox or summarising resolutions, lets you check the real value before extending to other departments.
How do you avoid errors in replies to residents?: Through scope and approval: the assistant only answers documented cases, proposes a draft reply that staff approve before sending, and every action is logged so it can be audited.

Let's talk about your public-service project

Let's discuss your use challenges, your data constraints and the use cases worth considering in your agency.

Book a call

To go further

Other formats complement this offer, depending on where you're starting from.

A request comes in. The right department gets an already-prepared file.

Inside the path of a request, not beside it

What it changes, role by role

Named the way you name them

Handling a shared inbox

Preparing a committee meeting

Answering a recurring request

The other tasks we keep seeing

A public-agency assistant answers no one.

Before — what is allowed

After — what it did

Where an agency would start

How a project runs

Frame

Test

Deploy

Measure

What we're asked before getting started

Let's talk about your public-service project

To go further

Legal sector

Bespoke AI agent development

AI and automation audit

AI agent training