How we build it
Agentic AI, where agents act inside your processes, has to meet the standards of any production system: identified, secured, measured and accountable. This page sets out how we design it, without the jargon and without the secrets.
Principles
Process first, model second
We start from a process and the decisions it carries, then choose where AI helps. The model is a component, never the strategy.
A person for every binding act
Paying, signing, sending, deleting: anything that commits you waits for a named person. Agents prepare, people decide.
Nothing we cannot measure
Quality, cost and time are measured from the first day, against a baseline taken before the pilot.
Your data, your perimeter
Data stays where your rules say it must, and the architecture lets you change model or provider without starting again.
A reference architecture
Every engagement adapts it, but the layers stay the same. Each one answers a question an auditor, a CISO or a CFO will ask.
- Your business systems
ERP, CRM, document stores, mailboxes, line-of-business applications. They stay where they are.
- One gateway, one protocol
Systems are exposed to agents through a single API gateway and the Model Context Protocol. A new agent reuses what exists instead of adding an integration.
- Agents with their own identity
Each agent has its own identity and only the roles it needs. Every call can be traced to one agent and revoked without touching the others.
- A model gateway
Calls go through a gateway that routes to several providers, or to open-weight models hosted in your environment, with caching and cost tracking on every call.
- Workflow orchestration
AI models read and propose; Microsoft Power Automate workflows, part of Microsoft Power Platform, carry the work from one system to the next, with retries, approvals and a full run history.
- Human approval
Binding actions stop at an approval step, with the context, the proposal and the reasons laid out for the person who decides.
- Observability across the stack
Logs, metrics, traces and alerts cover every layer, so you see what each agent did, when, at what cost and with what result.
Model strategy
The right model is the cheapest one that holds the quality on your real data.
Measured, not assumed
Candidate models are compared on real documents and real cases, and run twice on the same input: stability matters as much as accuracy.
Native document reading
Scanned documents go to models that read the PDF itself. A model that cannot see the page must return nothing, never a guess.
Prompt caching
Stable instructions and reference material are placed first and cached, so repeated calls cost a fraction of the first.
Plausibility before writing
Any extraction that writes into a system passes a plausibility check first. An answer that does not fit stops the flow.
Security by design
Agents get access to systems that matter. Security is built in from the first line, then proven.
Zero-trust APIs
Every boundary validates the token, the issuer, the audience, the expiry and the role. Knowing a key is never enough.
Least privilege
One identity per trust boundary, roles scoped to the task, access reviewed and revocable.
Secrets in a vault
Keys and credentials live in a vault with rotation, never in code, logs or telemetry.
Restricted networks
Services only accept traffic from the gateway, so no one can go around the controls.
Proof by the negative
We test that what must be refused is refused. An access that works proves nothing about the restriction.
Your data in Europe
Your data is hosted in European regions. When an AI model runs outside Europe, it runs under zero data retention: the provider stores nothing. When confidentiality requires it, we use models hosted in Europe or in your own environment.
European hosting by default
Applications, databases and documents are deployed in European cloud regions unless you decide otherwise.
Zero retention outside Europe
When a model runs outside Europe, the request is processed and forgotten: nothing is stored or used for training.
EU or in-house models when it matters
Confidential processes use models hosted in Europe or open models running in your own environment.
Governance and compliance
Every proposal traced
What the agent saw, what it proposed, who approved it and when: the record answers an auditor at any time.
EU AI Act upfront
Each use case is classified before work starts. High-risk uses are flagged with the safeguards they require, or left aside.
GDPR by default
Personal data is minimised, kept in its perimeter and never sent where it does not need to go.
Documentation that stays true
Architecture, data flows and decisions are documented when they change, so the next person can take over.
Running it
Logs that persist
Every service ships its logs and traces to a central workspace from day one.
Alerts on what matters
Error rates, latency, failed extractions and unusual cost trigger an alert before users notice.
Cost per process
AI spend is reported per process and per month, next to the time and errors it removes.
Kill switch
Every automation can be paused in one step, and resumes without losing work.
AI Act by design
The EU AI Act applies to the AI we build with you. We assess, prepare and document what it requires from the first workshop, without promising more than an assessment can prove.
Risk classified upfront
Every use case is assessed against the AI Act risk categories before work starts, and the result is documented.
No prohibited practices
We do not build systems that fall under the practices the AI Act prohibits.
High-risk uses handled with care
Uses such as employment decisions or creditworthiness are flagged and not deployed without the safeguards the regulation requires.
Transparency for people
When people interact with an AI system or receive AI-generated content, they are told so.
Human oversight and logs
A named person oversees each system, and its actions are logged so they can be explained afterwards.
AI literacy for your teams
The teams concerned are trained to use AI with judgement, an obligation that applies since February 2025.
From one process to production
- 01
Diagnostic
We map one process with the people who run it and agree the success criteria.
- 02
Pilot in real conditions
In five weeks the automation runs on your data, with your team, inside your environment.
- 03
Production
Security, access rights, monitoring, documentation and a named owner for every step.
- 04
Adoption
New ways of working co-designed with your teams: updated procedures, clear roles, training by role and internal champions.
- 05
Measurement
Results are compared with the baseline, then we choose the next process together.
People first
A tool that nobody adopts delivers nothing. Change management is built into the method, in step 4, and starts at the diagnostic.
Processes redesigned with the people who run them
New ways of working are co-designed with the teams on the ground, then written down as updated procedures.
Who validates what
Roles and responsibilities are set before go-live, including the person who approves every binding action.
Training by role
Each role learns what changes for it, with the AI literacy the AI Act expects from teams that use AI.
Champions in each team
Internal champions relay questions, collect feedback and help colleagues through the first weeks.
Social dialogue early
Where relevant, the works council and employee representatives are informed and consulted early, not at the end.
Adoption tracked after launch
Usage, feedback and results are followed after go-live, and the process is adjusted until it becomes the team's reflex.
What we do not do
No black boxes
You can see what every agent does and why. Nothing runs that you cannot inspect.
No tool sprawl
One gateway, one way to connect, one place to measure. Each new agent makes the platform stronger, not more fragile.
No binding act without a person
An agent proposes. A named person decides anything that commits you.
Let's look at your first process.
A working session to choose the process where Digital Intelligence pays off first, and how we would measure it.