Automatic model selection
A router chooses the model for each request by task, sensitivity and cost, and keeps work on the providers and locations you have approved. People ask; they do not have to pick a model.
The AI internal control system
Harness for companies. Zero-trust Private Queries.
ZPQ is the mandatory control layer between your organization and every model, every agent and every data source. Nothing reaches a model, and nothing comes back, without passing your own rules first.
identity context model action one permitted AI transaction
What ZPQ does
The layer that enforces your rules also carries the everyday work: it chooses the model, assembles the context, opens the right topics to the right people — and keeps every provider's spend in one place.
Company knowledge in every answer
Each request is assembled with the context it needs. Relevant documents from your own knowledge base are retrieved and added, limited to what the person asking is entitled to see.
Access that follows the org chart
Topics and projects are visible only to the units and roles that own them. Documents carry their status, in progress or completed, so people work from what is current.
One place for AI limits
Set spending limits by company, unit and provider, and see usage across every provider in a single view, instead of reconciling separate consoles.
Dashboards for every level
Roll-ups of topics, decisions and activity for teams, departments and leadership. Each level sees the picture it is entitled to, and nothing beyond it.
Skills for recurring work
A starter set of organizational skills: preparing for meetings with your manager, clients or peers; handling email; writing the daily briefing; managing your calendar.
Every AI arrives with rules your organization did not write
The behaviour of a deployed model is decided before your policy is ever consulted — and the gap only widens as connectors and tools are switched on.
Someone else's policy runs first
System prompts and provider policies shape behaviour ahead of your own rules. Your organization is a guest inside its own workflow.
Connectors expose what is reachable
Integrations open whatever the credential can reach — not what is appropriate for this role, on this topic, for this purpose.
Tool access turns answers into changes
Once a model can call tools, a reply is no longer a reply. It is a change written into your systems.
Teams diverge while leadership assumes alignment
Different teams get different context and different answers, and management continues to believe everyone is working from the same picture.
If AI cannot be governed as an organizational actor, it should not be deployed.
Every AI request becomes a permitted organizational transaction
One organizational context, one execution path, one evidence record. No match across identity, context and policy means no execution.
- Request A person or an agent asks for something.
- Resolve Role, topic and organizational context are established before anything else happens.
- Authorize Data, model and action are each permitted — or they are not.
- Route Execution is directed to cloud, VPC or on-premise according to the decision.
- Prove The decision, its sources and its result are recorded as evidence.
Your org chart, compiled into executable AI policy
ZPQ turns how the organization is actually structured into rules that run at request time.
Organization model
Units and roles, topics and context areas, entitlements to sources, and identities for agents — the organization described in terms a policy can evaluate.
Policy compiler
Resolves identity and purpose, assembles the rules that apply to the sources and topics in play, authorizes model and action, and then denies, approves or escalates.
Execution network
Semantic model routing across cloud, VPC and on-premise, with approvals, tool limits, and tamper-evident proof of what happened.
- Employees see a single point of entry.
- Leadership sees AI that is consistent across the organization.
- Security sees one decision path instead of many.
Who it is for
One requirement runs through all of them: context, topics and data sources must be governed explicitly, by unit and by organizational level.
Cloud, 10–249 people
Companies that need AI under organizational control from the start, without standing up infrastructure to get it.
Cloud, large organizations
Multiple business units, SSO and SCIM, and a VPC option where isolation is required.
On-premise, regulated industries
Private models, data residency requirements and audit obligations that rule out shared infrastructure.
On-premise, sovereign and isolated
Environments that must operate entirely within their own perimeter.
A control directors already treat as mandatory
Nothing here is a new expectation. It is the standard every other part of the business is already held to.
- No financial close without rules and accountability.
- No system without access rights and the ability to revoke them.
- No proof without an immutable record of decisions and sources.
Context
Who is entitled to know and to speak about what, and for which organizational purpose.
Execution
Which model, running where, permitted to take which action.
Evidence
What was permitted, why, on which sources, and with what result.
This is not AI security. It is organizational permission for AI.
Team
Mikhail Kharichko
Founder & CEO
20+ years building software and scaling operations. Scaled Dostavista / Borzo, led operations at YClients and Rabbit Care. Background in applied mathematics and computer science. Built AI operations in insurance handling roughly 20,000 calls a day.
Aleksey Gladkov
Co-founder & CAIO
Founder of the AI Advent Challenge.
Where we are
ZPQ is an early-stage startup, currently in development. The product described on this page is what we are building. We are talking to organizations that already treat AI as something that has to be governed.