Implementation · Level 3 · Custom development

Custom-built, running in your own infrastructure.

An automation written for your process, paired with the right technical profile: AI engineer, developer. The business logic is modelled before it is coded. The code is produced with Claude Code, reviewed line by line, then delivered containerised into your infrastructure: repository, image and documentation are yours. Designed for organisations whose technical team can deploy and maintain an application.

Business interviews UML modelling Claude Code Fall-back Docker Deployed on your side
The prerequisite

Designed for organisations with a technical team.

A developed application lives well when someone on your side hosts it, monitors it and updates it. Three capabilities are enough; all are standard operations work, none requires AI expertise.

Capability 1

Deploy

Someone on your side can run a Docker container, on an internal server or in your cloud, and expose it to its users.

Capability 2

Maintain

Apply an update, read the logs, react to an alert, manage secrets and API keys. Standard operations work.

Capability 3

Evolve

Open the repository, find a rule in the model, adjust a prompt or switch models without rebuilding the application.

The formats

Four stages, committed one at a time.

Each stage is committed separately. The technical scoping stands on its own, and it sometimes points to a simpler level: that is a useful outcome.

Scoping · 3 h

Confirm the right level

Feasibility, target architecture, maintenance load, estimated running cost, and the question asked plainly: which level best answers the need?

The business owner
and your technical team
Modelling · 1 to 2 weeks

Business interviews and UML

The process as it is actually run, gathered from the people who run it, then formalised into diagrams that serve as the specification.

3 to 6 people
interviewed separately
DEVELOPMENT · 4 TO 6 WEEKS

Build and harden

Written with Claude Code under human review, a test set on your real cases, fall-back strategy, containerisation.

AI engineer or developer
weekly checkpoints
Handover · on delivery

Deploy on your side

Go-live in your environment with your technical team at the controls, operations documentation, then support on request.

Your technical team
and the business owner
How it runs

From the described process to the application running in your infrastructure.

Six stages, in this order. The first two make the rest work: an automation lasts when the process was understood before it was coded.

01

Check that custom development is the right answer

Custom development is the largest investment of the three levels. Chosen well, it is also the most transformative — and this stage is there to make sure before any development starts.

  • What calls for development: high volume, many business rules, systems to orchestrate, execution without human intervention, or a data sovereignty constraint
  • What belongs to another level: a one-off need or a process still in flux, where an assistant or a no-code agent answers better, and faster
  • Target architecture: what calls for a model and what stays deterministic code — the latter costs less and gives a constant result
  • Running cost: estimated per run and per month, before committing, not at the first invoice
02

Business interviews: the process as it really is

The real process lives in the heads of the people who run it: the exceptions, the implicit trade-offs, the tacit rules. The interviews capture them, and that is what makes an automation hold up in the real world.

  • Who we meet: the people who do the task, the person who checks it, the person who picks up the errors. Separately: the versions differ, and that is instructive
  • What we look for: the real step-by-step flow, the inputs and their quality, the decisions made and on what criteria, the edge cases that are owned
  • The tacit rules: what nobody wrote down because “everyone knows it”. That is where the value of the automation sits, and its difficulty
  • Volumes and time spent: the measure of the gain, and the baseline after go-live
  • What stays human: the decisions that are never delegated, settled at this stage
03

UML modelling: agreement before any code is written

The interviews become diagrams. The business validates what it recognises, the technical side validates what it can build. The model then plays three roles: scope contract, development specification, and documentation that outlives the team.

  • Activity diagrams: the process flow, its branches, its decision points and its retry loops
  • Use cases: who triggers what, with which rights, and what the system refuses to do
  • Sequence diagrams: the exchanges with your existing systems: ERP, CRM, document base, mail
  • Data model: the business objects handled, their states and the permitted transitions
  • Out of scope: written down as explicitly as the scope itself. That is what prevents drift mid-project
04

Development with Claude Code, under human review

Claude Code writes much of the code from the validated model. An engineer drives, reviews and arbitrates. The gain: speed on what is mechanical, and human time concentrated on what matters.

  • Readable, documented code: the constraint is that a developer on your team can pick it up without us
  • Prompts and business rules isolated: kept in configuration files, editable without touching the rest
  • Tests on your real cases: the files brought during the interviews, including the awkward ones
  • Model choice: the smallest model that holds the task, swappable for another provider without a rewrite
  • Versioned in your repository: Git, full history, from week one, not at delivery
05

Fall-back: what happens when it fails

A live application differs from a prototype in one respect: it knows what to do when a model answers off-target, when an API goes down, when a document arrives in an unexpected format. Here every case has a defined behaviour.

  • Output validation: every model response is checked against an expected schema; what fails never reaches production
  • Retries and failover: a new attempt, another model or another provider, queuing if the service is unavailable
  • Confidence threshold: below it, the case goes to human review rather than producing a doubtful decision
  • Graceful degradation: the manual process stays available at all times; the automation never becomes a single point of failure
  • Logging and alerts: every run traceable, every failure notified to a named owner. Never a silent error
  • Costs under watch: caps per run and per period, alert on overrun
06

Docker, deployment on your side, handover

The application is delivered containerised and deploys into your infrastructure, with your technical team at the controls. Your data stays within your perimeter: only the model calls leave, and you know which ones.

  • Docker image and compose file: a reproducible deployment, identical in test and production
  • Configuration and secrets: environment variables, API keys managed on your side, no credentials in the code
  • Hosting: internal server, your cloud, or a self-hosted model if sovereignty requires it
  • Operations: health endpoint, logs, usage and cost metrics, update and rollback procedure
  • Operations documentation: install, monitor, diagnose, redeploy. Written for your team
  • Handover session: at the end, your technical team redeploys, changes a rule and switches models with no outside help
Discuss this format: 30 min, free →
What you take away

Six deliverables, all in your hands.

No dependency on a subscription, a platform or a supplier. If you decide to carry on alone, everything is there.

01 · Code

The complete Git repository

History, prompts, configuration and test set. Hosted on your side.

02 · Runtime

The containerised image

Docker and compose file, redeployable identically with one command.

03 · Business

The UML model

The process formalised and validated: the reference for any future change.

04 · Operations

The operations guide

Install, monitor, diagnose, update, roll back.

05 · Robustness

The fall-back matrix

Every failure mode, its expected behaviour, who is alerted.

06 · Autonomy

The handover to your team

A transfer session, then support on request, at your pace.

The principle
Built for you, delivered to you: the tool is entirely yours, and your team knows how to keep it alive.
Use cases

The needs that call for development.

High volume, many rules, several systems to orchestrate, execution without intervention. Below those criteria, a no-code agent usually answers the need.

Manufacturing · Sales administration

Processing incoming orders

Hundreds of orders a week, arriving as PDFs, emails or spreadsheets: extraction, checks against the reference data, injection into the ERP, human hold on any discrepancy.

Outcome · Data entry disappears; the team handles only the exceptions
Real estate · Legal

Reviewing an entire contract portfolio

Several thousand leases reviewed against your analysis rules, with structured output per asset and a link back to the source clause.

Outcome · Months of work brought down to a few days of processing
Healthcare · Compliance

Producing regulatory dossiers

Assembling documents to a mandated format, completeness and consistency checks, traceability of every decision made along the chain.

Outcome · Compliant, defensible dossiers, produced without manual monitoring
Services · Customer care

Qualifying and routing requests

Every incoming request analysed, categorised, enriched with customer context from the CRM, then routed to the right team with a draft reply.

Outcome · First-response times cut, with no support reorganisation
Finance · Accounts payable

Classifying scanned documents

Invoices, delivery notes and scanned customer files: document type identification, field extraction, matching against the order, filing in the right folder.

Outcome · A paper flow sorted and matched with no manual entry
Leadership · Market intelligence

Agentic press review

A system that reads your sources every morning — press, sector publications, regulatory texts — cross-checks them against your watch topics and delivers a sourced summary to the right people.

Outcome · Thorough daily monitoring, for the cost of a ten-minute read
Discuss this format: 30 min, free →
Get in touch

Describe the process and we will tell you which level answers it best.

The direct route

Book 30 minutes

A scoping call: the target process, your systems, your technical team, the level that fits best.

Book 30 minFree · No commitment
Paired with an AI engineerDeployed in your infrastructureEnglish · French · Spanish
In writing

Describe your need

Three lines is enough: reply within 24 h.