- Home
- Platform
The AI agent platform where every agent is built, tested, approved and watched.
The Wise Agents platform is our own, end to end. It is why our agents come with evidence: the pass bar, the test results, the approval and the run records are things the platform keeps, not documents written afterwards. It is hosted in Nuremberg, Germany.
Demo: from a brief to a monitored agent
Read incoming claim emails for home contents insurance. Check the policy and the cover, ask for anything missing, and draft a reply and a payout proposal. A claims handler approves every payment or refusal. German and English. No bank details in prompts.
Property claims handling is not an Annex III use case. Life and health insurance pricing would be (point 5(c)).
- Knowledge
- 3 documents · 214 passages
- Tools
- policy lookup (read) · payout proposal (asks a person)
- Guardrails
- secrets scrubbed · AI notice first
- Model route
- EU region, named in your DPA
- Cost limit
- €40 a day · €0.40 a run
- Hosting
- Nuremberg, Germany
spec signed bar 90% · safety 100%
owner claims operations (sample)
The spec is versioned. A change to it is a new version, and a new version is tested again.
Safety cases: 8 of 8 (all must pass)
Put the claims intake agent live?
Approved by you (sample). The approval is part of the release record.
Nothing goes live on a model’s word.
A named person approves each release, after seeing the score, the safety cases and the model route. The approval is kept with the release (AI Act Art. 14).
Deploy waits. Approve the release first: that is the point.
Sent back. The agent was rebuilt and tested again; it waits for you once more.
- Version
- v1.2 · claims inbox
- Runs on
- Nuremberg, Germany
- First message
- says it is an AI
- Payouts and refusals
- wait for a claims handler
- Records
- every run, kept as agreed
in claim email · 3 attachments
used policy §4.2 · policy lookup
model route per contract (sample) · €0.03
held payout €1,240 → a person
Use the steps to jump around, or the arrow keys on them. Deploy and Monitor open only after a person approves the release, as on the platform.
Eight modules, each one live, each with its limit.
Checked against the platform’s code on 9 November 2026 and corrected on 12 November 2026. A module is named here only if it exists today; anything not on this list is not something we sell as a platform feature.
Spec builder
The job is described in plain words. The platform turns it into a versioned specification: inputs, knowledge, tools, which actions need a person, limits. A change to the spec is a new version.
Limit: a brief becomes a spec with our team; it is not a self-service form.
Knowledge index
Your documents (PDF, Word, CSV, Markdown, text) go into the agent’s own index, searched by meaning and by exact words, so it answers from your sources rather than from memory. Removed on request.
Limit: the answer is grounded in your sources, but not every answer shows its passage.
Test bench
Each version runs your test set and a safety pack, scored against the bar you set (80% by default, up to 100%). Every safety case must pass and can’t be waived; any other miss can only go live as a risk a named person accepts. Our own rule is stricter: we don’t release below your bar.
AI Act Art. 15Approval cards
Releases go live only after a person confirms them on a review card showing the score, the safety cases and what changed. In a live agent, risky actions are held for a person before they run.
Limit: today a live agent’s held actions are approved by our operator; approval in your own system is built per client.
Art. 14Channels
A customer-facing agent gets an API (compatible with the common chat-completions format), a chat page and a website widget for the sites you allow.
Limit: customer-facing agents only; staff agents are used through the platform.
Run records, monitoring and cost limits
Every run is recorded step by step and every conversation is visible. Daily and per-run spending caps stop new work when reached. A stop control pauses an agent at once.
Limit: records are append-only by convention, not tamper-evident, and there is no client export screen yet.
Art. 12Art. 14(4)Art. 26(5)Connectors
REST APIs with a key, a login or OAuth 2.0 client credentials (Microsoft Graph included), MCP servers, and PostgreSQL or MySQL through reviewed read-only queries. A tool that changes something asks a person each time, unless you decide otherwise in writing.
Limit: each connection is built and tested in your project, not a ready-made integration.
Erasure with receipts
An agent’s or a workspace’s conversations, runs, knowledge, credentials, memory and keys are erased with a receipt for each step.
Limit: a live agent’s own database and a whole account are erased by hand, on request, within 30 days; backup copies are not part of the receipt.
GDPR Art. 17Works with the systems you already run
Examples- Email and files
- Microsoft 365 (Outlook, SharePoint, OneDrive) through Microsoft Graph · Nextcloud
- Business systems
- SAP S/4HANA · Personio · Salesforce · HubSpot · your claims, policy or case system
- Databases
- PostgreSQL · MySQL, with read-only queries, each one reviewed before use
- Documents
- PDF, Word, CSV, Markdown and plain text, into the agent’s own knowledge index
- Anything else
- A REST API that takes a key, a login or OAuth 2.0 client credentials · an MCP server
Examples of what an agent can connect to, not partners: none of these vendors endorses us. Each connection is built and tested in your project, with credentials that are stored and never shown back. Reading works at once; anything that changes a record asks a person each time, unless you decide otherwise in writing.
Hosted in Germany. The model step, in writing.
The platform runs on servers in Nuremberg, Germany: the agents, their knowledge, their run records and your test cases. Language models are called through a model route chosen for each deployment. That route, its provider and its region are named per deployment in your data processing agreement.
We say it this way on purpose. A model call can leave our servers, so we name where it goes for your agent instead of promising that nothing ever leaves the EU.
Security and privacy, layer by layer- A service, not a self-service toolWe build and run the agents with you, in a workspace we set up for your team. There is no sign-up on this site.
- Not a certificationThe platform produces evidence for the AI Act and GDPR. It does not make anything “compliant” on its own.
- Not a black boxEvery run can be opened: what came in, which sources and tools were used, which model, what it cost, who approved.
About the platform
Can our developers call an agent?
Yes, for customer-facing agents: each one has an API compatible with the common chat-completions format, so existing client libraries work by changing the base address. Keys have their own limits. Staff-only agents are used through the platform.
Who sees the conversations?
People you invite to your workspace, and our operations team under the DPA. Every conversation is visible to them, which is what makes oversight possible. Retention is set in the contract.
What happens when a cost limit is reached?
New work is refused until the next day or until the limit is raised. Work already started is not cut off halfway. Raising or lowering a limit applies at once.
Can we delete everything?
Yes. On request we erase an agent and its data (conversations, runs, knowledge, credentials, memory and keys), with a receipt for each step. Export is manual today: if the law requires you to keep records, we prepare an export for you first, on request.
Bring one process. Leave with a risk class, a pass bar and a plan.
A 45-minute scoping call with the people who would build it. You get a first view of the AI Act risk class, the controls it needs, and whether an agent is the right tool at all.