# RatioCore > RatioCore is a platform for redesigning whole business functions around AI agents. Agents run the day-to-day operation, from an online store to a security operations center. Your people set the limits, approve what matters and keep the record. It runs on your infrastructure, with your models, and without lock-in. RatioCore is an agentic platform for redesigning business functions around AI agents. Agents run the work, people set the limits and approve what matters, and you keep your infrastructure, models and data. Conversations shown on the website are illustrative examples with sample data. ## Six questions the Core answers once - **Who can order the agent to do what?** Tenants, identities, groups and grants. Every agent gets its own service identity, and every authorization decision is logged. - **What needs human approval?** Versioned policies and declared tool risk (read, write, destructive) decide. Risky actions wait in an approval queue. - **What happened?** An audit trail, turn traces and tool events show what each agent did, step by step. - **What did it cost?** Tokens are metered per agent per day, with daily budgets and quotas. - **How do we recover when something fails?** Durable admission and delivery, leases and backup checkpoints keep work recoverable. - **Who is responsible?** Named principals, separation of duties and decision logs keep ownership clear. ## Key capabilities - **Build the dialog** (Design): Guide people through what the agent needs to know. Use structured intake for work that starts from messaging, a business system or an API. - **Draw the workflow** (Orchestrate): Set the sequence, branches and approval steps in the visual workflow builder. Use fixed rules for predictable work and agent judgment where context matters. - **Give agents the right context** (Knowledge): Retrieve from enterprise documents and records with vector and lexical search. Add knowledge graph relationships where the workflow calls for them. - **Connect native, API and MCP tools** (Action): Define which tools agents can use and which actions require approval. Extend workflows with integrations, APIs and MCP tools. - **Assign owners and boundaries** (Control): Tenant and role permissions limit access. Sensitive actions route to the people responsible, and authorization is rechecked before execution. - **Manage agents after launch** (Operate): Use a shared agent catalog, release controls and operational views. Inspect runs, decisions and external effects as your fleet grows. ## Governance: how a protected action is handled - **Before execution — Define authority once**: Tenant-scoped identity, role permissions and versioned policies establish who can use a tool and when approval is required. Missing authority blocks dispatch. - **At execution — Check the action again**: Approval is specific to the intended action. The runtime rechecks permissions before dispatch and records the action and its outcome. - **After execution — Give operators a record**: Follow policy decisions, approvals, tool calls and outcomes in context. An uncertain external result requires reconciliation before anyone retries it. ## Deployment - **Run it where you choose**: On premises, in your private cloud or on AMD Cloud. The runtime, data, records and backups stay where you decide. - **Use the models you choose**: Local models for restricted work, approved external models where policy permits, and different routes for different workflows. Models never carry the authority to act. - **Keep your data and your systems**: Data stays in your own PostgreSQL database. Connect the systems you already run through APIs and MCP tools, and review every path before you enable it. Note: External models and connectors add their own data paths. Self-hosting is a deployment option; sovereignty depends on the full configuration. ## Builders - **Dialogue builder**: Choose what to ask, branch on the answers and capture the details and evidence the next step needs, from an approved channel or guided interface. Example: Access request: which system, which role, when access should expire. - **Visual workflow builder**: Map decisions, agent tasks, approvals and system actions. Fixed rules control routing; agents gather evidence or recommend a next step. Example: Check eligibility, route to the system owner, grant time-limited access. - **Agents and tools**: Give specialist agents a clear job and connect native, API or MCP tools. Policy and role permissions govern what each one can run. Example: Invoice exception: an agent pulls evidence from invoice and purchase order. ## Reference applications - [Commerce](https://ratiocore.ai/applications/commerce.html): A store run by agents. Agents source products, prepare listings and follow every order to delivery. The owner sets the limits and approves what goes live. - [Security Operations](https://ratiocore.ai/applications/security.html): An agentic SOC. Agents triage every alert, investigate it and propose the response. Their autonomy grows as their performance earns it, and analysts approve containment. - [IT Operations](https://ratiocore.ai/applications/it-operations.html): IT operations, run by agents. Agents group alerts into one situation, find the likely cause and prepare the fix. The incident lead approves the change. - [Vendor Management](https://ratiocore.ai/applications/vendor-management.html): Supplier management, run by agents. Agents check every supplier against its contract, compute remedies and prepare the follow-up. The relationship owner approves. - [GeoSupply](https://ratiocore.ai/applications/geosupply.html): A supply chain that watches the world. Agents map global disruptions to your suppliers and model the options. Your planners choose the response. - [Predictive Maintenance](https://ratiocore.ai/applications/predictive-maintenance.html): Maintenance that plans ahead. Agents spot early signs of failure and draft the work order. Supervisors approve every intervention. - [Adversarial Emulation](https://ratiocore.ai/applications/adversarial-emulation.html): Agents that test your defenses. Agents run authorized attack exercises in a scoped range and show what was detected, what was missed and what to retest. ## Application differentiators - **Security Operations**: An autonomy ladder from AL0 to AL3, with approvals bound to the exact action and its parameters. - **Vendor Management**: No LLM in the arithmetic: remedies are computed deterministically, with a hash-chained audit trail. - **GeoSupply**: 32 public data feeds on sanctions, ports, hazards and trade, turned into scenarios and decisions. - **Predictive Maintenance**: Every work order is approved by a person. - **Adversarial Emulation**: A safety core with 20+ checks, signed rules of engagement, a kill switch and four-eyes approval. ## Design stance - **Fail closed on authority**: Admission, identity, the execution gate, tool bindings and approvals refuse when in doubt. - **Fail open on enrichment**: Memory, grounding and catalog overlays degrade gracefully instead of blocking a turn. - **Durable before visible**: Nothing is shown to a user that has not first been stored. - **The LLM proposes, the runtime validates, a human publishes**: Applied alike to workflows, dialogs and agent profiles. ## Product controls - **Workflows never half-apply**: Every effect, receipt and journal entry commits in one fenced transaction. - **Approvals that expire**: Destructive steps get an approval bound to the specific run, step and input, with an expiry. - **Immutable versions**: Published workflows and dialogs cannot change; each run snapshots the definition and its hash. - **Deterministic policy**: Deny beats approval beats allow, and every policy can be simulated before activation. - **Zero-token replies**: Dialog respond steps return a fixed answer with no model call at all. - **An instant off switch**: Removing a dialog's binding from an agent switches that dialog off immediately. ## Cost control - Daily token budget per agent; the next turn is refused once it is spent - Prompt and completion tokens metered per agent per day - Per-turn token budgets with announced degradation - A kill switch for any agent: disabling it also stops its runtime ## Reliability - Inputs are admitted once: retrying the same message never runs it twice - Answers are stored before they are shown - Final answers are delivered at least once, and possible repeats are flagged - GPU capacity is shared through a fair queue, so one busy agent cannot starve the others - Model servers fail over to a compatible server when one goes down ## Who it is for - **CEO / CFO** (A business that runs differently): Functions run by agents, with budgets, limits and an off switch. - **CISO / compliance** (Evidence and control): Every action checked, approved where needed and logged. - **CIO / architect** (One foundation): One platform for every function, on your infrastructure, with no lock-in. - **Operations leaders** (Capacity): Less time collecting information, more time deciding. - **IT / platform team** (Operability): One console to deploy, monitor and stop agents. - **Employees** (A trustworthy assistant): Help in familiar channels, with clear approval prompts. ## How engagements work 1. **Reimagine the process**: Decide what agents run, what people keep and which limits apply. 2. **Run a controlled pilot**: In your environment, with a few users and approved tools. 3. **Go to production**: Integrations, policies and support, tested for security and performance. 4. **Extend across the company**: Reinvent the next function on the same platform and controls. ## What a pilot tests - **Agent lifecycle**: Create, configure, disable and redeploy an agent. Evidence: Audit and revision trail - **Workflow control**: Run a multi-step process with human approval. Evidence: Step journal, decision and result - **Dialog experience**: Collect inputs and hand off to an agent. Evidence: Scenario replay and completion record - **Integration**: Read and write through a constrained adapter. Evidence: API outcome and authorization proof - **Model choice**: Route to an approved, compatible model. Evidence: Served-model check and quality test - **Memory**: Recall permitted information, reject cross-scope retrieval. Evidence: Isolation and provenance tests - **Failure handling**: Interrupt and retry a turn, replay durable state. Evidence: Consistent final state, no unsafe duplicate effects - **Security**: Attempt unauthorized admin and tool actions. Evidence: Denial and audit evidence - **Operations**: Back up, restore and detect a service failure. Evidence: Recovery test and alert record - **Economics**: Measure a representative workload. Evidence: Tokens, infrastructure cost and outcome metrics ## Deployment paths - [AMD cloud](https://ratiocore.ai/deployment/amd-cloud.html): Serve open-weight models on AMD Instinct GPUs from Azure, Oracle Cloud, DigitalOcean and other providers, with the Core in your cloud account. - [On premises: AMD or NVIDIA](https://ratiocore.ai/deployment/on-premises.html): Install RatioCore on your own servers with AMD Instinct, NVIDIA, or both, for regulated and network-restricted environments. - [Fully cloud](https://ratiocore.ai/deployment/cloud.html): Run every tier in your AWS, Azure, Google Cloud or Oracle account, with self-served models, managed model APIs, or both. ## RatioCore is / is not Is: An agentic platform for redesigning business functions; A governance and execution layer for agents that act; A way to build controlled workflows and dialogs; A platform you can run on your own infrastructure; Open to the models and systems you choose; A foundation for optional business applications. Is not: A system of record that only stores what people did; A single chatbot or a collection of personas; An AI foundation model; An unrestricted tool-running assistant; A lock-in to one cloud, model or vendor; A requirement to buy every vertical application. ## FAQ ### What kind of work is RatioCore good for? Any business process that moves across several systems and ends in decisions or actions. The applications built so far run an online store, security operations, IT operations, supplier management, supply-chain risk, maintenance and security testing. The same approach works for other processes in your company: agents run the work, and people approve what matters. ### Is RatioCore an ERP? Not in the traditional sense. An ERP records the transactions people carry out. RatioCore runs the work itself: agents operate business functions across your existing systems, including your ERP, while people set the limits and approve. You can think of it as an agentic ERP, one platform for running many functions with agents. ### How does RatioCore avoid vendor lock-in? You choose where it runs: on premises, in your private cloud or on AMD Cloud. You choose the models, local or through an API, and different workflows can use different models. Your data stays in your own PostgreSQL database, and your systems connect through APIs and MCP tools. ### Do you help us redesign our processes? Yes. We work with your leaders and teams to decide what agents should run, which limits and approvals apply and how to measure the result. We start with one function, prove it in a pilot, and extend from there. ### How do we get started? Book a demo and bring one process your team runs today. We map what an agent could prepare and what your people keep, then scope a controlled pilot in your environment. ### Do we need AI engineers to use it? No. Applications are designed with guided dialog and visual workflow builders. Your IT team is involved in connecting systems and setting policies, and we support that work during the pilot. ### Can an agent take an action without approval? Only where your policy allows it. RatioCore checks tenant, role and policy before a protected tool runs. Sensitive actions can wait for a named approver, and missing authority blocks dispatch. You decide which steps agents take alone and which need a person. ### What happens if permissions change after an approval is given? Approval is specific to the intended action, and the runtime rechecks permissions just before dispatch. If the required authority is missing at that moment, the tool call is blocked. The action and its outcome are recorded either way. ### Can we run RatioCore in our own environment? Yes. RatioCore can be self-hosted on customer-managed infrastructure, on premises or in a private cloud. The current reference uses PostgreSQL 17 with pgvector and lexical search. Deployment details depend on the workload, and sovereignty depends on the full configuration, including models and connectors. ### Which models does RatioCore use, and does our data leave the boundary? You choose. Evaluate local inference for restricted workflows, or connect an approved external model where policy permits. Different workflows can use different model routes, and models never gain authority to execute actions. External models, channels and connectors add their own data paths, so review each before enabling it. ## Pages - [Home](https://ratiocore.ai/) - [Platform](https://ratiocore.ai/platform.html) - [Applications](https://ratiocore.ai/applications.html) - [Deployment](https://ratiocore.ai/deployment.html) - [Build your own](https://ratiocore.ai/build.html) - [Book a product demo](https://ratiocore.ai/demo.html)