Recommended setup for a small studio

A proportionate, defensible AI setup for a small firm — which contract to choose, what it really costs, and what correctly stays on you.

The short version

If you want to use AI on real work — the kind with client names, filings and contracts in it — the question isn't "which model is smartest." It's having a contract that holds up under the GDPR. The concrete recommendation: use Claude Cowork (Anthropic's "agentic" desktop app, which carries out multi-step work on your files) on a Claude Team contract — not on a personal Pro or Max subscription.

Part of it comes down to one acronym: the DPA — the first requirement, necessary but not sufficient (on its own it doesn't close the transfer of data outside the EU, which stays on you, and is manageable). The rest of this page explains why Team, what it costs, and where to be careful.

What Cowork gives you (and why it fits a firm)

Cowork is Anthropic's product that lets the AI work like a collaborator operating on your documents: it takes a task, breaks it into steps, works on the files, and reports back. It's available on all paid Claude plans through the desktop app. For a small firm it's the most direct way to bring capable AI into your real workflow, with nothing technical to set up.

But which paid plan you choose changes everything on the privacy side. That's where the DPA comes in.

Why Team (and not Pro/Max): the DPA

When you handle client data, the GDPR requires you to have a data processing agreement with your vendor — the DPA, set out in Article 28 GDPR. It's the document that governs Anthropic as a processor acting on your behalf. Without it, you're sending personal data to a party with whom you don't have the agreement the law requires.

Here's the point that decides the setup:

• The consumer plans — Pro and Max — have no DPA. They're built for personal use.
• The Team plan is the cheapest one that includes the Article 28 DPA, auto-incorporated through Anthropic's Commercial Terms. The same holds, upward, for Enterprise and for API access. How to view and sign the DPA.

One detail that matters if you're solo: the DPA attaches to the organization, not to a headcount. You don't need five distinct humans for the DPA to apply — it applies to the Team account as such. (The "five seats" as a contractual minimum is a separate knot, covered below.)

On the security side, Anthropic states it holds recognized certifications (among them ISO 42001 and SOC 2 Type II): you'll find the current status in its privacy and security center. An important note: zero data retention and EU residency are guarantees typical of the Enterprise plan, and are not included in Team. Team gives you the DPA; it does not in itself give you those two extra guarantees.

Before you sign: contractual terms change. Confirm on Anthropic's privacy center that the DPA is included in the plan at the moment you subscribe, and read what it covers.

What it costs: the minimum

Team has a 5-seat minimum. You can't activate it for a single seat: the contract starts at five.

Indicative list pricing for the Standard plan is roughly $20–25 per seat per month — the lower figure with annual billing, the higher with monthly. The current list is at claude.com/pricing. Prices are in dollars (Anthropic bills in $) and change: treat these numbers as an order of magnitude, not a quote.

In plain terms: the spending floor to enter the "with-DPA" world is five seats, even if only a few people use them.

"One lawyer, 5 seats?"

A fair question for a solo practitioner: can I take the Team plan — five-seat minimum — and hold all of them myself? Let's answer honestly, because it's a grey area.

• Seats are tied to an email address. Each seat is an identity.
• What Anthropic explicitly forbids is sharing a single login across several people — not one person holding several seats.
• Holding all five seats yourself is not explicitly prohibited. But Anthropic describes Team as a plan "not intended for individual use." So: grey area. Before you build your firm's operations on top of it, it's worth asking Anthropic's sales team directly and getting confirmation that the use you have in mind is allowed.

On capacity, the real numbers: a Team seat is worth roughly the same as a Pro plan in terms of usage — the Standard tier offers about 1.25× a Pro, the Premium tier about 6.25× a Pro. So five Standard seats give you, added up, something on the order of six times a Pro — plus, above all, the DPA.

No-training: the default per plan, and what to check

The other question that matters for a firm: is my data used to train the models? It depends on the plan — and in two cases out of three the default already protects you.

• Pro (and consumer plans in general): here, whether your content is used to improve the models depends on a setting — "Help Improve Claude", under Settings → Privacy. Since October 2025 it's an explicit user choice; if you handle client data, make sure it's turned off. One exception always applies: conversations flagged for safety review can be used regardless.
• Team: your data falls under the Commercial Terms, and there Anthropic does not use your data to train its models — that's the default. The only way it would be used is if an admin voluntarily joins the Development Partner Program (opt-in, under Settings → Privacy controls): just make sure it isn't enabled. On Team, no-training is the starting condition, not something you toggle on or off by hand.
• Enterprise: your data is not used for training, and it's the strongest guarantee (configurable retention, down to a negotiable zero-retention).

In short: on Team and Enterprise, no-training is the default under the commercial terms; on Pro it's a switch you have to flip off yourself. Either way, as a lawyer, check the setting and the terms in force when you sign — and if you need a locked-down contractual guarantee for your most sensitive data, Enterprise (or written confirmation from Anthropic) remains the most solid route.

The source-verification layer

A setup with the right contract solves the where your data ends up. It doesn't solve the other problem with AI in legal work: invented citations — rulings and articles that sound plausible but don't exist.

For that you need an extra layer: a tool that checks every legal reference against real databases. For Italian law, the one we recommend is BeccarIA — designed precisely as a source-verification layer on top of the AI. How it works, and which databases power it (BuddaLaw, Simpliciter), is explained in the dedicated guide: MCP and source verification.

Not (yet) ready for a Team contract? Try the sandbox first

The setup above — Cowork on a Team contract with a DPA — is what we recommend for working seriously on client data. But if you don't want to commit to a Team contract yet — maybe you already use another tool (Claude, ChatGPT, something else) and just want to see how a skill works on your own case, without starting over — there's a low-threshold path: the Skillagram sandbox.

In short: you take a skill or a prompt from the catalog and run it on your own text with your own API key, straight from the browser. It's zero-server: the key and the text never pass through our servers.

The privacy / GDPR side

• On the Skillagram side it's protective: no data passes through us (zero-server, key held by you).
• The sandbox uses your own API key — the developer API, not the consumer app (ChatGPT / Claude.ai). API use is covered by the DPA (Article 28) — the provider as processor — and does not use your data to train the models. That settles the relationship with the provider (Article 28); it is not yet the question of the transfer of data outside the EU, which is a separate matter (below).
• What to watch for: (a) the DPA is between you and the provider — accepting the Commercial Terms already incorporates it; (b) data is processed on the provider's servers, often outside the EU (US): here Chapter V of the GDPR applies. The SCCs (included in the DPA) are the safeguard mechanism, but assessing the transfer stays on you as controller — it is not "closed" by the DPA alone. For a trial with minimized data it's a low threshold; for day-to-day work on client data the managed setup (Team) is the way, and for the most sensitive data consider EU residency / zero-retention (Enterprise plans or dedicated arrangements); (c) choose the provider carefully — some (e.g. DeepSeek, China) offer no adequate transfer basis; (d) minimize and pseudonymize (Recode) the data: it reduces risk and documents your diligence, but does not remove your obligations or make the transfer lawful in itself.

In one line: the sandbox is a testing tool (one prompt from the browser); the DPA covers the relationship with the provider, while the transfer outside the EU stays your assessment (simple, with minimized data). Cowork + Team remains the managed setup for day-to-day work on client files — with a DPA and organization controls already included.

Costs

• The sandbox uses your own key and your own credit with the provider: you pay per use (usually a few cents per run), not a subscription to us.
• Watch out: creating the key isn't enough — the account with the provider must have credit / active billing, otherwise the request fails.

It's the way to try the skills without changing everything: if you already use Claude, ChatGPT or something else, you bring your own key and test in a minute. When you want to bring it into the firm's day-to-day work, you move to the Cowork + Team setup above — managed and integrated. Open the sandbox.

Terms, pricing and plan availability change fast. Treat the numbers on this page as an order of magnitude, not a quote, and always confirm the conditions in force on Anthropic's official channels before subscribing.