Skip to main content
nexeai
Our specialty

Bespoke AI agent development

An AI agent is a system that chains steps, uses your tools and stops when a consequential action needs approval. We design these systems for your processes, with the permissions and controls you choose.

Talk about your project
architecture · inbound request agentExample

What triggers it

Message received

shared inbox

File dropped

watched folder

Every morning

07:00

AgentAnalyses the request and gathers what's needed

breaks it down · calls its tools · stops at committing actions

memory · the case stays open from one run to the next

What it consults, and on what basis

CRM

read

Documents

read

Calendar

read · write

HumainValidation before any committing action

sending · editing a file · financial commitment

An agent is a system that can break a request into steps, chain actions and use tools, such as your inbox, your CRM or your documents. It does not replace a team member: it takes on the repetitive steps, under the rules you set.

In practice, an agent can prepare the work, gather the information and propose an action. When the decision commits your organisation, approval stays human. That boundary, set during framing, is what makes the system reliable and accepted by teams.

We build these systems, including agent harnesses, end to end: framing the process, choosing a local, hybrid or cloud architecture, integrating with your tools and documenting the rules. All of it is built for French SMEs and independent professionals, with data confidentiality as the starting constraint. Tell us about your process in a first conversation.

What an agent can do

One task, followed from the message received to your approval

The mechanism does not change from one task to the next: something arrives, the agent gathers, it prepares, and it stops in front of you. Pick a task to see it run in full.

agent · inbound requestsExample
09:41 · subject: contract amendmentMessage received

“Hi, could you tell me where things stand on the amendment requested last month?”

AgentGathers before replying
  • client identified in the CRM
  • contract and amendment retrieved
  • status of the request checked
HumainReply prepared

draft · 1 attachment · awaiting your approval

The other tasks we keep seeing

Most projects start with one or two well-defined tasks.

Answering routine emails
The agent drafts replies to frequent requests based on your standard answers, and prepares the edge cases for your review before sending.
Preparing quotes
Starting from a brief or an exchange, the agent gathers the information, checks availability and produces a first structured quote.
Extracting data from documents
Invoices, contracts, supporting documents: the agent reads what comes in, extracts the useful data and reports it in your tools, with no manual entry.
Chasing late payments
The agent identifies overdue invoices, prepares a reminder that scales with how late the payment is, and follows the replies through to settlement.
Drafting meeting reports
After a meeting, the agent turns your notes into a structured report: decisions, actions, owners and deadlines.
Filing and archiving documents
The agent files incoming documents into the planned structure, applies the naming rules and flags what is missing.
Updating the CRM
The agent enriches contact records, logs recent exchanges and prepares the next sales actions.
Preparing meetings
The agent gathers the agenda, the related documents and a summary of open points, ready to use on the morning of the meeting.
Qualifying incoming enquiries
First-level reply to forms and messages: the agent checks the information, routes the enquiry to the right person and keeps the requester informed.
What we build

Capabilities designed for your processes

Each agent is configured for a precise scope: the tasks it handles, the tools it consults and the controls that bound its autonomy.

Autonomous agents
An agent breaks a request into steps, chains them and calls the tools it needs, within the permissions set for the project. It stops as soon as a consequential action needs a human's approval.
Multi-agent orchestration
When a task goes beyond a single scope, we split the steps across several specialised agents: one gathers, another analyses, a third drafts. A supervisor coordinates the whole and escalates the checkpoints.
Tool calls and integrations
The agent connects to your business tools: inbox, CRM, calendar, databases, internal APIs or spreadsheets. Each connection is configured with the minimal rights the task needs.
Persistent memory
The agent keeps the useful context from one session to the next: preferences, past decisions, the state of a file in progress. You decide what is kept, for how long and in which storage space.
Scheduled and triggered agents
An agent can run at a fixed time, every morning or every week, or be triggered by an event: an email received, a document uploaded, a row added to a spreadsheet.
Cloud or local deployment
We can propose a local, hybrid or cloud architecture, depending on the design chosen. We document the data flows, the services, the permissions, the approval rules and the controls planned.
What bounds the agent

An agent does not do as it pleases.

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

Before — what is allowed

Each tool gets the rights the task needs, and no more.

Example permissions granted to an agent, tool by tool.
CRMMessagingBank
allowed : read a recordallowed : readdenied : read
allowed : log an exchangeallowed : draft a replydenied : modify
denied : deletedenied : send unsuperviseddenied : pay

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

After — what it did

Every action can be logged, down to the second.

  1. 09:41:03message received
  2. 09:41:04CRM checked · customer record
  3. 09:41:05contract retrieved
  4. 09:41:07draft prepared
  5. 09:41:07approval requested
  6. awaiting human approval

Example. The agent stopped on its own: sending the message is a committing action, and it wasn't within its scope.

Confidentiality and architecture

Documented choices, not a black box

Confidentiality is not a sales pitch, it is an architecture decision. Every project comes with precise documentation.

Data flows
We map the data the agent reads, processes and produces, along with the services it passes through, so no transfer is implicit.
Permissions and access
Each connected tool gets minimal rights: the agent reads what it needs for the task, nothing more. Access can be revoked at any time.
Approval rules
We formalise the rules the agent applies: the scope of allowed actions, trigger thresholds, expected formats, cases to hand to a human.
Human approval
Any consequential action, such as sending a signed email or a payment, goes through human approval configured for your organisation.
Local, hybrid or cloud hosting
Depending on how sensitive the data is, the agent runs on your servers, on dedicated infrastructure or in the cloud. The choice is documented and reversible.
Logging
Every action the agent takes is traced: date, tool consulted, decision made, human approval where relevant. You can audit the system at any time.
Where the agent runs

Three architectures, and what each one costs

The question is not “is the cloud dangerous” but “what is allowed to leave it”. Here are the three possible answers, with their trade-off.

Local

Everything stays on your machines.

Your documents

your machines

The agent

your machines

The model

your machines

Requires hardware and someone to run it. Open models that run on a single machine are more modest than the largest hosted models.

Hybrid

Sensitive data stays, the rest goes out.

Your documents

your machines

The agent

your machines

The model

depends on the data

Every piece of data must be ruled on individually. This is the longest scoping of the three, and the one that needs the most precise documentation.

Cloud

Nothing to host on your side.

Your documents

your machines

The agent

hosted

The model

hosted

The fastest to set up, and the only one where your documents pass through a third party's infrastructure. Reserve it for what you'd accept sending there.

Who it is for

Agents designed for your trade

The cases below are illustrative: every project starts from your real process, your tools and your constraints.

Accounting firm

  1. Receipt received
  2. Items matched
  3. File prepared
  4. Approval

An accounting firm could hand an agent the pre-sorting of invoice receipts before human approval: matching items, spotting duplicates and missing amounts, preparing files for the accountant.

Law firm or solicitor

  1. Contract received
  2. Clauses flagged
  3. Summary note
  4. Review

A law firm could use an agent for the first read of incoming contracts: flagging sensitive clauses, summarising the negotiation points and preparing the summary note for the solicitor.

Estate agency

  1. Enquiry in
  2. Criteria checked
  3. Viewing proposed
  4. Follow-up

An estate agency could delegate the qualification of incoming enquiries to an agent: checking criteria, preparing viewings and automatically following up prospects who have not replied.

Tradesperson or small business

  1. Quote request
  2. Details completed
  3. Estimate
  4. Slot proposed

A tradesperson could hand an agent the reply to simple quote requests: gathering the missing details, a first estimate based on the business's rates and proposing an appointment slot.

Industrial SME

  1. Order placed
  2. Dates checked
  3. Delay flagged
  4. Weekly update

An industrial SME could use an agent to track supplier orders: checking announced delivery dates, flagging likely delays and preparing the weekly updates.

Association

  1. Payment received
  2. Receipt sent
  3. Dues reminder
  4. List updated

An association could delegate membership administration to an agent: confirming payments, sending receipts, chasing overdue dues and keeping the member list up to date.

Our method

From framing to pilot, in four steps

We move in stages, to measure the value before investing in a rollout.

  1. Step 1

    Framing

    • The process to handle, described step by step.
    • The tools involved and the sensitive data.
    • The approval rules, and written success criteria.

  2. Step 2

    Prototype

    • A first agent on the chosen case, on real data.
    • We test the results in a bounded pilot.
    • You then adjust its behaviour with us.

  3. Step 3

    Rollout

    • Integration with your systems and the controls put in place.
    • Supervision and documentation of the system.
    • Human approval, when the project provides for it.

  4. Step 4

    Follow-up and iteration

    • The pilot's results reviewed with your team.
    • The gap against the success criteria, noted.
    • The configuration adjusted on an ongoing basis.
Frequently asked questions

What we get asked before starting

Straight answers on scope, data and setting up a first agent.

Does an AI agent replace an employee?

No. An agent takes on repetitive, documented steps; it replaces neither judgement nor the client relationship. The goal is to free up time on low-value tasks and strengthen controls, not to cut headcount.

How long does a first agent take?

A first agent working on a bounded case takes a few weeks, framing included. The timeline mostly depends on the availability of your data and your tools, more than on technical complexity.

Do we need to know how to code?

No. Our approach is no-code: the agents are configured and set up by our teams, and your staff interact with them through simple interfaces, with no technical skill needed.

Where does our data go?

It depends on the architecture chosen: local on your servers, hybrid or cloud. The choice is made during framing, based on how sensitive the data is, and every flow is documented. No transfer is implicit.

How do we check the agent is not getting it wrong?

First through scope: the agent only takes the actions it is allowed to. Then through human approval on consequential actions, and finally through logging, which lets you audit every decision. The pilot exists precisely to check the results before widening the scope.

Can we start small?

Yes, and that is what we recommend. A pilot bounded to one or two tasks lets you check the real value before extending. The framing workshop is a good entry point for choosing your first use case.

Let's talk about the first task to hand to an agent

Tell us about the process that takes up your time. Together we identify the scope of a pilot, the access it needs and the approval rules.

Book a call

To go further

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

Workshop: frame and prototype an AI agent

Half a day in a group to choose your use cases, map your data and leave with a prototype, no technical background needed.

AI and automation audit

Two weeks to identify the tasks you can automate, whether they are feasible, and a 90-day roadmap with before-and-after measurement.

Process automation

Simple chains of actions to handle your repetitive tasks, alongside or ahead of AI agents.

AI agent training

Build your team's skills on AI agents and their business use, to frame your projects better and steer your teams.