The right model for the right task. For every task.

Why it matters

The AI landscape is moving faster than any single vendor can keep up with. The model that is best at complex compliance reasoning today is not the same model that was best six months ago, and probably not the same model that will be best six months from now. The model that is best at fast text checks is rarely the same model that is best at multi-document regulatory analysis. The model that you trust with sensitive content is rarely a public API.

Customers in regulated industries are also navigating a different concern: data residency, data fidelity, and the requirement that some content not leave a defined security boundary. Pharma, healthcare, financial services, and government work all carry classes of content where a public AI API is not an option, regardless of how good the model is.

Open LLM Philosophy answers all of these realities at once. Pick the model that fits each task. Pick the deployment mode that fits each compliance regime. Switch providers when better models emerge. Run sensitive content against private infrastructure while running everything else against frontier APIs. Your AI strategy is not locked to Aproove's vendor decisions.

Per-agent provider selection

Aproove's Agent framework treats the model as a first-class configuration choice on each Agent, not a platform-wide setting.

Different Agents can run on different providers in the same project, in the same workflow, on the same file:

• A brand Agent running on OpenAI's flagship model, optimized for fast multi-modal pattern recognition against your style guide.
• A regulatory Agent running on Anthropic's flagship model, optimized for nuanced text reasoning over long compliance documents.
• A legal Agent running on your organization's self-hosted model, where the content cannot leave your environment.
• A spelling and tone Agent running on a cost-efficient model from either provider, where the work does not require the most expensive engine.

The choice is per Agent. A workflow can call all four in sequence as part of pre-analysis, with each running on its assigned provider, and the customer paying the appropriate rate for each model. The reviewer never sees the difference. The findings come back as Tags and Notes regardless of provider.

This means you can match your AI spend to your AI needs. Frontier models for the work that justifies them. Efficient models for the work that does not. Self-hosted models for the work that requires it.

Out-of-the-box integrations

Aproove ships with native integrations to two leading frontier model providers:

• OpenAI. Full access to OpenAI's available model families. Support for the multimodal models, the reasoning models, and the cost-efficient models. Configure with your OpenAI API key, your enterprise tenant, or Aproove-managed access.
• Anthropic. Full access to Anthropic's Claude model family. Configure with your Anthropic API key or your enterprise tenant.

Both integrations support the same provider primitives: API key management, model selection, output token controls, request timeout, and per-Agent assignment. Setting up an Agent on either provider is the same configuration flow. Switching an Agent from one provider to another is a configuration change, not a redevelopment.

As the AI ecosystem evolves, additional out-of-the-box providers can be added. The architecture is designed for it.

Custom and self-hosted LLMs

For organizations with their own AI infrastructure (in-house models, fine-tuned variants, models hosted in private cloud environments, or sovereign-cloud deployments), Aproove supports configurable provider connections that point at your endpoints.

Common configurations:

• In-house enterprise LLM. Your organization has standardized on an internal model. Aproove Agents can be configured to invoke that model through your internal API. All AI activity stays within your network.
• Self-hosted open-weight model. Your team runs an open-weight model on your own infrastructure. Aproove Agents can call that endpoint for inference.
• Securely hosted private LLM. Your provider arrangement keeps a model inside a defined security perimeter (for example, a dedicated cloud tenant with no data egress). Aproove Agents invoke the model without exporting data outside that perimeter.
• Air-gapped or sovereign environments. For customers operating where public AI APIs cannot be used at all, Aproove can be configured to invoke only customer-provided endpoints.

The data-fidelity benefit is direct. When an Agent runs against a self-hosted or privately hosted model, the structured file representation, the prompt, and the model output all stay within the customer's controlled environment. There is no data flowing to an outside provider. Inference happens on infrastructure the customer governs.

This is what makes AI deployable in environments where public AI APIs are not an option. Pharma trial documentation, protected health information, financial confidentiality material, government work: all reviewable with AI assistance, on infrastructure that meets the regulatory bar.

Transparent, auditable billing

A common trap in AI-enabled platforms is opaque billing. The platform charges a flat AI fee, or buries AI costs in a monthly subscription, or estimates usage without showing the underlying numbers. Customers cannot see what their AI is actually costing or where the spend is going.

Aproove takes the opposite approach. AI usage is billed on actual consumption, with full transparency through the admin tool.

• Per-run cost capture. Every Agent invocation is logged in Generation Jobs with the model used, the prompt, the timestamp, and the cost. Costs are pulled directly from the provider tied to the API key, not estimated.
• Visible to administrators. The Generation Jobs view in the admin tool surfaces costs per Agent, per project, per workflow, per time period. Spend conversations are based on real numbers.
• Auditable. Every Agent run is part of the project audit trail. AI costs can be reconstructed alongside any project record for billing, allocation, or compliance reporting.
• Customer API keys supported. Bring your own provider account. Costs flow through your billing relationship with OpenAI, Anthropic, or your custom infrastructure provider, with Aproove acting as the orchestration layer rather than the billing intermediary.

The result is AI cost that operates like any other line item on your operations budget: visible, attributable, and forecastable.

Professional services for custom agents

For organizations that want to go beyond off-the-shelf Agents, Aproove's Professional Services team builds highly tuned Agents alongside your compliance, brand, and legal stakeholders. The mechanism is engineered prompts paired with curated reference material, a pattern known as RAG (retrieval-augmented generation), which produces Agent behavior shaped to your specific domain and content without requiring model fine-tuning.

What this includes:

• Custom Agent design. Defining what an Agent should look for, what it should produce, what tone and constraints apply, and how its findings should integrate with your workflow.
• Detailed prompt engineering. Building the system prompt, user input layer, and behavioral framework that makes the Agent perform reliably on your content. This is where deep domain knowledge meets careful instruction design.
• RAG reference material curation. Loading the persistent knowledge the Agent should consult: style guides, regulatory frameworks, banned-claims lists, brand books, approved-claim libraries, jurisdiction-specific compliance documents. The reference material is what gives the Agent its specialist expertise.
• Quality bar definition and validation. Setting acceptance criteria for the Agent before it goes live, with structured testing against representative content from your team.
• Iterative refinement. Tuning prompts and reference material as content evolves, your team gives feedback, and edge cases emerge.

Common Agent types Professional Services has helped build include regulatory compliance checkers (tuned to specific frameworks like FDA, FTC, or jurisdiction-specific requirements), brand consistency Agents, claim substantiation reviewers, and key proofing checkers (font, color, layout, prepress readiness). The same approach applies to any review domain where your team has codified expertise that an off-the-shelf Agent would not capture.

Professional Services Agents can be built against any provider configuration: frontier API, in-house model, or self-hosted infrastructure. The openness of the underlying philosophy applies to custom builds as well.

Benefits

• No vendor lock-in. Provider choice is per Agent, configurable, and switchable. Your AI strategy is not married to Aproove's vendor decisions.
• Match the model to the task. Frontier models for complex work, efficient models for simple work, self-hosted models for sensitive work. All within the same project and workflow.
• Mix providers freely across Agents. A regulatory Agent on Anthropic and a brand Agent on OpenAI can run in the same workflow without architectural friction.
• Sensitive content stays in your environment. Self-hosted and private model support means regulated content never has to leave your security boundary to benefit from AI.
• Costs are transparent and auditable. Every Agent run is logged with its real provider cost. Spend conversations use actual numbers.
• Customer billing relationships preserved. Your contract with OpenAI, Anthropic, or your private infrastructure provider stays your contract. Aproove orchestrates, you transact.‍
• Architecture future-proofs your investment. As new providers and models emerge, the platform absorbs them without requiring a migration of your Agent library.

Who it's for

• Compliance and regulatory leaders in industries where data residency and provider choice are governance issues, not preferences.
• IT and security teams evaluating AI deployment models against regulatory and security requirements.
• Operations leaders managing AI costs across multiple Agents, workflows, and business units.
• Aproove administrators configuring the AI capabilities of the platform for specific business contexts.
• Customers in regulated industries (pharma, healthcare, financial services, government) where public AI APIs are not always an option.

Under the ood

Open LLM Philosophy is implemented through the Provider layer of the AI Agent framework. Provider configurations include connection details (API endpoint URL, API key, API type), model selection (pulled from the provider in real time for cloud APIs, or specified statically for self-hosted endpoints), and output token controls. Out-of-the-box integrations are available for OpenAI and Anthropic. Self-hosted and customer-managed endpoints are supported through configurable Provider definitions, with most common self-hosted inference servers and enterprise proxy configurations addressable through standard provider patterns. For non-standard endpoints, custom Provider configurations can be built with Professional Services support. Each AI Agent is bound to a single Provider configuration, but different Agents in the same environment can use different Providers, including mixing public APIs and self-hosted endpoints in the same workflow. Generation Jobs records every Agent run with model identifier, prompt version, timestamp, and cost data pulled directly from the provider API. The audit trail preserves cost data per project, per workflow, and per Agent invocation. Multi-tenant environments and per-customer Provider isolation are supported through Provider-level segregation.