An AI agent that only does what you agreed.
The real question is not whether AI is clever enough. It is whether you stay in control. Here is exactly how we make sure you do: autonomy levels, human approval, tests before launch, logs, and your data stored in Europe and kept out of model training.
How does theagency47 keep AI agents safe?
Every agent theagency47 builds has a written autonomy level (suggests, acts with approval, or acts on its own inside limits), must pass 90%+ of our internal eval (how we test) before launch, and logs every action it takes. Access is scope-limited, personal data is minimised, and a Data Processing Agreement is signed before any client data is touched. We are EU-based, work under the GDPR, store client data in the EU, and never train models on it.
theagency47 · Last updated: 23 October 2026A person stays in control
Each agent has a written autonomy level. It can only do what that level allows, and it hands anything unusual to a person with the full context.
02Tested before it touches real work
Every agent must pass 90%+ of our internal eval, including adversarial tests, before launch. The suite re-runs on every change; a failed run rolls back automatically.
03Every action logged
What came in, what the agent decided, which tool it used, what went out. You can audit any action, and escalations arrive with the trail attached.
04Personal data kept to a minimum
Scope-limited access, excluded categories, masking where the workflow allows, and a signed DPA before any client data touches our systems.
05EU storage, GDPR by default
We are based in the EU. Your knowledge base, logs and configuration are stored in the EU; model calls may be processed outside it under SCCs. Every sub-processor and its location is listed.
06Never trained on your data
We never train or fine-tune models on your data, and we use the providers’ business APIs, whose terms exclude training on API data.
07Your own database, never shared
On AI Company Brain builds and company-wide packages, every client gets their own EU database, knowledge vault and isolated environment.
08Locked-down staff access
2FA, least-privilege access and quarterly reviews for our team; Cloudflare Access Zero Trust sign-in for your staff on Cloudflare builds.
09You own what we build for you
Prompts, knowledge base, eval suite, configuration and integration code transfer to you upon final payment at handover.
Three levels. You choose, in writing.
Every agent page on this site shows its level. In your build, the level is part of the signed spec. No agent is built until you have signed its one-page spec, which sets its autonomy level, what it may decide alone and what it must always escalate, and none moves up a level without your sign-off.
Suggests
Drafts and recommends. A person decides and sends.
Example: a content agent drafts the article, an editor publishes it.
Acts with approval
Prepares the action. A person approves before it goes out.
Example: an email agent writes the reply, you press send.
Acts on its own
Acts inside written limits and escalates anything outside them.
Example: a support agent answers “where is my order?” and escalates a complaint.
Agents move up a level only when two things are true: the eval suite passes at the new level, and you sign off. Most start one level lower than where they end up. You can move an agent back down at any time, and every “acts on its own” agent has a written list of what it must always escalate: money above a limit, complaints, legal or medical questions, anything it is unsure about.
No eval, no launch.
Each agent has an eval suite of 20 or more cases: 60% everyday work, 20% edge cases, 15% adversarial attempts (prompt injection, false data, requests outside its job) and 5% regressions we have already fixed once. It must pass 90%+ of our internal eval (how we test) before launch, before it touches real work. The suite re-runs after every prompt or knowledge-base change, and a failed run reverts the agent to its last passing version. Read the full method.
Every action, on the record.
- What is logged: the input, the agent’s decision, the tools it called and the output it sent.
- Escalations: arrive with the full trail, so the person picking up never starts from zero.
- Retention: operational data for the engagement plus 90 days for handover; archive logs for 12 months unless your DPA says otherwise; deleted on request after that.
The least data that does the job.
- Scope-limited access. Agents connect with the narrowest permissions the task needs, read-only wherever possible.
- Exclusions. Sensitive categories can be kept out entirely, by sender, domain, folder or keyword.
- Masking. Where the workflow allows, names, phone numbers and IDs are masked before text reaches a model.
- DPA first. A Data Processing Agreement is signed before any client data touches our systems, naming every sub-processor.
European company, European rules.
theagency47 is based in the European Union and operates under the GDPR. We separate where your data is stored from where it is processed.
- Stored in the EU. Across engagements, your knowledge base, agent records, logs and configuration are stored in the EU. On AI Company Brain builds and company-wide packages (see where OS47 runs), databases run in Cloudflare D1 created in the EU jurisdiction or, for clients with strict EU requirements, in managed PostgreSQL (Supabase) in Frankfurt, with vault repositories hosted in Europe.
- Processed outside the EU when needed. Model inference, the moment an agent sends text to a model and gets an answer back, may run outside the EU. Claude requests are processed by Anthropic in the US or the EU under its terms, and open models on Cloudflare Workers AI run on Cloudflare’s global network, as do the Workers that run the agents. Where personal data leaves the EU, the European Commission’s Standard Contractual Clauses apply, and names, phone numbers and IDs are masked before text reaches a model where the workflow allows.
- How long it is kept. Anthropic keeps API inputs and outputs for 7 days and does not use them to train models. Our archive logs are stored in the EU and kept for 12 months across engagements unless your DPA says otherwise.
Every sub-processor, what it does and where it processes data is listed in our privacy policy, and material changes are announced to active clients 30 days in advance with a right to object.
Your data runs your agent. Nothing else.
We never train or fine-tune models on your data. The agents use the model providers’ business APIs, whose terms exclude using API data to train their models: Anthropic’s commercial terms for Claude, and Cloudflare’s Workers AI terms for open models. Your data is used only to run the agents you contracted, under a signed Data Processing Agreement.
Setting the same rules across your own company? Our AI governance framework turns them into owners, autonomy levels and a one-page agent charter.
Your own database. Never shared.
On AI Company Brain builds and company-wide packages (see where OS47 runs), every client has their own database, their own knowledge vault and their own isolated environment. Nothing is pooled with another client, and no two clients ever share a database.
- Your own database. A dedicated Cloudflare D1 database created in the EU jurisdiction, or, for clients with strict EU requirements, a dedicated managed PostgreSQL database (Supabase) in Frankfurt.
- Your own vault. The markdown files your agents read from sit in an EU-hosted repository of their own, never in a shared one.
- Your own isolated environment, on company-wide packages. Each client deployment runs in a dedicated per-client environment with its own EU D1 database, vault, connectors and access rules, so separation holds across compute, storage and sign-in.
- Separation inside your company, too. Public, internal and confidential knowledge are kept apart in the index, so a customer-facing agent cannot reach what only staff may see.
Other engagements are set out in your DPA and our sub-processor list, which name where your data is kept and who processes it.
Locked-down access to every environment.
Every theagency47 account that can reach client data uses two-factor authentication, holds only the access the job needs, and is reviewed every quarter. On AI Company Brain builds and company-wide packages on Cloudflare, your staff sign in through Cloudflare Access, Cloudflare’s Zero Trust gateway: nobody reaches the environment without passing the access policy, and policies are set per person or team.
What we build for you is yours.
The prompts, knowledge base, eval suite, configuration and integration code we build for you transfer to you upon final payment at handover, and you receive the full source so you can run the agents yourself or move to another partner. On company-wide packages, your dedicated EU D1 database, knowledge vault, configuration and custom agent code are handed over to you on the same terms. Your data is deleted on request after the handover period. Third-party tools and the models themselves stay under their providers’ licences.
Questions we hear before every project.
Will an AI agent do something on its own without asking?
Only what you have agreed in writing. Every agent has one of three autonomy levels: it suggests (a person decides), it acts with approval (a person approves each action), or it acts on its own inside written limits and escalates everything else. Agents start at the level agreed in the spec and move up only after passing their eval suite and with your sign-off.
Is my data used to train AI models?
No. We do not train models on client data, and the agents use the model providers’ business APIs, whose terms exclude using API data for model training. Your data is used only to run the agents you contracted, under a signed Data Processing Agreement.
Where is my data stored?
In the EU. theagency47 is based in the European Union and works under the GDPR. Across engagements, your knowledge base, agent records, logs and configuration are stored in the EU (on AI Company Brain builds and company-wide packages, in Cloudflare D1 in the EU jurisdiction or managed PostgreSQL in Frankfurt). Model inference may be processed outside the EU, for example by Anthropic in the US, under Standard Contractual Clauses. Anthropic keeps API data for 7 days; our archive logs are stored in the EU and kept for 12 months unless your DPA says otherwise. Every sub-processor and its location is listed in our privacy policy and in your DPA.
Can I see what the agent did?
Yes. Every action an agent takes is logged: what came in, what it decided, which tool it used and what went out. You can review the log, and escalations arrive with the full context attached. Archive logs are kept for 12 months unless your DPA says otherwise.
What happens to personal data the agent sees?
Agents get the narrowest access that does the job, through scope-limited connections. Sensitive categories (for example financial, legal or family email) can be excluded by sender, domain or keyword, and personal details can be masked before text reaches a model where the workflow allows it. Nothing is shared beyond the sub-processors named in your DPA.
Who owns the agent and its data?
You do. The prompts, knowledge base, eval suite, configuration and integration code built for you transfer to you upon final payment at handover, with full source, and your data is deleted on request after the handover period.
