AI for your whole team – with rules that apply before every answer.

Connect the models you already pay for, decide who may use them with which data, and keep costs and records in one place.

With your own keys or on EU infrastructure – you decide which provider sees which data.

Your team already uses AI. The only question is where.

Many people work with AI in a private browser tab. Nobody sees which data ends up there. And a policy on paper doesn't help at the moment someone hits “Send”.

What paperball.ai does for you

Your models, one place

Anthropic, OpenAI, Google, Mistral and xAI with your own key – or EU infrastructure such as Scaleway, OVHcloud and IONOS.

Assignment instead of sprawl

Models, knowledge sources, templates and skills go specifically to an organisation, department, project or person. A block always beats a grant.

Data classes, enforced

A confidential conversation only goes to a model cleared for it. The check happens before sending, not afterwards.

Budgets that hold

Limits for organisation, unit and person. The maximum cost is reserved before the request, and everyone sees their own limit.

Your own knowledge

Upload documents or connect Confluence – the assistant answers from your material.

Evidence for compliance

An AI register that derives itself, retention periods per data class and an audit log for every change.

How every request works

1Is the model assigned to you?
2Does the conversation stay with its model?
3Are profile and source active?
4May the model see this data class?
5Is the budget reserved?
Only then does your question go to the provider.

All features

Open a category to see what's included.

Models and providers

Your own keys — Connect Anthropic, OpenAI, Google, Mistral or xAI with your own contract. You pay the provider directly.
EU infrastructure — Scaleway, OVHcloud or IONOS as a source in the same catalogue; Azure OpenAI, AWS Bedrock and Google Vertex with your own instance.
Keys stay encrypted — Stored encrypted and rotatable; only the last four characters are shown.
Model profiles — The same model can appear several times in the catalogue, e.g. “for internal use” and “for confidential data, without file upload”.
Aliases instead of model names — Your team picks “Legal AI”. You swap the model behind it centrally.

Assignment and permissions

Assign to organisation, department, project or person — Models, knowledge sources, templates, skills and connectors are assigned, not unlocked for everyone.
Inheritance downwards — Whatever a department may do, every member may do. A block always beats a grant.
Delegation with a ceiling — Hand over administration, but never more than you have yourself. When the basis ends, the delegation ends.
Projects with an end date — Archive a project and its assignments stop working at that very moment – no clean-up job needed.
Roles and audit log — Owners, administrators and members; every change is logged.

Data classes and guardrails

Four data classes — Public, internal, confidential, strictly confidential. Every model profile states what it may see.
Checked before sending — A conversation only goes to a model cleared for its classification – checked before every answer.
Clearance per person — You decide who may work with confidential data – on top of the model's clearance.
Guardrails — Detect bank details, card numbers, tax IDs as well as health and HR information in inputs. What's stored is the type of finding, never the finding itself.
Classification hint on upload — If a document looks more sensitive than selected, you get a hint – without automatic upgrading.

Costs and budgets

Limits at every level — Budgets for organisation, unit and person; all of them apply at the same time.
Reserved, not just checked — The maximum cost is booked before the request. Two open tabs can't break a limit.
Everyone sees their limit — With bars and a breakdown, so a refusal is never left unexplained.
Warning before the end — At 80% and 100% of the quota, owners and administrators get a message.

Knowledge and work

Chat with streaming — Answers appear word by word; an ongoing conversation keeps its model.
Your own documents — PDF, Word, text, Markdown and HTML up to 50 MB; the assistant answers from your material. Duplicate uploads are detected.
Connect Confluence — A space is synchronised manually or at an interval you choose.
Personas — Roles with their own instruction, their own model and their own classification.
Workflows — Up to five steps in a row; every step goes through the same checks.
Prompt templates with approval — Create privately, submit, get approved. Every change is a new version.
Skills — Your own code runs in an isolated sandbox without network access.
Memory — Notes the assistant keeps about you – only with your consent and only what the model may see.

Evidence and compliance

AI register — Which AI runs here, on which model, with which provider, in which region and with which data – derived rather than entered, so never out of date.
Retention per data class — Transcripts are deleted after a deadline, with a grace period and a confirmation that states the number of affected conversations.
Analytics with a works agreement — Per-person analytics only with a works agreement (Section 87 BetrVG), visible to those affected, every access logged. Groups under five people are combined.
Labelling under the AI Act — AI outputs are labelled as such (Art. 50).
Tenant separation in the database — Every organisation only sees its own data – enforced at row level.

Switching models without risk

Test run before switching — Test sets ask a new model fixed questions before it reaches real people – optionally with a second model as judge.
Gradual roll-out — Part of the conversations go to the new model; the same conversation always stays with the same one.
One-click rollback — Back to the previous state; every switch is logged.
Approval rate — Your team's ratings sit right at the decision – from five votes onwards.
Deprecation watch — Asks the provider daily whether a model is still offered.

IT and integration

SCIM — Entra, Okta or Google add and remove members; directory groups become circles.
Second factor — For all accounts, mandatory in the organisation if you want.
REST API — Documents, chats and usage with your own key and fixed permissions.
MCP — The same features for Claude and other AI clients, with OAuth for custom connectors.
Webhooks — Signed events such as “document processed” or “quota almost used up”, with retries and a delivery log.

For your IT

SCIM for Entra, Okta and Google, a second factor, tenant separation in the database, a REST API and an MCP server with your own key, plus signed webhooks.

Go to the developer documentation →