Control Plane · works on its own
What may run, what it costs, and the evidence.
The Control Plane is Saible's service: in Stockholm for EU organisations, in Ohio for US ones. On its own it reads your providers' bills, your contracts and what you declare and, read only, your repositories and AI accounts. Add the gateway later and each call your tools send through it joins the record.
How it works
What it can do
01Know what is in use and what it costs
- Spend overviewCost by provider, model and day, from the providers' own billing feeds.Spend → Spend overview
- ContractsEach commitment as a burn-down: consumed so far, projected to the term's end, what forfeits or runs over.Spend → Contracts
- BudgetsEach team's month against the bill, projected at the run rate, with the 80% and 100% thresholds.Spend → Budgets
- Landscape reportOne report on the AI in use: providers, models, spend by team, and the readiness gaps, ranked.Evidence → Landscape report
02Know what each tool is wired to
- ContextWhat each tool is wired to: models, prompts, tools, MCP servers, sources, memory, guardrails. Read from its own repository and accounts, confirmed as a signed manifest.Control → Context
- Connected systemsContext a tool keeps in Confluence, SharePoint or another system, listed through that system's MCP server, only within the scope an admin names.Control → Context
- Change eventsConnect a GitHub push webhook or an EventBridge rule for Bedrock, and a change tells Saible, which reads the tool's sources again. The nightly read stays the backstop.Control → Registry → a tool
- OntologyEvery business element your AI touches, where it is read from and which tools use it, with related context ranked by stated reasons.Control → Ontology
03Decide what may run
- RegistryEvery tool with its owner and where it stands: provisional, approved, stopped. Silence never widens access.Control → Registry
- PolicyThe rules in force as sentences, in one signed revision, proposed by one person and approved by a second.Control → Policy
- Review queueEvery change waits here for an org admin to decide; whoever proposed a policy change never approves it. With the gateway, its impact is replayed over your own requests.Control → Review queue
- Assessment runPast usage from the bills assessed against a versioned policy, endpoint by endpoint, before any gateway enforces it.Evidence → Assessment run
04Check it before and after it runs
- EvaluationCandidate models run over a tool's golden set, each call decided by your policy and recorded. Figures side by side, no verdict; sealed as a validation pack.Control → Evaluation
- ExposureWhat each AI platform's configuration lets through, read with the credentials you already gave Saible. Each finding with the script that fixes it.Control → Exposure
05Prove it
- RegistersWhat you declared against what the records show, and the parts of the DORA register, the GDPR Art. 30 record and the AI Act inventory they can evidence.Evidence → Registers
- ReadinessPer article of the GDPR, the AI Act and DORA: held, partly or not yet, and whose move closes the gap. A map, never a score.Evidence → Readiness
- Policy packsObligations quoted from the acts with their EUR-Lex links, and the policy rules each pack adds.Evidence → Policy packs
- Evidence packThe zip you hand an auditor: assessment records, the policy behind them, registers, readiness, and who approved what.Evidence → Evidence pack
06In the new console, by control coverage
- CatalogEvery AI system found in every repository of your GitHub organisation or GitLab group, on your connected platforms, or registered, each unclaimed until an owner claims it and adds only its purpose and risk class. Agents, MCP servers and tools in their own registries.Catalog → Systems
- Data mapWhich personal and special category data each system reaches, through which tool, and where it lives. Not classified is never read as safe.A system → Map
- CoverageEach system's controls in seven areas: covered, partly or not yet, with the next control to add. Never a verdict of compliance.Home · a system → Coverage
- Write-backA fix on Amazon Bedrock or OpenAI: proposed, dry-run against what the platform holds, approved, applied under a separate write role, re-read. By default a person applies it.Policy → Write-back
- Calls from your logsOn request, through the API: calls read from your own AI platforms' logs as metadata, never a prompt or a person, and checked against the policy in force.A system → Calls
- Policy registryThe rules in force, the GDPR, AI Act and DORA packs, the policy each system gets, and rules suggested from a scan of your code.Policy → Registry
- Model labEach model's figures on your data classes and golden sets, beside cost. Figures, never a verdict; a switch goes to the review queue.Models
- OntologyYour terms from dbt, Snowflake, Unity Catalog, Jira and Confluence; the health of the context your AI uses; an API your agents ask.Ontology
- PeoplePaid seats nobody used, from the seat export you upload, by team and service; AI services people use in the browser once the browser extension comes. Never a person.Catalog → People
- SavingsEvery saving found, in one list, with how it was computed and whether it was realised.Spend → Savings
- EvidenceOne tamper-evident ledger of every kind of record, readiness by article, sealed reports, incidents with their clocks, AI literacy.Evidence
Later, if you want
What the Gateway adds
Same console, same policy. The gateway pulls it and acts on each call sent through it.
The Gateway →- Enforce, the seventh area of coverage: a decision on each call sent through a gateway, allow, redact, route or refuse
- A chained record of each decision the gateway makes, for the ledger and the evidence pack
- Emergency stops, checked by every gateway every 15 seconds, in every mode
- Bills matched to recorded requests by provider, model and day
- Model recommendations from a replay of your own requests, once 100 of them are in the eval corpus
Where an EU organisation's data is held
- In Stockholm: the organisation's database, files and evidence, their backups, its keys, the sign-in identities of its people, and the mail it sends.
- Saible's own model calls for it are processed inside the EU, in any of six EU regions, never through a global profile.
- AWS's own log of activity in Saible's EU account is kept in the US: which identity called which AWS service, when, from which IP address, with the call's parameters, which for user administration can include a person's sign-in name. It holds no prompts, records, evidence or files.
- An organisation whose first admin uses a personal email address (gmail.com and the like) is created in Ohio, in the US. Use a work address for an organisation in Stockholm; people invited into it are held there, whatever their address.
Do it
- 1Ask for access. Saible creates your organisation in its region for you: Stockholm for the EU, Ohio for the US.
- 2Open Settings → Providers and connect a billing feed.
- 3Add your commitments on Contracts.
- 4Register your tools in the Registry.
- 5Read Readiness, then export the Evidence pack.
Control Plane topics
- Coverage01Control coverage, in seven areasWhich controls each system has. The Control Plane reaches six areas; the gateway adds Enforce.
- Inventory02Your AI inventory, from what you already runFound on your platforms, in your code and at the gateway (the browser extension is coming); claimed by its owner, who adds the purpose and the risk class.
- Spend03Spend feeds from your providersBilling data in, once a day. Matched against what the gateway recorded.
- Control04A change, approved by a second personPropose, replay over what already happened, approve. Then every gateway pulls it.
- Evidence05Chains, checks and the evidence packEach record is chained to the one before. Change one and the chain names it.
- Context06What each tool is wired toRead from its own sources, confirmed by a person, signed. A change waits for a decision.
- Evaluation07Candidate models, on your own promptsEvery call decided by your policy and recorded. Figures side by side, never a verdict.
- Exposure08What each AI platform lets throughRead only, with the credentials you already gave Saible. Each finding with the script that fixes it.