An agentic system takes over a process. Autonomy ends where risk begins.
Agents that retrieve context from your documents, choose an action within explicit boundaries and leave a trail. If evidence is missing, they stop and hand off.
- sources
- stop
- yours
I don’t sell a chatbot. I define how a system behaves.
An agentic system is not a chat with more tools. It retrieves context from documents, decides the next step within explicit boundaries, performs the allowed actions and leaves a verifiable trail.
We start from one process and an exit criterion. First: sources, permissions, abstention, human oversight and fallback. Then: the model and the build.
The starting point is work full of documents, not the model of the moment.
Companies and service firms with document-heavy work
Case files, reports, contracts and email hold the context, but retrieving and checking it takes up expert time.
Processes that need context and a decision
You need to understand where the case stands, choose an allowed action and stop when the evidence is not enough.
Partners looking for agentic capability in co-delivery
Software companies and consultants who know the domain and are looking for the architecture, evaluation and handover of an agentic system.
Four steps. The first one also exists to tell you not to do it.
Audit, blueprint, build, optimization: we always start from the bottleneck that costs the most hours today.
Scoping call, 30 minutes
One process at the center: how it works today, where it breaks, how many hours it costs. If it isn’t worth automating, I tell you right there.
Agentic Process Review, 90 minutes
We map sources, decisions, allowed actions, abstention, human oversight and fallback. The output is the contract for the first system.
Building the system on one process
The agent works on the agreed sources, uses only authorized tools, logs decisions and actions and hands off beyond the defined threshold.
Handover and field check
Documentation, repository and access stay in your name. Then we watch how the system runs without me.
Need to pass cost, payback and maintenance on to whoever signs off? The sign-off brief (in Italian) fits on one page, with no form.
If the team doesn’t use the system, the system doesn’t pay off. That’s why there is also team training.
Three proofs with different roles: outcome, capability and reliability.
An agent retrieves the state of a project from real documents
Anonymous client, live session, exit criterion written in advance. It took 108 minutes instead of 90, and one of the two checks was carried over to the next block.
Orchestration, operational memory and handoffs between runtimes (in Italian)
A case on the internal system, to check architecture and operations. It is not a client result.
The gates that decide whether a system can continue (in Italian)
Claim checks, tracing and stop rules: what makes an agent verifiable.
No logos and no praise without a name behind them: without consent, the case stays anonymous. The main case, step by step and with its declared numbers, is in the case study. For a repetitive flow that doesn’t need an agent, the right page is business process automation.
What I don’t do, written up front.
- Generic agents without a process, an owner and an exit criterion.
- High-risk systems under Annex III of the AI Act (employment, credit, emergency triage) without a lawyer on the project.
- 24/7 ongoing support: I work alone.
- Hidden autonomy: sources, permissions, logs, fallback and human accountability are part of the delivery.
- License reselling: subscriptions stay in your name.
Eight questions before the quote, with the short answer.
How much does it cost to design an agentic system?
The scoping call costs €97. The Agentic Process Review lasts 90 minutes and costs €297. An agentic system on one process costs between €5,000 and €10,000; multi-process or multi-agent systems are estimated after the review.
What do I take away from the first call if I don’t go ahead?
The decision on the first process: an agentic system, deterministic automation or no intervention. If you need the technical map, the next step is the 90-minute review.
Do you work only with Claude, or with other tools too?
Not only with Claude: the choice depends on the process, the data and the constraints. The model is a replaceable component; the contract stays the same: sources, permissions, abstention, human oversight, logs and fallback.
Does my data need to be in order before we start?
No, but you need to know where things are. The first job is almost always making the state of the project retrievable, so that a newcomer, human or agent, finds the right file on their own.
How long before something actually runs?
The review takes 90 minutes. The build moves through verifiable milestones: first one process holds up on real and negative cases, then we decide whether to extend it. Timelines and dependencies go into the quote.
How do you handle privacy and AI Act compliance?
Sensitive data is processed locally when possible. As for the AI Act, Art. 4, applicable since February 2, 2025, requires taking AI literacy measures for the people who use AI: it is an obligation to take measures, not to achieve a result. For those who use AI, Art. 50 covers specific cases: deepfakes, text published to inform the public on matters of public interest, emotion recognition and biometric categorization. The EU database concerns high-risk systems. For literacy, the measure I offer is AI training for teams. This is not legal advice: validation stays with your lawyer.
You work alone. What happens to my system if we lose touch?
Everything lives in your files, in your repository, with plain-language instructions and services registered to your accounts. No proprietary platform. If there are integrations tied to a vendor, I tell you up front.
Where do you work? Do we need to meet in person?
I work in English or Italian, remotely or on site. Remotely, over video call with a shared screen: you see the terminal and the trail stays. On-site work is agreed case by case.
Where the models run, how I handle data and what the AI Act says about an internal process: Data, models and the AI Act (in Italian).
Bring one process. We leave the call with a decision, not a quote.
One process, 30 minutes. You come out knowing whether you need an agentic system, whether deterministic automation is enough or whether that process should be left alone. The call runs in English or Italian, remotely or on site.