Title: Michele Natali · Agentic workflows and confidential AI for SMEs · Language: en · Canonical: https://michelenatali.it/ Custom agentic workflows for SMEs # The process eating your hours can be taken off your hands. I analyse the process, build the workflow around it and set up what's missing. The agent prepares, the decision stays with you. [Let's talk](https://michelenatali.it/en/#contact) · [How it works](https://michelenatali.it/en/#method) I reply within one working day. If there's no suitable case, I'll tell you. fig. 00 · adoption gap, 2025: 8% SMEs, 71% large companies. - **8%** — of Italian SMEs have started an AI project. Among large companies it's 71%. - **58.6%** — Among companies that assessed AI but didn't adopt it, the first barrier is a lack of skills: 58.6%. High costs come in at 43.0%. - **77%** — of the Italian AI market is made up of custom projects. Not off-the-shelf software. Sources: Politecnico di Milano Artificial Intelligence Observatory, 2025 research (published 5 February 2026); ISTAT, Imprese e ICT, 2025 (15 December 2025). The problem ## Generic AI isn't enough for the work you do. Fine for a first draft. Not for a client file, and not for a process nobody has lined up yet. The first barrier reported by companies that assessed AI and didn't adopt it isn't cost: it's skills (58.6%). That's the problem I solve. Source: ISTAT, Imprese e ICT, 2025. - **Confidentiality** — In general-purpose AI services your data leaves your perimeter. For a firm with duties of confidentiality the question isn't how good the model is: it's where the file ends up. - **Time** — The hours go on things a machine can prepare: retyping, sorting, searching, checking again. - **Processes** — Off-the-shelf software solves somebody else's process. Yours stays manual, because nobody lines it up. What I do ## Agentic workflows, automation and custom software. One person: analysis, code, integration and ongoing support. The agent prepares, your team decides. I don't sell off-the-shelf software: the process is yours, so I build the workflow around it. Where no-code tools stop, I write the code. If you have no servers, no environment, no API access, logs or backups, I set them up. On your servers when data has to stay in-house, in the cloud when that's enough and costs less. - **Agentic workflows** — An agent reads the request, prepares the draft and leaves the decision to whoever signs. - **Automation** — Email, PEC, documents and the handoffs between your practice management system and CRM that stop being manual. - **Custom software** — When the tool you need doesn't exist, it gets built. What I don't do - I don't guarantee halved timelines. - I don't guarantee an ROI. - I don't start anything without a human sign-off. Confidential AI ## State-of-the-art models, without handing over your data. Encrypted inference: the infrastructure running the model can't read your data. No hardware, no GPUs to buy. For law firms, accountants and SMEs that can't send a client file to a general-purpose AI service: you need a strong model, but the data has to stay unreadable to whoever runs the computation. ### Is my data safe? What you send travels encrypted and is processed in an environment that can't read it. That environment is checked before anything is sent: if the check fails, nothing leaves. The outline stays visible: how many requests, when, and how big they are. ### Do I need to buy hardware? No. Nothing to install and no GPU to put in the server room: it's used as a service, pay as you go. Whoever provides the service brings the hardware. ### Is this serious, or an experiment? These are state-of-the-art models, run in an environment designed not to read your data. It isn't a test bed: it's what I use in the workflows I put into production. Where it falls short, I'll tell you. What I don't promise - I won't claim it's watertight: there are technical limits nobody can rule out. - I won't claim it makes you compliant: compliance depends on your obligations, not on the supplier. - Before touching client data we put in writing who does what. [Let's talk](https://michelenatali.it/en/#contact) If confidentiality is the issue, that's where we start. plate 01 · recorded, not readable Method ## Three phases, from diagnosis to ongoing support. It starts with the diagnosis. Project and ongoing support are quoted afterwards, once we know what's needed. ### Diagnosis 01 · 1–2 weeks I map the processes, we choose 1–3 cases where the effect is measurable, I check access, data and tools. At the end you have a plan with timelines, costs and a chosen pilot case. ### Custom project 02 · 4–8 weeks · Quoted after the diagnosis I build the first workflow in production and integrate it with your tools. We test it together on real cases before going live. ### Ongoing support 03 · monthly, minimum 3 months · Quoted after the diagnosis Monitoring, fixes, small improvements and support. The workflow stays alive after the project. The diagnosis takes one or two weeks. Project and ongoing support depend on the process and are set out in the proposal. Payment by milestone. [Let's talk](https://michelenatali.it/en/#contact) About ## One person, not an agency. Independent IT consultant. I write the code, set up the infrastructure and talk to your team: no salespeople in between. Before that I worked on production software, automation and integrations, in companies of every size. Remote, on site when needed · Italian and English · Everything documented [Read the full story](https://michelenatali.it/en/about/) Contact ## Tell me what weighs on you most. Two lines about the process that takes up your time. I reply within one working day and tell you whether it's worth talking about. [Send me an email](mailto:michele.na7ali@gmail.com) · michele.na7ali@gmail.com · [Book a call](https://calendly.com/michele-na7ali/30min) — Opens on calendly.com · 20 min No form: the email comes to me. If there's no case, I'll tell you. Michele Natali · VAT No. 04719270987 - [Email](mailto:michele.na7ali@gmail.com) - [LinkedIn](https://www.linkedin.com/in/michele-natali-b5079a151) - [GitHub](https://github.com/miche1e) - [Privacy](https://michelenatali.it/en/privacy/) - [Plain text for agents](https://michelenatali.it/llms.txt) No profiling cookies. No third-party tracking.